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 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
134 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
136 // Custom lowering of vector stores is required for local address space
138 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
139 // XXX: Native v2i32 local address space stores are possible, but not
140 // currently implemented.
141 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
143 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
144 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
145 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
147 // XXX: This can be change to Custom, once ExpandVectorStores can
148 // handle 64-bit stores.
149 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
151 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
152 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
153 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
154 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
155 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
158 setOperationAction(ISD::LOAD, MVT::f32, Promote);
159 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
161 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
162 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
164 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
165 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
167 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
168 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
170 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
171 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
173 setOperationAction(ISD::LOAD, MVT::f64, Promote);
174 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
176 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
177 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
179 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
180 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
181 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
182 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
183 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
184 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
185 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
186 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
187 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
188 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
190 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
191 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
192 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
193 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
194 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
195 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
196 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
197 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
198 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
199 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
200 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
201 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
203 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
205 setOperationAction(ISD::FNEG, MVT::v2f32, Expand);
206 setOperationAction(ISD::FNEG, MVT::v4f32, Expand);
208 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
210 setOperationAction(ISD::MUL, MVT::i64, Expand);
211 setOperationAction(ISD::SUB, MVT::i64, Expand);
213 setOperationAction(ISD::UDIV, MVT::i32, Expand);
214 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
215 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
216 setOperationAction(ISD::UREM, MVT::i32, Expand);
217 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
218 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
220 static const MVT::SimpleValueType IntTypes[] = {
221 MVT::v2i32, MVT::v4i32
223 const size_t NumIntTypes = array_lengthof(IntTypes);
225 for (unsigned int x = 0; x < NumIntTypes; ++x) {
226 MVT::SimpleValueType VT = IntTypes[x];
227 //Expand the following operations for the current type by default
228 setOperationAction(ISD::ADD, VT, Expand);
229 setOperationAction(ISD::AND, VT, Expand);
230 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
231 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
232 setOperationAction(ISD::MUL, VT, Expand);
233 setOperationAction(ISD::OR, VT, Expand);
234 setOperationAction(ISD::SHL, VT, Expand);
235 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
236 setOperationAction(ISD::SRL, VT, Expand);
237 setOperationAction(ISD::SRA, VT, Expand);
238 setOperationAction(ISD::SUB, VT, Expand);
239 setOperationAction(ISD::UDIV, VT, Expand);
240 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
241 setOperationAction(ISD::UREM, VT, Expand);
242 setOperationAction(ISD::SELECT, VT, Expand);
243 setOperationAction(ISD::VSELECT, VT, Expand);
244 setOperationAction(ISD::XOR, VT, Expand);
247 static const MVT::SimpleValueType FloatTypes[] = {
248 MVT::v2f32, MVT::v4f32
250 const size_t NumFloatTypes = array_lengthof(FloatTypes);
252 for (unsigned int x = 0; x < NumFloatTypes; ++x) {
253 MVT::SimpleValueType VT = FloatTypes[x];
254 setOperationAction(ISD::FABS, VT, Expand);
255 setOperationAction(ISD::FADD, VT, Expand);
256 setOperationAction(ISD::FCOS, VT, Expand);
257 setOperationAction(ISD::FDIV, VT, Expand);
258 setOperationAction(ISD::FPOW, VT, Expand);
259 setOperationAction(ISD::FFLOOR, VT, Expand);
260 setOperationAction(ISD::FTRUNC, VT, Expand);
261 setOperationAction(ISD::FMUL, VT, Expand);
262 setOperationAction(ISD::FRINT, VT, Expand);
263 setOperationAction(ISD::FSQRT, VT, Expand);
264 setOperationAction(ISD::FSIN, VT, Expand);
265 setOperationAction(ISD::FSUB, VT, Expand);
266 setOperationAction(ISD::SELECT, VT, Expand);
269 setTargetDAGCombine(ISD::MUL);
272 //===----------------------------------------------------------------------===//
273 // Target Information
274 //===----------------------------------------------------------------------===//
276 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
280 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
282 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
285 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
286 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
288 return ((LScalarSize <= CastScalarSize) ||
289 (CastScalarSize >= 32) ||
293 //===---------------------------------------------------------------------===//
295 //===---------------------------------------------------------------------===//
297 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
298 assert(VT.isFloatingPoint());
299 return VT == MVT::f32;
302 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
303 assert(VT.isFloatingPoint());
304 return VT == MVT::f32;
307 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
308 // Truncate is just accessing a subregister.
309 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
312 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
313 // Truncate is just accessing a subregister.
314 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
315 (Dest->getPrimitiveSizeInBits() % 32 == 0);
318 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
319 const DataLayout *DL = getDataLayout();
320 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
321 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
323 return SrcSize == 32 && DestSize == 64;
326 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
327 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
328 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
329 // this will enable reducing 64-bit operations the 32-bit, which is always
331 return Src == MVT::i32 && Dest == MVT::i64;
334 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
335 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
336 // limited number of native 64-bit operations. Shrinking an operation to fit
337 // in a single 32-bit register should always be helpful. As currently used,
338 // this is much less general than the name suggests, and is only used in
339 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
340 // not profitable, and may actually be harmful.
341 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
344 //===---------------------------------------------------------------------===//
345 // TargetLowering Callbacks
346 //===---------------------------------------------------------------------===//
348 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
349 const SmallVectorImpl<ISD::InputArg> &Ins) const {
351 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
354 SDValue AMDGPUTargetLowering::LowerReturn(
356 CallingConv::ID CallConv,
358 const SmallVectorImpl<ISD::OutputArg> &Outs,
359 const SmallVectorImpl<SDValue> &OutVals,
360 SDLoc DL, SelectionDAG &DAG) const {
361 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
364 //===---------------------------------------------------------------------===//
365 // Target specific lowering
366 //===---------------------------------------------------------------------===//
368 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
369 SmallVectorImpl<SDValue> &InVals) const {
370 SDValue Callee = CLI.Callee;
371 SelectionDAG &DAG = CLI.DAG;
373 const Function &Fn = *DAG.getMachineFunction().getFunction();
375 StringRef FuncName("<unknown>");
377 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
378 FuncName = G->getSymbol();
379 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
380 FuncName = G->getGlobal()->getName();
382 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
383 DAG.getContext()->diagnose(NoCalls);
387 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
389 switch (Op.getOpcode()) {
391 Op.getNode()->dump();
392 llvm_unreachable("Custom lowering code for this"
393 "instruction is not implemented yet!");
395 // AMDIL DAG lowering
396 case ISD::SDIV: return LowerSDIV(Op, DAG);
397 case ISD::SREM: return LowerSREM(Op, DAG);
398 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
399 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
400 // AMDGPU DAG lowering
401 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
402 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
403 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
404 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
405 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
406 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
411 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
412 SmallVectorImpl<SDValue> &Results,
413 SelectionDAG &DAG) const {
414 switch (N->getOpcode()) {
415 case ISD::SIGN_EXTEND_INREG:
416 // Different parts of legalization seem to interpret which type of
417 // sign_extend_inreg is the one to check for custom lowering. The extended
418 // from type is what really matters, but some places check for custom
419 // lowering of the result type. This results in trying to use
420 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
421 // nothing here and let the illegal result integer be handled normally.
424 SDValue Op = SDValue(N, 0);
426 EVT VT = Op.getValueType();
427 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
428 N->getOperand(0), N->getOperand(1));
429 Results.push_back(UDIVREM);
433 SDValue Op = SDValue(N, 0);
435 EVT VT = Op.getValueType();
436 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
437 N->getOperand(0), N->getOperand(1));
438 Results.push_back(UDIVREM.getValue(1));
442 SDValue Op = SDValue(N, 0);
444 EVT VT = Op.getValueType();
445 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
447 SDValue one = DAG.getConstant(1, HalfVT);
448 SDValue zero = DAG.getConstant(0, HalfVT);
451 SDValue LHS = N->getOperand(0);
452 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
453 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
455 SDValue RHS = N->getOperand(1);
456 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
457 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
459 // Get Speculative values
460 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
461 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
463 SDValue REM_Hi = zero;
464 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
466 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
467 SDValue DIV_Lo = zero;
469 const unsigned halfBitWidth = HalfVT.getSizeInBits();
471 for (unsigned i = 0; i < halfBitWidth; ++i) {
472 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
473 // Get Value of high bit
475 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
476 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
478 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
479 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
482 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
483 DAG.getConstant(halfBitWidth - 1, HalfVT));
484 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
485 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
487 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
488 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
491 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
493 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
494 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
496 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
500 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
502 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
503 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
504 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
507 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
508 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
509 Results.push_back(DIV);
510 Results.push_back(REM);
518 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
519 const GlobalValue *GV,
520 const SDValue &InitPtr,
522 SelectionDAG &DAG) const {
523 const DataLayout *TD = getTargetMachine().getDataLayout();
525 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
526 EVT VT = EVT::getEVT(CI->getType());
527 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
528 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
529 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
530 TD->getPrefTypeAlignment(CI->getType()));
531 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
532 EVT VT = EVT::getEVT(CFP->getType());
533 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
534 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
535 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
536 TD->getPrefTypeAlignment(CFP->getType()));
537 } else if (Init->getType()->isAggregateType()) {
538 EVT PtrVT = InitPtr.getValueType();
539 unsigned NumElements = Init->getType()->getArrayNumElements();
540 SmallVector<SDValue, 8> Chains;
541 for (unsigned i = 0; i < NumElements; ++i) {
542 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
543 Init->getType()->getArrayElementType()), PtrVT);
544 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
545 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
546 GV, Ptr, Chain, DAG));
548 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
551 llvm_unreachable("Unhandled constant initializer");
555 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
557 SelectionDAG &DAG) const {
559 const DataLayout *TD = getTargetMachine().getDataLayout();
560 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
561 const GlobalValue *GV = G->getGlobal();
563 switch (G->getAddressSpace()) {
564 default: llvm_unreachable("Global Address lowering not implemented for this "
566 case AMDGPUAS::LOCAL_ADDRESS: {
567 // XXX: What does the value of G->getOffset() mean?
568 assert(G->getOffset() == 0 &&
569 "Do not know what to do with an non-zero offset");
572 if (MFI->LocalMemoryObjects.count(GV) == 0) {
573 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
574 Offset = MFI->LDSSize;
575 MFI->LocalMemoryObjects[GV] = Offset;
576 // XXX: Account for alignment?
577 MFI->LDSSize += Size;
579 Offset = MFI->LocalMemoryObjects[GV];
582 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
584 case AMDGPUAS::CONSTANT_ADDRESS: {
585 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
586 Type *EltType = GV->getType()->getElementType();
587 unsigned Size = TD->getTypeAllocSize(EltType);
588 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
590 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
591 const Constant *Init = Var->getInitializer();
592 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
593 SDValue InitPtr = DAG.getFrameIndex(FI,
594 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
595 SmallVector<SDNode*, 8> WorkList;
597 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
598 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
599 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
601 WorkList.push_back(*I);
603 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
604 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
605 E = WorkList.end(); I != E; ++I) {
606 SmallVector<SDValue, 8> Ops;
607 Ops.push_back(Chain);
608 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
609 Ops.push_back((*I)->getOperand(i));
611 DAG.UpdateNodeOperands(*I, Ops);
613 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
614 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
619 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
620 SelectionDAG &DAG) const {
621 SmallVector<SDValue, 8> Args;
622 SDValue A = Op.getOperand(0);
623 SDValue B = Op.getOperand(1);
625 DAG.ExtractVectorElements(A, Args);
626 DAG.ExtractVectorElements(B, Args);
628 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
631 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
632 SelectionDAG &DAG) const {
634 SmallVector<SDValue, 8> Args;
635 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
636 EVT VT = Op.getValueType();
637 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
638 VT.getVectorNumElements());
640 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
643 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
644 SelectionDAG &DAG) const {
646 MachineFunction &MF = DAG.getMachineFunction();
647 const AMDGPUFrameLowering *TFL =
648 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
650 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
653 unsigned FrameIndex = FIN->getIndex();
654 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
655 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
659 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
660 SelectionDAG &DAG) const {
661 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
663 EVT VT = Op.getValueType();
665 switch (IntrinsicID) {
667 case AMDGPUIntrinsic::AMDIL_abs:
668 return LowerIntrinsicIABS(Op, DAG);
669 case AMDGPUIntrinsic::AMDIL_exp:
670 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
671 case AMDGPUIntrinsic::AMDGPU_lrp:
672 return LowerIntrinsicLRP(Op, DAG);
673 case AMDGPUIntrinsic::AMDIL_fraction:
674 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
675 case AMDGPUIntrinsic::AMDIL_max:
676 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
678 case AMDGPUIntrinsic::AMDGPU_imax:
679 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
681 case AMDGPUIntrinsic::AMDGPU_umax:
682 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
684 case AMDGPUIntrinsic::AMDIL_min:
685 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
687 case AMDGPUIntrinsic::AMDGPU_imin:
688 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
690 case AMDGPUIntrinsic::AMDGPU_umin:
691 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
694 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
695 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
700 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
701 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
706 case AMDGPUIntrinsic::AMDGPU_bfi:
707 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
712 case AMDGPUIntrinsic::AMDGPU_bfm:
713 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
717 case AMDGPUIntrinsic::AMDIL_round_nearest:
718 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
722 ///IABS(a) = SMAX(sub(0, a), a)
723 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
724 SelectionDAG &DAG) const {
727 EVT VT = Op.getValueType();
728 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
731 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
734 /// Linear Interpolation
735 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
736 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
737 SelectionDAG &DAG) const {
739 EVT VT = Op.getValueType();
740 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
741 DAG.getConstantFP(1.0f, MVT::f32),
743 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
745 return DAG.getNode(ISD::FADD, DL, VT,
746 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
750 /// \brief Generate Min/Max node
751 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
752 SelectionDAG &DAG) const {
754 EVT VT = Op.getValueType();
756 SDValue LHS = Op.getOperand(0);
757 SDValue RHS = Op.getOperand(1);
758 SDValue True = Op.getOperand(2);
759 SDValue False = Op.getOperand(3);
760 SDValue CC = Op.getOperand(4);
762 if (VT != MVT::f32 ||
763 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
767 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
781 llvm_unreachable("Operation should already be optimised!");
789 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
791 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
800 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
802 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
804 case ISD::SETCC_INVALID:
805 llvm_unreachable("Invalid setcc condcode!");
810 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
811 SelectionDAG &DAG) const {
812 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
813 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
814 EVT EltVT = Op.getValueType().getVectorElementType();
815 EVT PtrVT = Load->getBasePtr().getValueType();
816 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
817 SmallVector<SDValue, 8> Loads;
820 for (unsigned i = 0, e = NumElts; i != e; ++i) {
821 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
822 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
823 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
824 Load->getChain(), Ptr,
825 MachinePointerInfo(Load->getMemOperand()->getValue()),
826 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
827 Load->getAlignment()));
829 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
832 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
833 SelectionDAG &DAG) const {
834 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
835 EVT MemVT = Store->getMemoryVT();
836 unsigned MemBits = MemVT.getSizeInBits();
838 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
839 // truncating store into an i32 store.
840 // XXX: We could also handle optimize other vector bitwidths.
841 if (!MemVT.isVector() || MemBits > 32) {
846 SDValue Value = Store->getValue();
847 EVT VT = Value.getValueType();
848 EVT ElemVT = VT.getVectorElementType();
849 SDValue Ptr = Store->getBasePtr();
850 EVT MemEltVT = MemVT.getVectorElementType();
851 unsigned MemEltBits = MemEltVT.getSizeInBits();
852 unsigned MemNumElements = MemVT.getVectorNumElements();
853 unsigned PackedSize = MemVT.getStoreSizeInBits();
854 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
856 assert(Value.getValueType().getScalarSizeInBits() >= 32);
859 for (unsigned i = 0; i < MemNumElements; ++i) {
860 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
861 DAG.getConstant(i, MVT::i32));
862 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
863 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
865 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
866 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
871 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
875 if (PackedSize < 32) {
876 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
877 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
878 Store->getMemOperand()->getPointerInfo(),
880 Store->isNonTemporal(), Store->isVolatile(),
881 Store->getAlignment());
884 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
885 Store->getMemOperand()->getPointerInfo(),
886 Store->isVolatile(), Store->isNonTemporal(),
887 Store->getAlignment());
890 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
891 SelectionDAG &DAG) const {
892 StoreSDNode *Store = cast<StoreSDNode>(Op);
893 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
894 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
895 EVT PtrVT = Store->getBasePtr().getValueType();
896 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
899 SmallVector<SDValue, 8> Chains;
901 for (unsigned i = 0, e = NumElts; i != e; ++i) {
902 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
903 Store->getValue(), DAG.getConstant(i, MVT::i32));
904 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
906 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
908 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
909 MachinePointerInfo(Store->getMemOperand()->getValue()),
910 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
911 Store->getAlignment()));
913 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
916 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
918 LoadSDNode *Load = cast<LoadSDNode>(Op);
919 ISD::LoadExtType ExtType = Load->getExtensionType();
920 EVT VT = Op.getValueType();
921 EVT MemVT = Load->getMemoryVT();
923 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
924 // We can do the extload to 32-bits, and then need to separately extend to
927 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
931 Load->getMemOperand());
932 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
935 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
936 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
937 // FIXME: Copied from PPC
938 // First, load into 32 bits, then truncate to 1 bit.
940 SDValue Chain = Load->getChain();
941 SDValue BasePtr = Load->getBasePtr();
942 MachineMemOperand *MMO = Load->getMemOperand();
944 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
945 BasePtr, MVT::i8, MMO);
946 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
949 // Lower loads constant address space global variable loads
950 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
952 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
954 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
955 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
956 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
957 DAG.getConstant(2, MVT::i32));
958 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
959 Load->getChain(), Ptr,
960 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
963 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
964 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
968 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
969 DAG.getConstant(2, MVT::i32));
970 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
971 Load->getChain(), Ptr,
972 DAG.getTargetConstant(0, MVT::i32),
974 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
976 DAG.getConstant(0x3, MVT::i32));
977 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
978 DAG.getConstant(3, MVT::i32));
980 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
982 EVT MemEltVT = MemVT.getScalarType();
983 if (ExtType == ISD::SEXTLOAD) {
984 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
985 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
988 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
991 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
993 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
994 if (Result.getNode()) {
998 StoreSDNode *Store = cast<StoreSDNode>(Op);
999 SDValue Chain = Store->getChain();
1000 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1001 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1002 Store->getValue().getValueType().isVector()) {
1003 return SplitVectorStore(Op, DAG);
1006 EVT MemVT = Store->getMemoryVT();
1007 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1008 MemVT.bitsLT(MVT::i32)) {
1010 if (Store->getMemoryVT() == MVT::i8) {
1012 } else if (Store->getMemoryVT() == MVT::i16) {
1015 SDValue BasePtr = Store->getBasePtr();
1016 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1017 DAG.getConstant(2, MVT::i32));
1018 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1019 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1021 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1022 DAG.getConstant(0x3, MVT::i32));
1024 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1025 DAG.getConstant(3, MVT::i32));
1027 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1030 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1032 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1033 MaskedValue, ShiftAmt);
1035 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1037 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1038 DAG.getConstant(0xffffffff, MVT::i32));
1039 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1041 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1042 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1043 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1048 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1049 SelectionDAG &DAG) const {
1051 EVT VT = Op.getValueType();
1053 SDValue Num = Op.getOperand(0);
1054 SDValue Den = Op.getOperand(1);
1056 // RCP = URECIP(Den) = 2^32 / Den + e
1057 // e is rounding error.
1058 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1060 // RCP_LO = umulo(RCP, Den) */
1061 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1063 // RCP_HI = mulhu (RCP, Den) */
1064 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1066 // NEG_RCP_LO = -RCP_LO
1067 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1070 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1071 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1074 // Calculate the rounding error from the URECIP instruction
1075 // E = mulhu(ABS_RCP_LO, RCP)
1076 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1078 // RCP_A_E = RCP + E
1079 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1081 // RCP_S_E = RCP - E
1082 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1084 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1085 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1088 // Quotient = mulhu(Tmp0, Num)
1089 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1091 // Num_S_Remainder = Quotient * Den
1092 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1094 // Remainder = Num - Num_S_Remainder
1095 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1097 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1098 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1099 DAG.getConstant(-1, VT),
1100 DAG.getConstant(0, VT),
1102 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1103 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1105 DAG.getConstant(-1, VT),
1106 DAG.getConstant(0, VT),
1108 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1109 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1112 // Calculate Division result:
1114 // Quotient_A_One = Quotient + 1
1115 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1116 DAG.getConstant(1, VT));
1118 // Quotient_S_One = Quotient - 1
1119 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1120 DAG.getConstant(1, VT));
1122 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1123 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1124 Quotient, Quotient_A_One, ISD::SETEQ);
1126 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1127 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1128 Quotient_S_One, Div, ISD::SETEQ);
1130 // Calculate Rem result:
1132 // Remainder_S_Den = Remainder - Den
1133 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1135 // Remainder_A_Den = Remainder + Den
1136 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1138 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1139 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1140 Remainder, Remainder_S_Den, ISD::SETEQ);
1142 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1143 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1144 Remainder_A_Den, Rem, ISD::SETEQ);
1149 return DAG.getMergeValues(Ops, DL);
1152 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1153 SelectionDAG &DAG) const {
1154 SDValue S0 = Op.getOperand(0);
1156 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1159 // f32 uint_to_fp i64
1160 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1161 DAG.getConstant(0, MVT::i32));
1162 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1163 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1164 DAG.getConstant(1, MVT::i32));
1165 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1166 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1167 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1168 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1172 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1174 SelectionDAG &DAG) const {
1175 MVT VT = Op.getSimpleValueType();
1177 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1178 // Shift left by 'Shift' bits.
1179 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1180 // Signed shift Right by 'Shift' bits.
1181 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1184 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1185 SelectionDAG &DAG) const {
1186 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1187 MVT VT = Op.getSimpleValueType();
1188 MVT ScalarVT = VT.getScalarType();
1193 SDValue Src = Op.getOperand(0);
1196 // TODO: Don't scalarize on Evergreen?
1197 unsigned NElts = VT.getVectorNumElements();
1198 SmallVector<SDValue, 8> Args;
1199 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1201 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1202 for (unsigned I = 0; I < NElts; ++I)
1203 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1205 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1208 //===----------------------------------------------------------------------===//
1209 // Custom DAG optimizations
1210 //===----------------------------------------------------------------------===//
1212 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1213 APInt KnownZero, KnownOne;
1214 EVT VT = Op.getValueType();
1215 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
1217 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1220 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1221 EVT VT = Op.getValueType();
1223 // In order for this to be a signed 24-bit value, bit 23, must
1225 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1226 // as unsigned 24-bit values.
1227 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1230 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1232 SelectionDAG &DAG = DCI.DAG;
1233 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1234 EVT VT = Op.getValueType();
1236 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1237 APInt KnownZero, KnownOne;
1238 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1239 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1240 DCI.CommitTargetLoweringOpt(TLO);
1243 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1244 DAGCombinerInfo &DCI) const {
1245 SelectionDAG &DAG = DCI.DAG;
1248 switch(N->getOpcode()) {
1251 EVT VT = N->getValueType(0);
1252 SDValue N0 = N->getOperand(0);
1253 SDValue N1 = N->getOperand(1);
1256 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1257 if (VT.isVector() || VT.getSizeInBits() > 32)
1260 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1261 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1262 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1263 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1264 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1265 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1266 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1267 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1272 // We need to use sext even for MUL_U24, because MUL_U24 is used
1273 // for signed multiply of 8 and 16-bit types.
1274 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1278 case AMDGPUISD::MUL_I24:
1279 case AMDGPUISD::MUL_U24: {
1280 SDValue N0 = N->getOperand(0);
1281 SDValue N1 = N->getOperand(1);
1282 simplifyI24(N0, DCI);
1283 simplifyI24(N1, DCI);
1290 //===----------------------------------------------------------------------===//
1292 //===----------------------------------------------------------------------===//
1294 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1297 const SmallVectorImpl<ISD::InputArg> &Ins,
1298 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1300 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1301 if (Ins[i].ArgVT == Ins[i].VT) {
1302 OrigIns.push_back(Ins[i]);
1307 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1308 // Vector has been split into scalars.
1309 VT = Ins[i].ArgVT.getVectorElementType();
1310 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1311 Ins[i].ArgVT.getVectorElementType() !=
1312 Ins[i].VT.getVectorElementType()) {
1313 // Vector elements have been promoted
1316 // Vector has been spilt into smaller vectors.
1320 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1321 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1322 OrigIns.push_back(Arg);
1326 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1327 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1328 return CFP->isExactlyValue(1.0);
1330 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1331 return C->isAllOnesValue();
1336 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1337 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1338 return CFP->getValueAPF().isZero();
1340 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1341 return C->isNullValue();
1346 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1347 const TargetRegisterClass *RC,
1348 unsigned Reg, EVT VT) const {
1349 MachineFunction &MF = DAG.getMachineFunction();
1350 MachineRegisterInfo &MRI = MF.getRegInfo();
1351 unsigned VirtualRegister;
1352 if (!MRI.isLiveIn(Reg)) {
1353 VirtualRegister = MRI.createVirtualRegister(RC);
1354 MRI.addLiveIn(Reg, VirtualRegister);
1356 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1358 return DAG.getRegister(VirtualRegister, VT);
1361 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1363 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1365 default: return nullptr;
1367 NODE_NAME_CASE(CALL);
1368 NODE_NAME_CASE(UMUL);
1369 NODE_NAME_CASE(DIV_INF);
1370 NODE_NAME_CASE(RET_FLAG);
1371 NODE_NAME_CASE(BRANCH_COND);
1374 NODE_NAME_CASE(DWORDADDR)
1375 NODE_NAME_CASE(FRACT)
1376 NODE_NAME_CASE(FMAX)
1377 NODE_NAME_CASE(SMAX)
1378 NODE_NAME_CASE(UMAX)
1379 NODE_NAME_CASE(FMIN)
1380 NODE_NAME_CASE(SMIN)
1381 NODE_NAME_CASE(UMIN)
1382 NODE_NAME_CASE(BFE_U32)
1383 NODE_NAME_CASE(BFE_I32)
1386 NODE_NAME_CASE(MUL_U24)
1387 NODE_NAME_CASE(MUL_I24)
1388 NODE_NAME_CASE(URECIP)
1389 NODE_NAME_CASE(DOT4)
1390 NODE_NAME_CASE(EXPORT)
1391 NODE_NAME_CASE(CONST_ADDRESS)
1392 NODE_NAME_CASE(REGISTER_LOAD)
1393 NODE_NAME_CASE(REGISTER_STORE)
1394 NODE_NAME_CASE(LOAD_CONSTANT)
1395 NODE_NAME_CASE(LOAD_INPUT)
1396 NODE_NAME_CASE(SAMPLE)
1397 NODE_NAME_CASE(SAMPLEB)
1398 NODE_NAME_CASE(SAMPLED)
1399 NODE_NAME_CASE(SAMPLEL)
1400 NODE_NAME_CASE(STORE_MSKOR)
1401 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1405 static void computeMaskedBitsForMinMax(const SDValue Op0,
1409 const SelectionDAG &DAG,
1411 APInt Op0Zero, Op0One;
1412 APInt Op1Zero, Op1One;
1413 DAG.ComputeMaskedBits(Op0, Op0Zero, Op0One, Depth);
1414 DAG.ComputeMaskedBits(Op1, Op1Zero, Op1One, Depth);
1416 KnownZero = Op0Zero & Op1Zero;
1417 KnownOne = Op0One & Op1One;
1420 void AMDGPUTargetLowering::computeMaskedBitsForTargetNode(
1424 const SelectionDAG &DAG,
1425 unsigned Depth) const {
1427 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1428 unsigned Opc = Op.getOpcode();
1430 case ISD::INTRINSIC_WO_CHAIN: {
1431 // FIXME: The intrinsic should just use the node.
1432 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1433 case AMDGPUIntrinsic::AMDGPU_imax:
1434 case AMDGPUIntrinsic::AMDGPU_umax:
1435 case AMDGPUIntrinsic::AMDGPU_imin:
1436 case AMDGPUIntrinsic::AMDGPU_umin:
1437 computeMaskedBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1438 KnownZero, KnownOne, DAG, Depth);
1446 case AMDGPUISD::SMAX:
1447 case AMDGPUISD::UMAX:
1448 case AMDGPUISD::SMIN:
1449 case AMDGPUISD::UMIN:
1450 computeMaskedBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1451 KnownZero, KnownOne, DAG, Depth);