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);
205 setOperationAction(ISD::SUB, MVT::i64, Expand);
207 setOperationAction(ISD::UDIV, MVT::i32, Expand);
208 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
209 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
210 setOperationAction(ISD::UREM, MVT::i32, Expand);
211 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
212 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
214 static const MVT::SimpleValueType IntTypes[] = {
215 MVT::v2i32, MVT::v4i32
217 const size_t NumIntTypes = array_lengthof(IntTypes);
219 for (unsigned int x = 0; x < NumIntTypes; ++x) {
220 MVT::SimpleValueType VT = IntTypes[x];
221 //Expand the following operations for the current type by default
222 setOperationAction(ISD::ADD, VT, Expand);
223 setOperationAction(ISD::AND, VT, Expand);
224 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
225 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
226 setOperationAction(ISD::MUL, VT, Expand);
227 setOperationAction(ISD::OR, VT, Expand);
228 setOperationAction(ISD::SHL, VT, Expand);
229 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
230 setOperationAction(ISD::SRL, VT, Expand);
231 setOperationAction(ISD::SRA, VT, Expand);
232 setOperationAction(ISD::SUB, VT, Expand);
233 setOperationAction(ISD::UDIV, VT, Expand);
234 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
235 setOperationAction(ISD::UREM, VT, Expand);
236 setOperationAction(ISD::SELECT, VT, Expand);
237 setOperationAction(ISD::VSELECT, VT, Expand);
238 setOperationAction(ISD::XOR, VT, Expand);
241 static const MVT::SimpleValueType FloatTypes[] = {
242 MVT::v2f32, MVT::v4f32
244 const size_t NumFloatTypes = array_lengthof(FloatTypes);
246 for (unsigned int x = 0; x < NumFloatTypes; ++x) {
247 MVT::SimpleValueType VT = FloatTypes[x];
248 setOperationAction(ISD::FABS, VT, Expand);
249 setOperationAction(ISD::FADD, VT, Expand);
250 setOperationAction(ISD::FCOS, VT, Expand);
251 setOperationAction(ISD::FDIV, VT, Expand);
252 setOperationAction(ISD::FPOW, VT, Expand);
253 setOperationAction(ISD::FFLOOR, VT, Expand);
254 setOperationAction(ISD::FTRUNC, VT, Expand);
255 setOperationAction(ISD::FMUL, VT, Expand);
256 setOperationAction(ISD::FRINT, VT, Expand);
257 setOperationAction(ISD::FSQRT, VT, Expand);
258 setOperationAction(ISD::FSIN, VT, Expand);
259 setOperationAction(ISD::FSUB, VT, Expand);
260 setOperationAction(ISD::SELECT, VT, Expand);
263 setTargetDAGCombine(ISD::MUL);
266 //===----------------------------------------------------------------------===//
267 // Target Information
268 //===----------------------------------------------------------------------===//
270 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
274 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
276 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
279 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
280 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
282 return ((LScalarSize <= CastScalarSize) ||
283 (CastScalarSize >= 32) ||
287 //===---------------------------------------------------------------------===//
289 //===---------------------------------------------------------------------===//
291 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
292 assert(VT.isFloatingPoint());
293 return VT == MVT::f32;
296 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
297 assert(VT.isFloatingPoint());
298 return VT == MVT::f32;
301 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
302 // Truncate is just accessing a subregister.
303 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
306 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
307 // Truncate is just accessing a subregister.
308 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
309 (Dest->getPrimitiveSizeInBits() % 32 == 0);
312 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
313 const DataLayout *DL = getDataLayout();
314 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
315 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
317 return SrcSize == 32 && DestSize == 64;
320 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
321 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
322 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
323 // this will enable reducing 64-bit operations the 32-bit, which is always
325 return Src == MVT::i32 && Dest == MVT::i64;
328 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
329 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
330 // limited number of native 64-bit operations. Shrinking an operation to fit
331 // in a single 32-bit register should always be helpful. As currently used,
332 // this is much less general than the name suggests, and is only used in
333 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
334 // not profitable, and may actually be harmful.
335 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
338 //===---------------------------------------------------------------------===//
339 // TargetLowering Callbacks
340 //===---------------------------------------------------------------------===//
342 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
343 const SmallVectorImpl<ISD::InputArg> &Ins) const {
345 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
348 SDValue AMDGPUTargetLowering::LowerReturn(
350 CallingConv::ID CallConv,
352 const SmallVectorImpl<ISD::OutputArg> &Outs,
353 const SmallVectorImpl<SDValue> &OutVals,
354 SDLoc DL, SelectionDAG &DAG) const {
355 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
358 //===---------------------------------------------------------------------===//
359 // Target specific lowering
360 //===---------------------------------------------------------------------===//
362 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
363 SmallVectorImpl<SDValue> &InVals) const {
364 SDValue Callee = CLI.Callee;
365 SelectionDAG &DAG = CLI.DAG;
367 const Function &Fn = *DAG.getMachineFunction().getFunction();
369 StringRef FuncName("<unknown>");
371 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
372 FuncName = G->getSymbol();
373 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
374 FuncName = G->getGlobal()->getName();
376 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
377 DAG.getContext()->diagnose(NoCalls);
381 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
383 switch (Op.getOpcode()) {
385 Op.getNode()->dump();
386 llvm_unreachable("Custom lowering code for this"
387 "instruction is not implemented yet!");
389 // AMDIL DAG lowering
390 case ISD::SDIV: return LowerSDIV(Op, DAG);
391 case ISD::SREM: return LowerSREM(Op, DAG);
392 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
393 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
394 // AMDGPU DAG lowering
395 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
396 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
397 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
398 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
399 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
400 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
405 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
406 SmallVectorImpl<SDValue> &Results,
407 SelectionDAG &DAG) const {
408 switch (N->getOpcode()) {
409 case ISD::SIGN_EXTEND_INREG:
410 // Different parts of legalization seem to interpret which type of
411 // sign_extend_inreg is the one to check for custom lowering. The extended
412 // from type is what really matters, but some places check for custom
413 // lowering of the result type. This results in trying to use
414 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
415 // nothing here and let the illegal result integer be handled normally.
418 SDValue Op = SDValue(N, 0);
420 EVT VT = Op.getValueType();
421 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
422 N->getOperand(0), N->getOperand(1));
423 Results.push_back(UDIVREM);
427 SDValue Op = SDValue(N, 0);
429 EVT VT = Op.getValueType();
430 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
431 N->getOperand(0), N->getOperand(1));
432 Results.push_back(UDIVREM.getValue(1));
436 SDValue Op = SDValue(N, 0);
438 EVT VT = Op.getValueType();
439 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
441 SDValue one = DAG.getConstant(1, HalfVT);
442 SDValue zero = DAG.getConstant(0, HalfVT);
445 SDValue LHS = N->getOperand(0);
446 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
447 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
449 SDValue RHS = N->getOperand(1);
450 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
451 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
453 // Get Speculative values
454 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
455 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
457 SDValue REM_Hi = zero;
458 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
460 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
461 SDValue DIV_Lo = zero;
463 const unsigned halfBitWidth = HalfVT.getSizeInBits();
465 for (unsigned i = 0; i < halfBitWidth; ++i) {
466 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
467 // Get Value of high bit
469 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
470 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
472 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
473 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
476 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
477 DAG.getConstant(halfBitWidth - 1, HalfVT));
478 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
479 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
481 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
482 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
485 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
487 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
488 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
490 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
494 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
496 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
497 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
498 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
501 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
502 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
503 Results.push_back(DIV);
504 Results.push_back(REM);
512 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
513 const GlobalValue *GV,
514 const SDValue &InitPtr,
516 SelectionDAG &DAG) const {
517 const DataLayout *TD = getTargetMachine().getDataLayout();
519 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
520 EVT VT = EVT::getEVT(CI->getType());
521 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
522 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
523 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
524 TD->getPrefTypeAlignment(CI->getType()));
525 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
526 EVT VT = EVT::getEVT(CFP->getType());
527 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
528 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
529 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
530 TD->getPrefTypeAlignment(CFP->getType()));
531 } else if (Init->getType()->isAggregateType()) {
532 EVT PtrVT = InitPtr.getValueType();
533 unsigned NumElements = Init->getType()->getArrayNumElements();
534 SmallVector<SDValue, 8> Chains;
535 for (unsigned i = 0; i < NumElements; ++i) {
536 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
537 Init->getType()->getArrayElementType()), PtrVT);
538 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
539 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
540 GV, Ptr, Chain, DAG));
542 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
545 llvm_unreachable("Unhandled constant initializer");
549 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
551 SelectionDAG &DAG) const {
553 const DataLayout *TD = getTargetMachine().getDataLayout();
554 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
555 const GlobalValue *GV = G->getGlobal();
557 switch (G->getAddressSpace()) {
558 default: llvm_unreachable("Global Address lowering not implemented for this "
560 case AMDGPUAS::LOCAL_ADDRESS: {
561 // XXX: What does the value of G->getOffset() mean?
562 assert(G->getOffset() == 0 &&
563 "Do not know what to do with an non-zero offset");
566 if (MFI->LocalMemoryObjects.count(GV) == 0) {
567 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
568 Offset = MFI->LDSSize;
569 MFI->LocalMemoryObjects[GV] = Offset;
570 // XXX: Account for alignment?
571 MFI->LDSSize += Size;
573 Offset = MFI->LocalMemoryObjects[GV];
576 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
578 case AMDGPUAS::CONSTANT_ADDRESS: {
579 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
580 Type *EltType = GV->getType()->getElementType();
581 unsigned Size = TD->getTypeAllocSize(EltType);
582 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
584 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
585 const Constant *Init = Var->getInitializer();
586 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
587 SDValue InitPtr = DAG.getFrameIndex(FI,
588 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
589 SmallVector<SDNode*, 8> WorkList;
591 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
592 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
593 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
595 WorkList.push_back(*I);
597 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
598 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
599 E = WorkList.end(); I != E; ++I) {
600 SmallVector<SDValue, 8> Ops;
601 Ops.push_back(Chain);
602 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
603 Ops.push_back((*I)->getOperand(i));
605 DAG.UpdateNodeOperands(*I, Ops);
607 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
608 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
613 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
614 SelectionDAG &DAG) const {
615 SmallVector<SDValue, 8> Args;
616 SDValue A = Op.getOperand(0);
617 SDValue B = Op.getOperand(1);
619 DAG.ExtractVectorElements(A, Args);
620 DAG.ExtractVectorElements(B, Args);
622 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
625 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
626 SelectionDAG &DAG) const {
628 SmallVector<SDValue, 8> Args;
629 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
630 EVT VT = Op.getValueType();
631 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
632 VT.getVectorNumElements());
634 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
637 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
638 SelectionDAG &DAG) const {
640 MachineFunction &MF = DAG.getMachineFunction();
641 const AMDGPUFrameLowering *TFL =
642 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
644 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
647 unsigned FrameIndex = FIN->getIndex();
648 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
649 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
653 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
654 SelectionDAG &DAG) const {
655 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
657 EVT VT = Op.getValueType();
659 switch (IntrinsicID) {
661 case AMDGPUIntrinsic::AMDIL_abs:
662 return LowerIntrinsicIABS(Op, DAG);
663 case AMDGPUIntrinsic::AMDIL_exp:
664 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
665 case AMDGPUIntrinsic::AMDGPU_lrp:
666 return LowerIntrinsicLRP(Op, DAG);
667 case AMDGPUIntrinsic::AMDIL_fraction:
668 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
669 case AMDGPUIntrinsic::AMDIL_max:
670 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
672 case AMDGPUIntrinsic::AMDGPU_imax:
673 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
675 case AMDGPUIntrinsic::AMDGPU_umax:
676 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
678 case AMDGPUIntrinsic::AMDIL_min:
679 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
681 case AMDGPUIntrinsic::AMDGPU_imin:
682 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
684 case AMDGPUIntrinsic::AMDGPU_umin:
685 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
688 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
689 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
694 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
695 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
700 case AMDGPUIntrinsic::AMDGPU_bfi:
701 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
706 case AMDGPUIntrinsic::AMDGPU_bfm:
707 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
711 case AMDGPUIntrinsic::AMDIL_round_nearest:
712 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
716 ///IABS(a) = SMAX(sub(0, a), a)
717 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
718 SelectionDAG &DAG) const {
721 EVT VT = Op.getValueType();
722 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
725 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
728 /// Linear Interpolation
729 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
730 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
731 SelectionDAG &DAG) const {
733 EVT VT = Op.getValueType();
734 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
735 DAG.getConstantFP(1.0f, MVT::f32),
737 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
739 return DAG.getNode(ISD::FADD, DL, VT,
740 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
744 /// \brief Generate Min/Max node
745 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
746 SelectionDAG &DAG) const {
748 EVT VT = Op.getValueType();
750 SDValue LHS = Op.getOperand(0);
751 SDValue RHS = Op.getOperand(1);
752 SDValue True = Op.getOperand(2);
753 SDValue False = Op.getOperand(3);
754 SDValue CC = Op.getOperand(4);
756 if (VT != MVT::f32 ||
757 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
761 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
775 llvm_unreachable("Operation should already be optimised!");
783 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
785 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
794 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
796 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
798 case ISD::SETCC_INVALID:
799 llvm_unreachable("Invalid setcc condcode!");
804 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
805 SelectionDAG &DAG) const {
806 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
807 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
808 EVT EltVT = Op.getValueType().getVectorElementType();
809 EVT PtrVT = Load->getBasePtr().getValueType();
810 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
811 SmallVector<SDValue, 8> Loads;
814 for (unsigned i = 0, e = NumElts; i != e; ++i) {
815 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
816 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
817 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
818 Load->getChain(), Ptr,
819 MachinePointerInfo(Load->getMemOperand()->getValue()),
820 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
821 Load->getAlignment()));
823 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
826 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
827 SelectionDAG &DAG) const {
828 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
829 EVT MemVT = Store->getMemoryVT();
830 unsigned MemBits = MemVT.getSizeInBits();
832 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
833 // truncating store into an i32 store.
834 // XXX: We could also handle optimize other vector bitwidths.
835 if (!MemVT.isVector() || MemBits > 32) {
840 SDValue Value = Store->getValue();
841 EVT VT = Value.getValueType();
842 EVT ElemVT = VT.getVectorElementType();
843 SDValue Ptr = Store->getBasePtr();
844 EVT MemEltVT = MemVT.getVectorElementType();
845 unsigned MemEltBits = MemEltVT.getSizeInBits();
846 unsigned MemNumElements = MemVT.getVectorNumElements();
847 unsigned PackedSize = MemVT.getStoreSizeInBits();
848 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
850 assert(Value.getValueType().getScalarSizeInBits() >= 32);
853 for (unsigned i = 0; i < MemNumElements; ++i) {
854 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
855 DAG.getConstant(i, MVT::i32));
856 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
857 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
859 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
860 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
865 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
869 if (PackedSize < 32) {
870 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
871 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
872 Store->getMemOperand()->getPointerInfo(),
874 Store->isNonTemporal(), Store->isVolatile(),
875 Store->getAlignment());
878 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
879 Store->getMemOperand()->getPointerInfo(),
880 Store->isVolatile(), Store->isNonTemporal(),
881 Store->getAlignment());
884 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
885 SelectionDAG &DAG) const {
886 StoreSDNode *Store = cast<StoreSDNode>(Op);
887 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
888 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
889 EVT PtrVT = Store->getBasePtr().getValueType();
890 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
893 SmallVector<SDValue, 8> Chains;
895 for (unsigned i = 0, e = NumElts; i != e; ++i) {
896 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
897 Store->getValue(), DAG.getConstant(i, MVT::i32));
898 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
900 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
902 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
903 MachinePointerInfo(Store->getMemOperand()->getValue()),
904 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
905 Store->getAlignment()));
907 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
910 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
912 LoadSDNode *Load = cast<LoadSDNode>(Op);
913 ISD::LoadExtType ExtType = Load->getExtensionType();
914 EVT VT = Op.getValueType();
915 EVT MemVT = Load->getMemoryVT();
917 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
918 // We can do the extload to 32-bits, and then need to separately extend to
921 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
925 Load->getMemOperand());
926 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
929 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
930 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
931 // FIXME: Copied from PPC
932 // First, load into 32 bits, then truncate to 1 bit.
934 SDValue Chain = Load->getChain();
935 SDValue BasePtr = Load->getBasePtr();
936 MachineMemOperand *MMO = Load->getMemOperand();
938 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
939 BasePtr, MVT::i8, MMO);
940 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
943 // Lower loads constant address space global variable loads
944 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
946 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
948 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
949 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
950 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
951 DAG.getConstant(2, MVT::i32));
952 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
953 Load->getChain(), Ptr,
954 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
957 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
958 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
962 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
963 DAG.getConstant(2, MVT::i32));
964 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
965 Load->getChain(), Ptr,
966 DAG.getTargetConstant(0, MVT::i32),
968 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
970 DAG.getConstant(0x3, MVT::i32));
971 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
972 DAG.getConstant(3, MVT::i32));
974 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
976 EVT MemEltVT = MemVT.getScalarType();
977 if (ExtType == ISD::SEXTLOAD) {
978 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
979 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
982 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
985 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
987 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
988 if (Result.getNode()) {
992 StoreSDNode *Store = cast<StoreSDNode>(Op);
993 SDValue Chain = Store->getChain();
994 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
995 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
996 Store->getValue().getValueType().isVector()) {
997 return SplitVectorStore(Op, DAG);
1000 EVT MemVT = Store->getMemoryVT();
1001 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1002 MemVT.bitsLT(MVT::i32)) {
1004 if (Store->getMemoryVT() == MVT::i8) {
1006 } else if (Store->getMemoryVT() == MVT::i16) {
1009 SDValue BasePtr = Store->getBasePtr();
1010 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1011 DAG.getConstant(2, MVT::i32));
1012 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1013 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1015 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1016 DAG.getConstant(0x3, MVT::i32));
1018 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1019 DAG.getConstant(3, MVT::i32));
1021 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1024 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1026 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1027 MaskedValue, ShiftAmt);
1029 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1031 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1032 DAG.getConstant(0xffffffff, MVT::i32));
1033 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1035 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1036 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1037 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1042 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1043 SelectionDAG &DAG) const {
1045 EVT VT = Op.getValueType();
1047 SDValue Num = Op.getOperand(0);
1048 SDValue Den = Op.getOperand(1);
1050 // RCP = URECIP(Den) = 2^32 / Den + e
1051 // e is rounding error.
1052 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1054 // RCP_LO = umulo(RCP, Den) */
1055 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1057 // RCP_HI = mulhu (RCP, Den) */
1058 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1060 // NEG_RCP_LO = -RCP_LO
1061 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1064 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1065 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1068 // Calculate the rounding error from the URECIP instruction
1069 // E = mulhu(ABS_RCP_LO, RCP)
1070 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1072 // RCP_A_E = RCP + E
1073 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1075 // RCP_S_E = RCP - E
1076 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1078 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1079 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1082 // Quotient = mulhu(Tmp0, Num)
1083 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1085 // Num_S_Remainder = Quotient * Den
1086 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1088 // Remainder = Num - Num_S_Remainder
1089 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1091 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1092 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1093 DAG.getConstant(-1, VT),
1094 DAG.getConstant(0, VT),
1096 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1097 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1099 DAG.getConstant(-1, VT),
1100 DAG.getConstant(0, VT),
1102 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1103 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1106 // Calculate Division result:
1108 // Quotient_A_One = Quotient + 1
1109 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1110 DAG.getConstant(1, VT));
1112 // Quotient_S_One = Quotient - 1
1113 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1114 DAG.getConstant(1, VT));
1116 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1117 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1118 Quotient, Quotient_A_One, ISD::SETEQ);
1120 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1121 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1122 Quotient_S_One, Div, ISD::SETEQ);
1124 // Calculate Rem result:
1126 // Remainder_S_Den = Remainder - Den
1127 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1129 // Remainder_A_Den = Remainder + Den
1130 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1132 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1133 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1134 Remainder, Remainder_S_Den, ISD::SETEQ);
1136 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1137 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1138 Remainder_A_Den, Rem, ISD::SETEQ);
1143 return DAG.getMergeValues(Ops, DL);
1146 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1147 SelectionDAG &DAG) const {
1148 SDValue S0 = Op.getOperand(0);
1150 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1153 // f32 uint_to_fp i64
1154 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1155 DAG.getConstant(0, MVT::i32));
1156 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1157 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1158 DAG.getConstant(1, MVT::i32));
1159 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1160 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1161 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1162 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1166 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1168 SelectionDAG &DAG) const {
1169 MVT VT = Op.getSimpleValueType();
1171 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1172 // Shift left by 'Shift' bits.
1173 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1174 // Signed shift Right by 'Shift' bits.
1175 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1178 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1179 SelectionDAG &DAG) const {
1180 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1181 MVT VT = Op.getSimpleValueType();
1182 MVT ScalarVT = VT.getScalarType();
1187 SDValue Src = Op.getOperand(0);
1190 // TODO: Don't scalarize on Evergreen?
1191 unsigned NElts = VT.getVectorNumElements();
1192 SmallVector<SDValue, 8> Args;
1193 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1195 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1196 for (unsigned I = 0; I < NElts; ++I)
1197 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1199 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1202 //===----------------------------------------------------------------------===//
1203 // Custom DAG optimizations
1204 //===----------------------------------------------------------------------===//
1206 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1207 APInt KnownZero, KnownOne;
1208 EVT VT = Op.getValueType();
1209 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
1211 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1214 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1215 EVT VT = Op.getValueType();
1217 // In order for this to be a signed 24-bit value, bit 23, must
1219 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1220 // as unsigned 24-bit values.
1221 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1224 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1226 SelectionDAG &DAG = DCI.DAG;
1227 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1228 EVT VT = Op.getValueType();
1230 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1231 APInt KnownZero, KnownOne;
1232 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1233 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1234 DCI.CommitTargetLoweringOpt(TLO);
1237 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1238 DAGCombinerInfo &DCI) const {
1239 SelectionDAG &DAG = DCI.DAG;
1242 switch(N->getOpcode()) {
1245 EVT VT = N->getValueType(0);
1246 SDValue N0 = N->getOperand(0);
1247 SDValue N1 = N->getOperand(1);
1250 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1251 if (VT.isVector() || VT.getSizeInBits() > 32)
1254 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1255 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1256 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1257 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1258 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1259 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1260 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1261 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1266 // We need to use sext even for MUL_U24, because MUL_U24 is used
1267 // for signed multiply of 8 and 16-bit types.
1268 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1272 case AMDGPUISD::MUL_I24:
1273 case AMDGPUISD::MUL_U24: {
1274 SDValue N0 = N->getOperand(0);
1275 SDValue N1 = N->getOperand(1);
1276 simplifyI24(N0, DCI);
1277 simplifyI24(N1, DCI);
1284 //===----------------------------------------------------------------------===//
1286 //===----------------------------------------------------------------------===//
1288 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1291 const SmallVectorImpl<ISD::InputArg> &Ins,
1292 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1294 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1295 if (Ins[i].ArgVT == Ins[i].VT) {
1296 OrigIns.push_back(Ins[i]);
1301 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1302 // Vector has been split into scalars.
1303 VT = Ins[i].ArgVT.getVectorElementType();
1304 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1305 Ins[i].ArgVT.getVectorElementType() !=
1306 Ins[i].VT.getVectorElementType()) {
1307 // Vector elements have been promoted
1310 // Vector has been spilt into smaller vectors.
1314 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1315 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1316 OrigIns.push_back(Arg);
1320 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1321 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1322 return CFP->isExactlyValue(1.0);
1324 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1325 return C->isAllOnesValue();
1330 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1331 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1332 return CFP->getValueAPF().isZero();
1334 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1335 return C->isNullValue();
1340 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1341 const TargetRegisterClass *RC,
1342 unsigned Reg, EVT VT) const {
1343 MachineFunction &MF = DAG.getMachineFunction();
1344 MachineRegisterInfo &MRI = MF.getRegInfo();
1345 unsigned VirtualRegister;
1346 if (!MRI.isLiveIn(Reg)) {
1347 VirtualRegister = MRI.createVirtualRegister(RC);
1348 MRI.addLiveIn(Reg, VirtualRegister);
1350 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1352 return DAG.getRegister(VirtualRegister, VT);
1355 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1357 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1359 default: return nullptr;
1361 NODE_NAME_CASE(CALL);
1362 NODE_NAME_CASE(UMUL);
1363 NODE_NAME_CASE(DIV_INF);
1364 NODE_NAME_CASE(RET_FLAG);
1365 NODE_NAME_CASE(BRANCH_COND);
1368 NODE_NAME_CASE(DWORDADDR)
1369 NODE_NAME_CASE(FRACT)
1370 NODE_NAME_CASE(FMAX)
1371 NODE_NAME_CASE(SMAX)
1372 NODE_NAME_CASE(UMAX)
1373 NODE_NAME_CASE(FMIN)
1374 NODE_NAME_CASE(SMIN)
1375 NODE_NAME_CASE(UMIN)
1376 NODE_NAME_CASE(BFE_U32)
1377 NODE_NAME_CASE(BFE_I32)
1380 NODE_NAME_CASE(MUL_U24)
1381 NODE_NAME_CASE(MUL_I24)
1382 NODE_NAME_CASE(URECIP)
1383 NODE_NAME_CASE(DOT4)
1384 NODE_NAME_CASE(EXPORT)
1385 NODE_NAME_CASE(CONST_ADDRESS)
1386 NODE_NAME_CASE(REGISTER_LOAD)
1387 NODE_NAME_CASE(REGISTER_STORE)
1388 NODE_NAME_CASE(LOAD_CONSTANT)
1389 NODE_NAME_CASE(LOAD_INPUT)
1390 NODE_NAME_CASE(SAMPLE)
1391 NODE_NAME_CASE(SAMPLEB)
1392 NODE_NAME_CASE(SAMPLED)
1393 NODE_NAME_CASE(SAMPLEL)
1394 NODE_NAME_CASE(STORE_MSKOR)
1395 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1399 static void computeMaskedBitsForMinMax(const SDValue Op0,
1403 const SelectionDAG &DAG,
1405 APInt Op0Zero, Op0One;
1406 APInt Op1Zero, Op1One;
1407 DAG.ComputeMaskedBits(Op0, Op0Zero, Op0One, Depth);
1408 DAG.ComputeMaskedBits(Op1, Op1Zero, Op1One, Depth);
1410 KnownZero = Op0Zero & Op1Zero;
1411 KnownOne = Op0One & Op1One;
1414 void AMDGPUTargetLowering::computeMaskedBitsForTargetNode(
1418 const SelectionDAG &DAG,
1419 unsigned Depth) const {
1421 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1422 unsigned Opc = Op.getOpcode();
1424 case ISD::INTRINSIC_WO_CHAIN: {
1425 // FIXME: The intrinsic should just use the node.
1426 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1427 case AMDGPUIntrinsic::AMDGPU_imax:
1428 case AMDGPUIntrinsic::AMDGPU_umax:
1429 case AMDGPUIntrinsic::AMDGPU_imin:
1430 case AMDGPUIntrinsic::AMDGPU_umin:
1431 computeMaskedBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1432 KnownZero, KnownOne, DAG, Depth);
1440 case AMDGPUISD::SMAX:
1441 case AMDGPUISD::UMAX:
1442 case AMDGPUISD::SMIN:
1443 case AMDGPUISD::UMIN:
1444 computeMaskedBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1445 KnownZero, KnownOne, DAG, Depth);