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"
33 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
34 CCValAssign::LocInfo LocInfo,
35 ISD::ArgFlagsTy ArgFlags, CCState &State) {
36 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
37 ArgFlags.getOrigAlign());
38 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
43 #include "AMDGPUGenCallingConv.inc"
45 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
46 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
48 // Initialize target lowering borrowed from AMDIL
51 // We need to custom lower some of the intrinsics
52 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
54 // Library functions. These default to Expand, but we have instructions
56 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
57 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
58 setOperationAction(ISD::FPOW, MVT::f32, Legal);
59 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
60 setOperationAction(ISD::FABS, MVT::f32, Legal);
61 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
62 setOperationAction(ISD::FRINT, MVT::f32, Legal);
63 setOperationAction(ISD::FROUND, MVT::f32, Legal);
64 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
66 // The hardware supports ROTR, but not ROTL
67 setOperationAction(ISD::ROTL, MVT::i32, Expand);
69 // Lower floating point store/load to integer store/load to reduce the number
70 // of patterns in tablegen.
71 setOperationAction(ISD::STORE, MVT::f32, Promote);
72 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
74 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
75 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
77 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
78 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
80 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
81 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
83 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
84 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
86 setOperationAction(ISD::STORE, MVT::f64, Promote);
87 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
89 // Custom lowering of vector stores is required for local address space
91 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
92 // XXX: Native v2i32 local address space stores are possible, but not
93 // currently implemented.
94 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
96 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
97 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
98 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
99 // XXX: This can be change to Custom, once ExpandVectorStores can
100 // handle 64-bit stores.
101 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
103 setOperationAction(ISD::LOAD, MVT::f32, Promote);
104 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
106 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
107 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
109 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
110 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
112 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
113 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
115 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
116 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
118 setOperationAction(ISD::LOAD, MVT::f64, Promote);
119 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
121 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
122 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
123 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
124 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
126 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
127 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
128 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
129 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
130 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
131 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
132 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
133 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
134 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
135 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
136 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
137 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
139 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
141 setOperationAction(ISD::FNEG, MVT::v2f32, Expand);
142 setOperationAction(ISD::FNEG, MVT::v4f32, Expand);
144 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
146 setOperationAction(ISD::MUL, MVT::i64, Expand);
148 setOperationAction(ISD::UDIV, MVT::i32, Expand);
149 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
150 setOperationAction(ISD::UREM, MVT::i32, Expand);
151 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
152 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
154 static const MVT::SimpleValueType IntTypes[] = {
155 MVT::v2i32, MVT::v4i32
157 const size_t NumIntTypes = array_lengthof(IntTypes);
159 for (unsigned int x = 0; x < NumIntTypes; ++x) {
160 MVT::SimpleValueType VT = IntTypes[x];
161 //Expand the following operations for the current type by default
162 setOperationAction(ISD::ADD, VT, Expand);
163 setOperationAction(ISD::AND, VT, Expand);
164 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
165 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
166 setOperationAction(ISD::MUL, VT, Expand);
167 setOperationAction(ISD::OR, VT, Expand);
168 setOperationAction(ISD::SHL, VT, Expand);
169 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
170 setOperationAction(ISD::SRL, VT, Expand);
171 setOperationAction(ISD::SRA, VT, Expand);
172 setOperationAction(ISD::SUB, VT, Expand);
173 setOperationAction(ISD::UDIV, VT, Expand);
174 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
175 setOperationAction(ISD::UREM, VT, Expand);
176 setOperationAction(ISD::VSELECT, VT, Expand);
177 setOperationAction(ISD::XOR, VT, Expand);
180 static const MVT::SimpleValueType FloatTypes[] = {
181 MVT::v2f32, MVT::v4f32
183 const size_t NumFloatTypes = array_lengthof(FloatTypes);
185 for (unsigned int x = 0; x < NumFloatTypes; ++x) {
186 MVT::SimpleValueType VT = FloatTypes[x];
187 setOperationAction(ISD::FABS, VT, Expand);
188 setOperationAction(ISD::FADD, VT, Expand);
189 setOperationAction(ISD::FDIV, VT, Expand);
190 setOperationAction(ISD::FPOW, VT, Expand);
191 setOperationAction(ISD::FFLOOR, VT, Expand);
192 setOperationAction(ISD::FTRUNC, VT, Expand);
193 setOperationAction(ISD::FMUL, VT, Expand);
194 setOperationAction(ISD::FRINT, VT, Expand);
195 setOperationAction(ISD::FSQRT, VT, Expand);
196 setOperationAction(ISD::FSUB, VT, Expand);
200 //===----------------------------------------------------------------------===//
201 // Target Information
202 //===----------------------------------------------------------------------===//
204 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
208 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
210 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
213 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
214 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
216 return ((LScalarSize <= CastScalarSize) ||
217 (CastScalarSize >= 32) ||
221 //===---------------------------------------------------------------------===//
223 //===---------------------------------------------------------------------===//
225 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
226 assert(VT.isFloatingPoint());
227 return VT == MVT::f32;
230 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
231 assert(VT.isFloatingPoint());
232 return VT == MVT::f32;
235 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
236 // Truncate is just accessing a subregister.
237 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
240 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
241 // Truncate is just accessing a subregister.
242 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
243 (Dest->getPrimitiveSizeInBits() % 32 == 0);
246 //===---------------------------------------------------------------------===//
247 // TargetLowering Callbacks
248 //===---------------------------------------------------------------------===//
250 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
251 const SmallVectorImpl<ISD::InputArg> &Ins) const {
253 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
256 SDValue AMDGPUTargetLowering::LowerReturn(
258 CallingConv::ID CallConv,
260 const SmallVectorImpl<ISD::OutputArg> &Outs,
261 const SmallVectorImpl<SDValue> &OutVals,
262 SDLoc DL, SelectionDAG &DAG) const {
263 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
266 //===---------------------------------------------------------------------===//
267 // Target specific lowering
268 //===---------------------------------------------------------------------===//
270 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
272 switch (Op.getOpcode()) {
274 Op.getNode()->dump();
275 llvm_unreachable("Custom lowering code for this"
276 "instruction is not implemented yet!");
278 // AMDIL DAG lowering
279 case ISD::SDIV: return LowerSDIV(Op, DAG);
280 case ISD::SREM: return LowerSREM(Op, DAG);
281 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
282 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
283 // AMDGPU DAG lowering
284 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
285 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
286 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
287 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
288 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
289 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
294 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
295 const GlobalValue *GV,
296 const SDValue &InitPtr,
298 SelectionDAG &DAG) const {
299 const DataLayout *TD = getTargetMachine().getDataLayout();
301 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
302 EVT VT = EVT::getEVT(CI->getType());
303 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
304 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
305 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
306 TD->getPrefTypeAlignment(CI->getType()));
307 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
308 EVT VT = EVT::getEVT(CFP->getType());
309 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
310 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
311 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
312 TD->getPrefTypeAlignment(CFP->getType()));
313 } else if (Init->getType()->isAggregateType()) {
314 EVT PtrVT = InitPtr.getValueType();
315 unsigned NumElements = Init->getType()->getArrayNumElements();
316 SmallVector<SDValue, 8> Chains;
317 for (unsigned i = 0; i < NumElements; ++i) {
318 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
319 Init->getType()->getArrayElementType()), PtrVT);
320 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
321 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
322 GV, Ptr, Chain, DAG));
324 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
328 llvm_unreachable("Unhandled constant initializer");
332 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
334 SelectionDAG &DAG) const {
336 const DataLayout *TD = getTargetMachine().getDataLayout();
337 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
338 const GlobalValue *GV = G->getGlobal();
340 switch (G->getAddressSpace()) {
341 default: llvm_unreachable("Global Address lowering not implemented for this "
343 case AMDGPUAS::LOCAL_ADDRESS: {
344 // XXX: What does the value of G->getOffset() mean?
345 assert(G->getOffset() == 0 &&
346 "Do not know what to do with an non-zero offset");
349 if (MFI->LocalMemoryObjects.count(GV) == 0) {
350 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
351 Offset = MFI->LDSSize;
352 MFI->LocalMemoryObjects[GV] = Offset;
353 // XXX: Account for alignment?
354 MFI->LDSSize += Size;
356 Offset = MFI->LocalMemoryObjects[GV];
359 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
361 case AMDGPUAS::CONSTANT_ADDRESS: {
362 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
363 Type *EltType = GV->getType()->getElementType();
364 unsigned Size = TD->getTypeAllocSize(EltType);
365 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
367 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
368 const Constant *Init = Var->getInitializer();
369 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
370 SDValue InitPtr = DAG.getFrameIndex(FI,
371 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
372 SmallVector<SDNode*, 8> WorkList;
374 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
375 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
376 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
378 WorkList.push_back(*I);
380 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
381 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
382 E = WorkList.end(); I != E; ++I) {
383 SmallVector<SDValue, 8> Ops;
384 Ops.push_back(Chain);
385 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
386 Ops.push_back((*I)->getOperand(i));
388 DAG.UpdateNodeOperands(*I, &Ops[0], Ops.size());
390 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
391 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
396 void AMDGPUTargetLowering::ExtractVectorElements(SDValue Op, SelectionDAG &DAG,
397 SmallVectorImpl<SDValue> &Args,
399 unsigned Count) const {
400 EVT VT = Op.getValueType();
401 for (unsigned i = Start, e = Start + Count; i != e; ++i) {
402 Args.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
403 VT.getVectorElementType(),
404 Op, DAG.getConstant(i, MVT::i32)));
408 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
409 SelectionDAG &DAG) const {
410 SmallVector<SDValue, 8> Args;
411 SDValue A = Op.getOperand(0);
412 SDValue B = Op.getOperand(1);
414 ExtractVectorElements(A, DAG, Args, 0,
415 A.getValueType().getVectorNumElements());
416 ExtractVectorElements(B, DAG, Args, 0,
417 B.getValueType().getVectorNumElements());
419 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(),
420 &Args[0], Args.size());
423 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
424 SelectionDAG &DAG) const {
426 SmallVector<SDValue, 8> Args;
427 EVT VT = Op.getValueType();
428 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
429 ExtractVectorElements(Op.getOperand(0), DAG, Args, Start,
430 VT.getVectorNumElements());
432 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(),
433 &Args[0], Args.size());
436 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
437 SelectionDAG &DAG) const {
439 MachineFunction &MF = DAG.getMachineFunction();
440 const AMDGPUFrameLowering *TFL =
441 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
443 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
446 unsigned FrameIndex = FIN->getIndex();
447 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
448 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
452 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
453 SelectionDAG &DAG) const {
454 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
456 EVT VT = Op.getValueType();
458 switch (IntrinsicID) {
460 case AMDGPUIntrinsic::AMDIL_abs:
461 return LowerIntrinsicIABS(Op, DAG);
462 case AMDGPUIntrinsic::AMDIL_exp:
463 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
464 case AMDGPUIntrinsic::AMDGPU_lrp:
465 return LowerIntrinsicLRP(Op, DAG);
466 case AMDGPUIntrinsic::AMDIL_fraction:
467 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
468 case AMDGPUIntrinsic::AMDIL_max:
469 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
471 case AMDGPUIntrinsic::AMDGPU_imax:
472 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
474 case AMDGPUIntrinsic::AMDGPU_umax:
475 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
477 case AMDGPUIntrinsic::AMDIL_min:
478 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
480 case AMDGPUIntrinsic::AMDGPU_imin:
481 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
483 case AMDGPUIntrinsic::AMDGPU_umin:
484 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
486 case AMDGPUIntrinsic::AMDIL_round_nearest:
487 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
491 ///IABS(a) = SMAX(sub(0, a), a)
492 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
493 SelectionDAG &DAG) const {
496 EVT VT = Op.getValueType();
497 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
500 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
503 /// Linear Interpolation
504 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
505 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
506 SelectionDAG &DAG) const {
508 EVT VT = Op.getValueType();
509 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
510 DAG.getConstantFP(1.0f, MVT::f32),
512 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
514 return DAG.getNode(ISD::FADD, DL, VT,
515 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
519 /// \brief Generate Min/Max node
520 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
521 SelectionDAG &DAG) const {
523 EVT VT = Op.getValueType();
525 SDValue LHS = Op.getOperand(0);
526 SDValue RHS = Op.getOperand(1);
527 SDValue True = Op.getOperand(2);
528 SDValue False = Op.getOperand(3);
529 SDValue CC = Op.getOperand(4);
531 if (VT != MVT::f32 ||
532 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
536 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
550 llvm_unreachable("Operation should already be optimised!");
558 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
560 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
569 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
571 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
573 case ISD::SETCC_INVALID:
574 llvm_unreachable("Invalid setcc condcode!");
579 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
580 SelectionDAG &DAG) const {
581 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
582 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
583 EVT EltVT = Op.getValueType().getVectorElementType();
584 EVT PtrVT = Load->getBasePtr().getValueType();
585 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
586 SmallVector<SDValue, 8> Loads;
589 for (unsigned i = 0, e = NumElts; i != e; ++i) {
590 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
591 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
592 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
593 Load->getChain(), Ptr,
594 MachinePointerInfo(Load->getMemOperand()->getValue()),
595 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
596 Load->getAlignment()));
598 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), &Loads[0],
602 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
603 SelectionDAG &DAG) const {
604 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
605 EVT MemVT = Store->getMemoryVT();
606 unsigned MemBits = MemVT.getSizeInBits();
608 // Byte stores are really expensive, so if possible, try to pack
609 // 32-bit vector truncatating store into an i32 store.
610 // XXX: We could also handle optimize other vector bitwidths
611 if (!MemVT.isVector() || MemBits > 32) {
616 const SDValue &Value = Store->getValue();
617 EVT VT = Value.getValueType();
618 const SDValue &Ptr = Store->getBasePtr();
619 EVT MemEltVT = MemVT.getVectorElementType();
620 unsigned MemEltBits = MemEltVT.getSizeInBits();
621 unsigned MemNumElements = MemVT.getVectorNumElements();
622 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits());
626 Mask = DAG.getConstant(0xFF, PackedVT);
629 Mask = DAG.getConstant(0xFFFF, PackedVT);
632 llvm_unreachable("Cannot lower this vector store");
635 for (unsigned i = 0; i < MemNumElements; ++i) {
636 EVT ElemVT = VT.getVectorElementType();
637 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
638 DAG.getConstant(i, MVT::i32));
639 Elt = DAG.getZExtOrTrunc(Elt, DL, PackedVT);
640 Elt = DAG.getNode(ISD::AND, DL, PackedVT, Elt, Mask);
641 SDValue Shift = DAG.getConstant(MemEltBits * i, PackedVT);
642 Elt = DAG.getNode(ISD::SHL, DL, PackedVT, Elt, Shift);
646 PackedValue = DAG.getNode(ISD::OR, DL, PackedVT, PackedValue, Elt);
649 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
650 MachinePointerInfo(Store->getMemOperand()->getValue()),
651 Store->isVolatile(), Store->isNonTemporal(),
652 Store->getAlignment());
655 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
656 SelectionDAG &DAG) const {
657 StoreSDNode *Store = cast<StoreSDNode>(Op);
658 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
659 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
660 EVT PtrVT = Store->getBasePtr().getValueType();
661 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
664 SmallVector<SDValue, 8> Chains;
666 for (unsigned i = 0, e = NumElts; i != e; ++i) {
667 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
668 Store->getValue(), DAG.getConstant(i, MVT::i32));
669 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
671 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
673 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
674 MachinePointerInfo(Store->getMemOperand()->getValue()),
675 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
676 Store->getAlignment()));
678 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, &Chains[0], NumElts);
681 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
683 LoadSDNode *Load = cast<LoadSDNode>(Op);
684 ISD::LoadExtType ExtType = Load->getExtensionType();
686 // Lower loads constant address space global variable loads
687 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
688 isa<GlobalVariable>(GetUnderlyingObject(Load->getPointerInfo().V))) {
690 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
691 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
692 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
693 DAG.getConstant(2, MVT::i32));
694 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
695 Load->getChain(), Ptr,
696 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
699 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
700 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
704 EVT VT = Op.getValueType();
705 EVT MemVT = Load->getMemoryVT();
707 if (Load->getMemoryVT() == MVT::i8) {
709 } else if (Load->getMemoryVT() == MVT::i16) {
712 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
713 DAG.getConstant(2, MVT::i32));
714 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
715 Load->getChain(), Ptr,
716 DAG.getTargetConstant(0, MVT::i32),
718 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
720 DAG.getConstant(0x3, MVT::i32));
721 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
722 DAG.getConstant(3, MVT::i32));
723 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
724 Ret = DAG.getNode(ISD::AND, DL, MVT::i32, Ret,
725 DAG.getConstant(Mask, MVT::i32));
726 if (ExtType == ISD::SEXTLOAD) {
727 SDValue SExtShift = DAG.getConstant(
728 VT.getSizeInBits() - MemVT.getSizeInBits(), MVT::i32);
729 Ret = DAG.getNode(ISD::SHL, DL, MVT::i32, Ret, SExtShift);
730 Ret = DAG.getNode(ISD::SRA, DL, MVT::i32, Ret, SExtShift);
736 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
738 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
739 if (Result.getNode()) {
743 StoreSDNode *Store = cast<StoreSDNode>(Op);
744 SDValue Chain = Store->getChain();
745 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
746 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
747 Store->getValue().getValueType().isVector()) {
748 return SplitVectorStore(Op, DAG);
751 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
752 Store->getMemoryVT().bitsLT(MVT::i32)) {
754 if (Store->getMemoryVT() == MVT::i8) {
756 } else if (Store->getMemoryVT() == MVT::i16) {
759 SDValue TruncPtr = DAG.getZExtOrTrunc(Store->getBasePtr(), DL, MVT::i32);
760 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, TruncPtr,
761 DAG.getConstant(2, MVT::i32));
762 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
763 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
764 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, TruncPtr,
765 DAG.getConstant(0x3, MVT::i32));
766 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
767 DAG.getConstant(3, MVT::i32));
768 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
770 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, SExtValue,
771 DAG.getConstant(Mask, MVT::i32));
772 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
773 MaskedValue, ShiftAmt);
774 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
776 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
777 DAG.getConstant(0xffffffff, MVT::i32));
778 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
780 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
781 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
782 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
787 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
788 SelectionDAG &DAG) const {
790 EVT VT = Op.getValueType();
792 SDValue Num = Op.getOperand(0);
793 SDValue Den = Op.getOperand(1);
795 SmallVector<SDValue, 8> Results;
797 // RCP = URECIP(Den) = 2^32 / Den + e
798 // e is rounding error.
799 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
801 // RCP_LO = umulo(RCP, Den) */
802 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
804 // RCP_HI = mulhu (RCP, Den) */
805 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
807 // NEG_RCP_LO = -RCP_LO
808 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
811 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
812 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
815 // Calculate the rounding error from the URECIP instruction
816 // E = mulhu(ABS_RCP_LO, RCP)
817 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
820 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
823 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
825 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
826 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
829 // Quotient = mulhu(Tmp0, Num)
830 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
832 // Num_S_Remainder = Quotient * Den
833 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
835 // Remainder = Num - Num_S_Remainder
836 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
838 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
839 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
840 DAG.getConstant(-1, VT),
841 DAG.getConstant(0, VT),
843 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
844 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
846 DAG.getConstant(-1, VT),
847 DAG.getConstant(0, VT),
849 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
850 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
853 // Calculate Division result:
855 // Quotient_A_One = Quotient + 1
856 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
857 DAG.getConstant(1, VT));
859 // Quotient_S_One = Quotient - 1
860 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
861 DAG.getConstant(1, VT));
863 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
864 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
865 Quotient, Quotient_A_One, ISD::SETEQ);
867 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
868 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
869 Quotient_S_One, Div, ISD::SETEQ);
871 // Calculate Rem result:
873 // Remainder_S_Den = Remainder - Den
874 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
876 // Remainder_A_Den = Remainder + Den
877 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
879 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
880 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
881 Remainder, Remainder_S_Den, ISD::SETEQ);
883 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
884 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
885 Remainder_A_Den, Rem, ISD::SETEQ);
889 return DAG.getMergeValues(Ops, 2, DL);
892 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
893 SelectionDAG &DAG) const {
894 SDValue S0 = Op.getOperand(0);
896 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
899 // f32 uint_to_fp i64
900 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
901 DAG.getConstant(0, MVT::i32));
902 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
903 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
904 DAG.getConstant(1, MVT::i32));
905 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
906 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
907 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
908 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
912 //===----------------------------------------------------------------------===//
914 //===----------------------------------------------------------------------===//
916 void AMDGPUTargetLowering::getOriginalFunctionArgs(
919 const SmallVectorImpl<ISD::InputArg> &Ins,
920 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
922 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
923 if (Ins[i].ArgVT == Ins[i].VT) {
924 OrigIns.push_back(Ins[i]);
929 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
930 // Vector has been split into scalars.
931 VT = Ins[i].ArgVT.getVectorElementType();
932 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
933 Ins[i].ArgVT.getVectorElementType() !=
934 Ins[i].VT.getVectorElementType()) {
935 // Vector elements have been promoted
938 // Vector has been spilt into smaller vectors.
942 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
943 Ins[i].OrigArgIndex, Ins[i].PartOffset);
944 OrigIns.push_back(Arg);
948 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
949 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
950 return CFP->isExactlyValue(1.0);
952 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
953 return C->isAllOnesValue();
958 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
959 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
960 return CFP->getValueAPF().isZero();
962 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
963 return C->isNullValue();
968 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
969 const TargetRegisterClass *RC,
970 unsigned Reg, EVT VT) const {
971 MachineFunction &MF = DAG.getMachineFunction();
972 MachineRegisterInfo &MRI = MF.getRegInfo();
973 unsigned VirtualRegister;
974 if (!MRI.isLiveIn(Reg)) {
975 VirtualRegister = MRI.createVirtualRegister(RC);
976 MRI.addLiveIn(Reg, VirtualRegister);
978 VirtualRegister = MRI.getLiveInVirtReg(Reg);
980 return DAG.getRegister(VirtualRegister, VT);
983 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
985 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
989 NODE_NAME_CASE(CALL);
990 NODE_NAME_CASE(UMUL);
991 NODE_NAME_CASE(DIV_INF);
992 NODE_NAME_CASE(RET_FLAG);
993 NODE_NAME_CASE(BRANCH_COND);
996 NODE_NAME_CASE(DWORDADDR)
997 NODE_NAME_CASE(FRACT)
1000 NODE_NAME_CASE(UMAX)
1001 NODE_NAME_CASE(FMIN)
1002 NODE_NAME_CASE(SMIN)
1003 NODE_NAME_CASE(UMIN)
1004 NODE_NAME_CASE(URECIP)
1005 NODE_NAME_CASE(EXPORT)
1006 NODE_NAME_CASE(CONST_ADDRESS)
1007 NODE_NAME_CASE(REGISTER_LOAD)
1008 NODE_NAME_CASE(REGISTER_STORE)
1009 NODE_NAME_CASE(LOAD_CONSTANT)
1010 NODE_NAME_CASE(LOAD_INPUT)
1011 NODE_NAME_CASE(SAMPLE)
1012 NODE_NAME_CASE(SAMPLEB)
1013 NODE_NAME_CASE(SAMPLED)
1014 NODE_NAME_CASE(SAMPLEL)
1015 NODE_NAME_CASE(STORE_MSKOR)
1016 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)