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 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
50 // Initialize target lowering borrowed from AMDIL
53 // We need to custom lower some of the intrinsics
54 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
56 // Library functions. These default to Expand, but we have instructions
58 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
59 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
60 setOperationAction(ISD::FPOW, MVT::f32, Legal);
61 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
62 setOperationAction(ISD::FABS, MVT::f32, Legal);
63 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
64 setOperationAction(ISD::FRINT, MVT::f32, Legal);
65 setOperationAction(ISD::FROUND, MVT::f32, Legal);
66 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
68 // The hardware supports ROTR, but not ROTL
69 setOperationAction(ISD::ROTL, MVT::i32, Expand);
71 // Lower floating point store/load to integer store/load to reduce the number
72 // of patterns in tablegen.
73 setOperationAction(ISD::STORE, MVT::f32, Promote);
74 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
76 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
77 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
79 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
80 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
82 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
83 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
85 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
86 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
88 setOperationAction(ISD::STORE, MVT::f64, Promote);
89 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
91 // Custom lowering of vector stores is required for local address space
93 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
94 // XXX: Native v2i32 local address space stores are possible, but not
95 // currently implemented.
96 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
98 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
99 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
100 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
101 // XXX: This can be change to Custom, once ExpandVectorStores can
102 // handle 64-bit stores.
103 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
105 setOperationAction(ISD::LOAD, MVT::f32, Promote);
106 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
108 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
109 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
111 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
112 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
114 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
115 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
117 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
118 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
120 setOperationAction(ISD::LOAD, MVT::f64, Promote);
121 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
123 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
124 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
125 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
126 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
127 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
128 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
129 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
130 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
131 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
132 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
134 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
135 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
136 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
137 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
138 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
139 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
140 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
141 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
142 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
143 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
144 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
145 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
147 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
149 setOperationAction(ISD::FNEG, MVT::v2f32, Expand);
150 setOperationAction(ISD::FNEG, MVT::v4f32, Expand);
152 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
154 setOperationAction(ISD::MUL, MVT::i64, Expand);
156 setOperationAction(ISD::UDIV, MVT::i32, Expand);
157 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
158 setOperationAction(ISD::UREM, MVT::i32, Expand);
159 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
160 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
162 static const MVT::SimpleValueType IntTypes[] = {
163 MVT::v2i32, MVT::v4i32
165 const size_t NumIntTypes = array_lengthof(IntTypes);
167 for (unsigned int x = 0; x < NumIntTypes; ++x) {
168 MVT::SimpleValueType VT = IntTypes[x];
169 //Expand the following operations for the current type by default
170 setOperationAction(ISD::ADD, VT, Expand);
171 setOperationAction(ISD::AND, VT, Expand);
172 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
173 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
174 setOperationAction(ISD::MUL, VT, Expand);
175 setOperationAction(ISD::OR, VT, Expand);
176 setOperationAction(ISD::SHL, VT, Expand);
177 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
178 setOperationAction(ISD::SRL, VT, Expand);
179 setOperationAction(ISD::SRA, VT, Expand);
180 setOperationAction(ISD::SUB, VT, Expand);
181 setOperationAction(ISD::UDIV, VT, Expand);
182 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
183 setOperationAction(ISD::UREM, VT, Expand);
184 setOperationAction(ISD::SELECT, VT, Expand);
185 setOperationAction(ISD::VSELECT, VT, Expand);
186 setOperationAction(ISD::XOR, VT, Expand);
189 static const MVT::SimpleValueType FloatTypes[] = {
190 MVT::v2f32, MVT::v4f32
192 const size_t NumFloatTypes = array_lengthof(FloatTypes);
194 for (unsigned int x = 0; x < NumFloatTypes; ++x) {
195 MVT::SimpleValueType VT = FloatTypes[x];
196 setOperationAction(ISD::FABS, VT, Expand);
197 setOperationAction(ISD::FADD, VT, Expand);
198 setOperationAction(ISD::FDIV, VT, Expand);
199 setOperationAction(ISD::FPOW, VT, Expand);
200 setOperationAction(ISD::FFLOOR, VT, Expand);
201 setOperationAction(ISD::FTRUNC, VT, Expand);
202 setOperationAction(ISD::FMUL, VT, Expand);
203 setOperationAction(ISD::FRINT, VT, Expand);
204 setOperationAction(ISD::FSQRT, VT, Expand);
205 setOperationAction(ISD::FSUB, VT, Expand);
206 setOperationAction(ISD::SELECT, VT, Expand);
210 //===----------------------------------------------------------------------===//
211 // Target Information
212 //===----------------------------------------------------------------------===//
214 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
218 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
220 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
223 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
224 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
226 return ((LScalarSize <= CastScalarSize) ||
227 (CastScalarSize >= 32) ||
231 //===---------------------------------------------------------------------===//
233 //===---------------------------------------------------------------------===//
235 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
236 assert(VT.isFloatingPoint());
237 return VT == MVT::f32;
240 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
241 assert(VT.isFloatingPoint());
242 return VT == MVT::f32;
245 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
246 // Truncate is just accessing a subregister.
247 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
250 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
251 // Truncate is just accessing a subregister.
252 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
253 (Dest->getPrimitiveSizeInBits() % 32 == 0);
256 //===---------------------------------------------------------------------===//
257 // TargetLowering Callbacks
258 //===---------------------------------------------------------------------===//
260 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
261 const SmallVectorImpl<ISD::InputArg> &Ins) const {
263 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
266 SDValue AMDGPUTargetLowering::LowerReturn(
268 CallingConv::ID CallConv,
270 const SmallVectorImpl<ISD::OutputArg> &Outs,
271 const SmallVectorImpl<SDValue> &OutVals,
272 SDLoc DL, SelectionDAG &DAG) const {
273 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
276 //===---------------------------------------------------------------------===//
277 // Target specific lowering
278 //===---------------------------------------------------------------------===//
280 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
282 switch (Op.getOpcode()) {
284 Op.getNode()->dump();
285 llvm_unreachable("Custom lowering code for this"
286 "instruction is not implemented yet!");
288 // AMDIL DAG lowering
289 case ISD::SDIV: return LowerSDIV(Op, DAG);
290 case ISD::SREM: return LowerSREM(Op, DAG);
291 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
292 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
293 // AMDGPU DAG lowering
294 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
295 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
296 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
297 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
298 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
299 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
304 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
305 const GlobalValue *GV,
306 const SDValue &InitPtr,
308 SelectionDAG &DAG) const {
309 const DataLayout *TD = getTargetMachine().getDataLayout();
311 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
312 EVT VT = EVT::getEVT(CI->getType());
313 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
314 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
315 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
316 TD->getPrefTypeAlignment(CI->getType()));
317 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
318 EVT VT = EVT::getEVT(CFP->getType());
319 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
320 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
321 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
322 TD->getPrefTypeAlignment(CFP->getType()));
323 } else if (Init->getType()->isAggregateType()) {
324 EVT PtrVT = InitPtr.getValueType();
325 unsigned NumElements = Init->getType()->getArrayNumElements();
326 SmallVector<SDValue, 8> Chains;
327 for (unsigned i = 0; i < NumElements; ++i) {
328 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
329 Init->getType()->getArrayElementType()), PtrVT);
330 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
331 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
332 GV, Ptr, Chain, DAG));
334 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
338 llvm_unreachable("Unhandled constant initializer");
342 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
344 SelectionDAG &DAG) const {
346 const DataLayout *TD = getTargetMachine().getDataLayout();
347 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
348 const GlobalValue *GV = G->getGlobal();
350 switch (G->getAddressSpace()) {
351 default: llvm_unreachable("Global Address lowering not implemented for this "
353 case AMDGPUAS::LOCAL_ADDRESS: {
354 // XXX: What does the value of G->getOffset() mean?
355 assert(G->getOffset() == 0 &&
356 "Do not know what to do with an non-zero offset");
359 if (MFI->LocalMemoryObjects.count(GV) == 0) {
360 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
361 Offset = MFI->LDSSize;
362 MFI->LocalMemoryObjects[GV] = Offset;
363 // XXX: Account for alignment?
364 MFI->LDSSize += Size;
366 Offset = MFI->LocalMemoryObjects[GV];
369 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
371 case AMDGPUAS::CONSTANT_ADDRESS: {
372 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
373 Type *EltType = GV->getType()->getElementType();
374 unsigned Size = TD->getTypeAllocSize(EltType);
375 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
377 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
378 const Constant *Init = Var->getInitializer();
379 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
380 SDValue InitPtr = DAG.getFrameIndex(FI,
381 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
382 SmallVector<SDNode*, 8> WorkList;
384 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
385 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
386 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
388 WorkList.push_back(*I);
390 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
391 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
392 E = WorkList.end(); I != E; ++I) {
393 SmallVector<SDValue, 8> Ops;
394 Ops.push_back(Chain);
395 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
396 Ops.push_back((*I)->getOperand(i));
398 DAG.UpdateNodeOperands(*I, &Ops[0], Ops.size());
400 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
401 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
406 void AMDGPUTargetLowering::ExtractVectorElements(SDValue Op, SelectionDAG &DAG,
407 SmallVectorImpl<SDValue> &Args,
409 unsigned Count) const {
410 EVT VT = Op.getValueType();
411 for (unsigned i = Start, e = Start + Count; i != e; ++i) {
412 Args.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
413 VT.getVectorElementType(),
414 Op, DAG.getConstant(i, MVT::i32)));
418 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
419 SelectionDAG &DAG) const {
420 SmallVector<SDValue, 8> Args;
421 SDValue A = Op.getOperand(0);
422 SDValue B = Op.getOperand(1);
424 ExtractVectorElements(A, DAG, Args, 0,
425 A.getValueType().getVectorNumElements());
426 ExtractVectorElements(B, DAG, Args, 0,
427 B.getValueType().getVectorNumElements());
429 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(),
430 &Args[0], Args.size());
433 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
434 SelectionDAG &DAG) const {
436 SmallVector<SDValue, 8> Args;
437 EVT VT = Op.getValueType();
438 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
439 ExtractVectorElements(Op.getOperand(0), DAG, Args, Start,
440 VT.getVectorNumElements());
442 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(),
443 &Args[0], Args.size());
446 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
447 SelectionDAG &DAG) const {
449 MachineFunction &MF = DAG.getMachineFunction();
450 const AMDGPUFrameLowering *TFL =
451 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
453 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
456 unsigned FrameIndex = FIN->getIndex();
457 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
458 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
462 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
463 SelectionDAG &DAG) const {
464 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
466 EVT VT = Op.getValueType();
468 switch (IntrinsicID) {
470 case AMDGPUIntrinsic::AMDIL_abs:
471 return LowerIntrinsicIABS(Op, DAG);
472 case AMDGPUIntrinsic::AMDIL_exp:
473 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
474 case AMDGPUIntrinsic::AMDGPU_lrp:
475 return LowerIntrinsicLRP(Op, DAG);
476 case AMDGPUIntrinsic::AMDIL_fraction:
477 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
478 case AMDGPUIntrinsic::AMDIL_max:
479 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
481 case AMDGPUIntrinsic::AMDGPU_imax:
482 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
484 case AMDGPUIntrinsic::AMDGPU_umax:
485 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
487 case AMDGPUIntrinsic::AMDIL_min:
488 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
490 case AMDGPUIntrinsic::AMDGPU_imin:
491 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
493 case AMDGPUIntrinsic::AMDGPU_umin:
494 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
496 case AMDGPUIntrinsic::AMDIL_round_nearest:
497 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
501 ///IABS(a) = SMAX(sub(0, a), a)
502 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
503 SelectionDAG &DAG) const {
506 EVT VT = Op.getValueType();
507 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
510 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
513 /// Linear Interpolation
514 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
515 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
516 SelectionDAG &DAG) const {
518 EVT VT = Op.getValueType();
519 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
520 DAG.getConstantFP(1.0f, MVT::f32),
522 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
524 return DAG.getNode(ISD::FADD, DL, VT,
525 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
529 /// \brief Generate Min/Max node
530 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
531 SelectionDAG &DAG) const {
533 EVT VT = Op.getValueType();
535 SDValue LHS = Op.getOperand(0);
536 SDValue RHS = Op.getOperand(1);
537 SDValue True = Op.getOperand(2);
538 SDValue False = Op.getOperand(3);
539 SDValue CC = Op.getOperand(4);
541 if (VT != MVT::f32 ||
542 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
546 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
560 llvm_unreachable("Operation should already be optimised!");
568 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
570 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
579 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
581 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
583 case ISD::SETCC_INVALID:
584 llvm_unreachable("Invalid setcc condcode!");
589 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
590 SelectionDAG &DAG) const {
591 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
592 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
593 EVT EltVT = Op.getValueType().getVectorElementType();
594 EVT PtrVT = Load->getBasePtr().getValueType();
595 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
596 SmallVector<SDValue, 8> Loads;
599 for (unsigned i = 0, e = NumElts; i != e; ++i) {
600 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
601 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
602 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
603 Load->getChain(), Ptr,
604 MachinePointerInfo(Load->getMemOperand()->getValue()),
605 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
606 Load->getAlignment()));
608 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), &Loads[0],
612 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
613 SelectionDAG &DAG) const {
614 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
615 EVT MemVT = Store->getMemoryVT();
616 unsigned MemBits = MemVT.getSizeInBits();
618 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
619 // truncating store into an i32 store.
620 // XXX: We could also handle optimize other vector bitwidths.
621 if (!MemVT.isVector() || MemBits > 32) {
626 const SDValue &Value = Store->getValue();
627 EVT VT = Value.getValueType();
628 const SDValue &Ptr = Store->getBasePtr();
629 EVT MemEltVT = MemVT.getVectorElementType();
630 unsigned MemEltBits = MemEltVT.getSizeInBits();
631 unsigned MemNumElements = MemVT.getVectorNumElements();
632 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits());
636 Mask = DAG.getConstant(0xFF, PackedVT);
639 Mask = DAG.getConstant(0xFFFF, PackedVT);
642 llvm_unreachable("Cannot lower this vector store");
645 for (unsigned i = 0; i < MemNumElements; ++i) {
646 EVT ElemVT = VT.getVectorElementType();
647 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
648 DAG.getConstant(i, MVT::i32));
649 Elt = DAG.getZExtOrTrunc(Elt, DL, PackedVT);
650 Elt = DAG.getNode(ISD::AND, DL, PackedVT, Elt, Mask);
651 SDValue Shift = DAG.getConstant(MemEltBits * i, PackedVT);
652 Elt = DAG.getNode(ISD::SHL, DL, PackedVT, Elt, Shift);
656 PackedValue = DAG.getNode(ISD::OR, DL, PackedVT, PackedValue, Elt);
659 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
660 MachinePointerInfo(Store->getMemOperand()->getValue()),
661 Store->isVolatile(), Store->isNonTemporal(),
662 Store->getAlignment());
665 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
666 SelectionDAG &DAG) const {
667 StoreSDNode *Store = cast<StoreSDNode>(Op);
668 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
669 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
670 EVT PtrVT = Store->getBasePtr().getValueType();
671 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
674 SmallVector<SDValue, 8> Chains;
676 for (unsigned i = 0, e = NumElts; i != e; ++i) {
677 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
678 Store->getValue(), DAG.getConstant(i, MVT::i32));
679 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
681 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
683 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
684 MachinePointerInfo(Store->getMemOperand()->getValue()),
685 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
686 Store->getAlignment()));
688 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, &Chains[0], NumElts);
691 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
693 LoadSDNode *Load = cast<LoadSDNode>(Op);
694 ISD::LoadExtType ExtType = Load->getExtensionType();
695 EVT VT = Op.getValueType();
696 EVT MemVT = Load->getMemoryVT();
698 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
699 // We can do the extload to 32-bits, and then need to separately extend to
702 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
706 Load->getMemOperand());
707 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
710 // Lower loads constant address space global variable loads
711 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
712 isa<GlobalVariable>(GetUnderlyingObject(Load->getPointerInfo().V))) {
714 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
715 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
716 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
717 DAG.getConstant(2, MVT::i32));
718 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
719 Load->getChain(), Ptr,
720 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
723 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
724 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
729 if (Load->getMemoryVT() == MVT::i8) {
731 } else if (Load->getMemoryVT() == MVT::i16) {
734 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
735 DAG.getConstant(2, MVT::i32));
736 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
737 Load->getChain(), Ptr,
738 DAG.getTargetConstant(0, MVT::i32),
740 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
742 DAG.getConstant(0x3, MVT::i32));
743 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
744 DAG.getConstant(3, MVT::i32));
745 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
746 Ret = DAG.getNode(ISD::AND, DL, MVT::i32, Ret,
747 DAG.getConstant(Mask, MVT::i32));
748 if (ExtType == ISD::SEXTLOAD) {
749 SDValue SExtShift = DAG.getConstant(
750 VT.getSizeInBits() - MemVT.getSizeInBits(), MVT::i32);
751 Ret = DAG.getNode(ISD::SHL, DL, MVT::i32, Ret, SExtShift);
752 Ret = DAG.getNode(ISD::SRA, DL, MVT::i32, Ret, SExtShift);
758 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
760 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
761 if (Result.getNode()) {
765 StoreSDNode *Store = cast<StoreSDNode>(Op);
766 SDValue Chain = Store->getChain();
767 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
768 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
769 Store->getValue().getValueType().isVector()) {
770 return SplitVectorStore(Op, DAG);
773 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
774 Store->getMemoryVT().bitsLT(MVT::i32)) {
776 if (Store->getMemoryVT() == MVT::i8) {
778 } else if (Store->getMemoryVT() == MVT::i16) {
781 SDValue TruncPtr = DAG.getZExtOrTrunc(Store->getBasePtr(), DL, MVT::i32);
782 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, TruncPtr,
783 DAG.getConstant(2, MVT::i32));
784 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
785 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
786 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, TruncPtr,
787 DAG.getConstant(0x3, MVT::i32));
788 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
789 DAG.getConstant(3, MVT::i32));
790 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
792 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, SExtValue,
793 DAG.getConstant(Mask, MVT::i32));
794 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
795 MaskedValue, ShiftAmt);
796 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
798 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
799 DAG.getConstant(0xffffffff, MVT::i32));
800 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
802 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
803 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
804 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
809 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
810 SelectionDAG &DAG) const {
812 EVT VT = Op.getValueType();
814 SDValue Num = Op.getOperand(0);
815 SDValue Den = Op.getOperand(1);
817 SmallVector<SDValue, 8> Results;
819 // RCP = URECIP(Den) = 2^32 / Den + e
820 // e is rounding error.
821 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
823 // RCP_LO = umulo(RCP, Den) */
824 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
826 // RCP_HI = mulhu (RCP, Den) */
827 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
829 // NEG_RCP_LO = -RCP_LO
830 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
833 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
834 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
837 // Calculate the rounding error from the URECIP instruction
838 // E = mulhu(ABS_RCP_LO, RCP)
839 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
842 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
845 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
847 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
848 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
851 // Quotient = mulhu(Tmp0, Num)
852 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
854 // Num_S_Remainder = Quotient * Den
855 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
857 // Remainder = Num - Num_S_Remainder
858 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
860 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
861 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
862 DAG.getConstant(-1, VT),
863 DAG.getConstant(0, VT),
865 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
866 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
868 DAG.getConstant(-1, VT),
869 DAG.getConstant(0, VT),
871 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
872 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
875 // Calculate Division result:
877 // Quotient_A_One = Quotient + 1
878 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
879 DAG.getConstant(1, VT));
881 // Quotient_S_One = Quotient - 1
882 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
883 DAG.getConstant(1, VT));
885 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
886 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
887 Quotient, Quotient_A_One, ISD::SETEQ);
889 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
890 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
891 Quotient_S_One, Div, ISD::SETEQ);
893 // Calculate Rem result:
895 // Remainder_S_Den = Remainder - Den
896 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
898 // Remainder_A_Den = Remainder + Den
899 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
901 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
902 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
903 Remainder, Remainder_S_Den, ISD::SETEQ);
905 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
906 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
907 Remainder_A_Den, Rem, ISD::SETEQ);
911 return DAG.getMergeValues(Ops, 2, DL);
914 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
915 SelectionDAG &DAG) const {
916 SDValue S0 = Op.getOperand(0);
918 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
921 // f32 uint_to_fp i64
922 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
923 DAG.getConstant(0, MVT::i32));
924 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
925 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
926 DAG.getConstant(1, MVT::i32));
927 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
928 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
929 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
930 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
934 //===----------------------------------------------------------------------===//
936 //===----------------------------------------------------------------------===//
938 void AMDGPUTargetLowering::getOriginalFunctionArgs(
941 const SmallVectorImpl<ISD::InputArg> &Ins,
942 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
944 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
945 if (Ins[i].ArgVT == Ins[i].VT) {
946 OrigIns.push_back(Ins[i]);
951 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
952 // Vector has been split into scalars.
953 VT = Ins[i].ArgVT.getVectorElementType();
954 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
955 Ins[i].ArgVT.getVectorElementType() !=
956 Ins[i].VT.getVectorElementType()) {
957 // Vector elements have been promoted
960 // Vector has been spilt into smaller vectors.
964 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
965 Ins[i].OrigArgIndex, Ins[i].PartOffset);
966 OrigIns.push_back(Arg);
970 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
971 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
972 return CFP->isExactlyValue(1.0);
974 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
975 return C->isAllOnesValue();
980 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
981 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
982 return CFP->getValueAPF().isZero();
984 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
985 return C->isNullValue();
990 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
991 const TargetRegisterClass *RC,
992 unsigned Reg, EVT VT) const {
993 MachineFunction &MF = DAG.getMachineFunction();
994 MachineRegisterInfo &MRI = MF.getRegInfo();
995 unsigned VirtualRegister;
996 if (!MRI.isLiveIn(Reg)) {
997 VirtualRegister = MRI.createVirtualRegister(RC);
998 MRI.addLiveIn(Reg, VirtualRegister);
1000 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1002 return DAG.getRegister(VirtualRegister, VT);
1005 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1007 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1011 NODE_NAME_CASE(CALL);
1012 NODE_NAME_CASE(UMUL);
1013 NODE_NAME_CASE(DIV_INF);
1014 NODE_NAME_CASE(RET_FLAG);
1015 NODE_NAME_CASE(BRANCH_COND);
1018 NODE_NAME_CASE(DWORDADDR)
1019 NODE_NAME_CASE(FRACT)
1020 NODE_NAME_CASE(FMAX)
1021 NODE_NAME_CASE(SMAX)
1022 NODE_NAME_CASE(UMAX)
1023 NODE_NAME_CASE(FMIN)
1024 NODE_NAME_CASE(SMIN)
1025 NODE_NAME_CASE(UMIN)
1026 NODE_NAME_CASE(URECIP)
1027 NODE_NAME_CASE(DOT4)
1028 NODE_NAME_CASE(EXPORT)
1029 NODE_NAME_CASE(CONST_ADDRESS)
1030 NODE_NAME_CASE(REGISTER_LOAD)
1031 NODE_NAME_CASE(REGISTER_STORE)
1032 NODE_NAME_CASE(LOAD_CONSTANT)
1033 NODE_NAME_CASE(LOAD_INPUT)
1034 NODE_NAME_CASE(SAMPLE)
1035 NODE_NAME_CASE(SAMPLEB)
1036 NODE_NAME_CASE(SAMPLED)
1037 NODE_NAME_CASE(SAMPLEL)
1038 NODE_NAME_CASE(STORE_MSKOR)
1039 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)