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::VSELECT, VT, Expand);
185 setOperationAction(ISD::XOR, VT, Expand);
188 static const MVT::SimpleValueType FloatTypes[] = {
189 MVT::v2f32, MVT::v4f32
191 const size_t NumFloatTypes = array_lengthof(FloatTypes);
193 for (unsigned int x = 0; x < NumFloatTypes; ++x) {
194 MVT::SimpleValueType VT = FloatTypes[x];
195 setOperationAction(ISD::FABS, VT, Expand);
196 setOperationAction(ISD::FADD, VT, Expand);
197 setOperationAction(ISD::FDIV, VT, Expand);
198 setOperationAction(ISD::FPOW, VT, Expand);
199 setOperationAction(ISD::FFLOOR, VT, Expand);
200 setOperationAction(ISD::FTRUNC, VT, Expand);
201 setOperationAction(ISD::FMUL, VT, Expand);
202 setOperationAction(ISD::FRINT, VT, Expand);
203 setOperationAction(ISD::FSQRT, VT, Expand);
204 setOperationAction(ISD::FSUB, VT, Expand);
208 //===----------------------------------------------------------------------===//
209 // Target Information
210 //===----------------------------------------------------------------------===//
212 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
216 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
218 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
221 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
222 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
224 return ((LScalarSize <= CastScalarSize) ||
225 (CastScalarSize >= 32) ||
229 //===---------------------------------------------------------------------===//
231 //===---------------------------------------------------------------------===//
233 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
234 assert(VT.isFloatingPoint());
235 return VT == MVT::f32;
238 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
239 assert(VT.isFloatingPoint());
240 return VT == MVT::f32;
243 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
244 // Truncate is just accessing a subregister.
245 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
248 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
249 // Truncate is just accessing a subregister.
250 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
251 (Dest->getPrimitiveSizeInBits() % 32 == 0);
254 //===---------------------------------------------------------------------===//
255 // TargetLowering Callbacks
256 //===---------------------------------------------------------------------===//
258 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
259 const SmallVectorImpl<ISD::InputArg> &Ins) const {
261 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
264 SDValue AMDGPUTargetLowering::LowerReturn(
266 CallingConv::ID CallConv,
268 const SmallVectorImpl<ISD::OutputArg> &Outs,
269 const SmallVectorImpl<SDValue> &OutVals,
270 SDLoc DL, SelectionDAG &DAG) const {
271 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
274 //===---------------------------------------------------------------------===//
275 // Target specific lowering
276 //===---------------------------------------------------------------------===//
278 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
280 switch (Op.getOpcode()) {
282 Op.getNode()->dump();
283 llvm_unreachable("Custom lowering code for this"
284 "instruction is not implemented yet!");
286 // AMDIL DAG lowering
287 case ISD::SDIV: return LowerSDIV(Op, DAG);
288 case ISD::SREM: return LowerSREM(Op, DAG);
289 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
290 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
291 // AMDGPU DAG lowering
292 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
293 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
294 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
295 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
296 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
297 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
302 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
303 const GlobalValue *GV,
304 const SDValue &InitPtr,
306 SelectionDAG &DAG) const {
307 const DataLayout *TD = getTargetMachine().getDataLayout();
309 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
310 EVT VT = EVT::getEVT(CI->getType());
311 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
312 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
313 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
314 TD->getPrefTypeAlignment(CI->getType()));
315 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
316 EVT VT = EVT::getEVT(CFP->getType());
317 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
318 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
319 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
320 TD->getPrefTypeAlignment(CFP->getType()));
321 } else if (Init->getType()->isAggregateType()) {
322 EVT PtrVT = InitPtr.getValueType();
323 unsigned NumElements = Init->getType()->getArrayNumElements();
324 SmallVector<SDValue, 8> Chains;
325 for (unsigned i = 0; i < NumElements; ++i) {
326 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
327 Init->getType()->getArrayElementType()), PtrVT);
328 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
329 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
330 GV, Ptr, Chain, DAG));
332 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
336 llvm_unreachable("Unhandled constant initializer");
340 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
342 SelectionDAG &DAG) const {
344 const DataLayout *TD = getTargetMachine().getDataLayout();
345 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
346 const GlobalValue *GV = G->getGlobal();
348 switch (G->getAddressSpace()) {
349 default: llvm_unreachable("Global Address lowering not implemented for this "
351 case AMDGPUAS::LOCAL_ADDRESS: {
352 // XXX: What does the value of G->getOffset() mean?
353 assert(G->getOffset() == 0 &&
354 "Do not know what to do with an non-zero offset");
357 if (MFI->LocalMemoryObjects.count(GV) == 0) {
358 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
359 Offset = MFI->LDSSize;
360 MFI->LocalMemoryObjects[GV] = Offset;
361 // XXX: Account for alignment?
362 MFI->LDSSize += Size;
364 Offset = MFI->LocalMemoryObjects[GV];
367 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
369 case AMDGPUAS::CONSTANT_ADDRESS: {
370 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
371 Type *EltType = GV->getType()->getElementType();
372 unsigned Size = TD->getTypeAllocSize(EltType);
373 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
375 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
376 const Constant *Init = Var->getInitializer();
377 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
378 SDValue InitPtr = DAG.getFrameIndex(FI,
379 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
380 SmallVector<SDNode*, 8> WorkList;
382 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
383 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
384 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
386 WorkList.push_back(*I);
388 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
389 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
390 E = WorkList.end(); I != E; ++I) {
391 SmallVector<SDValue, 8> Ops;
392 Ops.push_back(Chain);
393 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
394 Ops.push_back((*I)->getOperand(i));
396 DAG.UpdateNodeOperands(*I, &Ops[0], Ops.size());
398 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
399 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
404 void AMDGPUTargetLowering::ExtractVectorElements(SDValue Op, SelectionDAG &DAG,
405 SmallVectorImpl<SDValue> &Args,
407 unsigned Count) const {
408 EVT VT = Op.getValueType();
409 for (unsigned i = Start, e = Start + Count; i != e; ++i) {
410 Args.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
411 VT.getVectorElementType(),
412 Op, DAG.getConstant(i, MVT::i32)));
416 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
417 SelectionDAG &DAG) const {
418 SmallVector<SDValue, 8> Args;
419 SDValue A = Op.getOperand(0);
420 SDValue B = Op.getOperand(1);
422 ExtractVectorElements(A, DAG, Args, 0,
423 A.getValueType().getVectorNumElements());
424 ExtractVectorElements(B, DAG, Args, 0,
425 B.getValueType().getVectorNumElements());
427 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(),
428 &Args[0], Args.size());
431 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
432 SelectionDAG &DAG) const {
434 SmallVector<SDValue, 8> Args;
435 EVT VT = Op.getValueType();
436 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
437 ExtractVectorElements(Op.getOperand(0), DAG, Args, Start,
438 VT.getVectorNumElements());
440 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(),
441 &Args[0], Args.size());
444 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
445 SelectionDAG &DAG) const {
447 MachineFunction &MF = DAG.getMachineFunction();
448 const AMDGPUFrameLowering *TFL =
449 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
451 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
454 unsigned FrameIndex = FIN->getIndex();
455 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
456 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
460 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
461 SelectionDAG &DAG) const {
462 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
464 EVT VT = Op.getValueType();
466 switch (IntrinsicID) {
468 case AMDGPUIntrinsic::AMDIL_abs:
469 return LowerIntrinsicIABS(Op, DAG);
470 case AMDGPUIntrinsic::AMDIL_exp:
471 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
472 case AMDGPUIntrinsic::AMDGPU_lrp:
473 return LowerIntrinsicLRP(Op, DAG);
474 case AMDGPUIntrinsic::AMDIL_fraction:
475 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
476 case AMDGPUIntrinsic::AMDIL_max:
477 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
479 case AMDGPUIntrinsic::AMDGPU_imax:
480 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
482 case AMDGPUIntrinsic::AMDGPU_umax:
483 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
485 case AMDGPUIntrinsic::AMDIL_min:
486 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
488 case AMDGPUIntrinsic::AMDGPU_imin:
489 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
491 case AMDGPUIntrinsic::AMDGPU_umin:
492 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
494 case AMDGPUIntrinsic::AMDIL_round_nearest:
495 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
499 ///IABS(a) = SMAX(sub(0, a), a)
500 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
501 SelectionDAG &DAG) const {
504 EVT VT = Op.getValueType();
505 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
508 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
511 /// Linear Interpolation
512 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
513 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
514 SelectionDAG &DAG) const {
516 EVT VT = Op.getValueType();
517 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
518 DAG.getConstantFP(1.0f, MVT::f32),
520 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
522 return DAG.getNode(ISD::FADD, DL, VT,
523 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
527 /// \brief Generate Min/Max node
528 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
529 SelectionDAG &DAG) const {
531 EVT VT = Op.getValueType();
533 SDValue LHS = Op.getOperand(0);
534 SDValue RHS = Op.getOperand(1);
535 SDValue True = Op.getOperand(2);
536 SDValue False = Op.getOperand(3);
537 SDValue CC = Op.getOperand(4);
539 if (VT != MVT::f32 ||
540 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
544 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
558 llvm_unreachable("Operation should already be optimised!");
566 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
568 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
577 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
579 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
581 case ISD::SETCC_INVALID:
582 llvm_unreachable("Invalid setcc condcode!");
587 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
588 SelectionDAG &DAG) const {
589 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
590 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
591 EVT EltVT = Op.getValueType().getVectorElementType();
592 EVT PtrVT = Load->getBasePtr().getValueType();
593 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
594 SmallVector<SDValue, 8> Loads;
597 for (unsigned i = 0, e = NumElts; i != e; ++i) {
598 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
599 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
600 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
601 Load->getChain(), Ptr,
602 MachinePointerInfo(Load->getMemOperand()->getValue()),
603 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
604 Load->getAlignment()));
606 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), &Loads[0],
610 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
611 SelectionDAG &DAG) const {
612 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
613 EVT MemVT = Store->getMemoryVT();
614 unsigned MemBits = MemVT.getSizeInBits();
616 // Byte stores are really expensive, so if possible, try to pack
617 // 32-bit vector truncatating store into an i32 store.
618 // XXX: We could also handle optimize other vector bitwidths
619 if (!MemVT.isVector() || MemBits > 32) {
624 const SDValue &Value = Store->getValue();
625 EVT VT = Value.getValueType();
626 const SDValue &Ptr = Store->getBasePtr();
627 EVT MemEltVT = MemVT.getVectorElementType();
628 unsigned MemEltBits = MemEltVT.getSizeInBits();
629 unsigned MemNumElements = MemVT.getVectorNumElements();
630 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits());
634 Mask = DAG.getConstant(0xFF, PackedVT);
637 Mask = DAG.getConstant(0xFFFF, PackedVT);
640 llvm_unreachable("Cannot lower this vector store");
643 for (unsigned i = 0; i < MemNumElements; ++i) {
644 EVT ElemVT = VT.getVectorElementType();
645 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
646 DAG.getConstant(i, MVT::i32));
647 Elt = DAG.getZExtOrTrunc(Elt, DL, PackedVT);
648 Elt = DAG.getNode(ISD::AND, DL, PackedVT, Elt, Mask);
649 SDValue Shift = DAG.getConstant(MemEltBits * i, PackedVT);
650 Elt = DAG.getNode(ISD::SHL, DL, PackedVT, Elt, Shift);
654 PackedValue = DAG.getNode(ISD::OR, DL, PackedVT, PackedValue, Elt);
657 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
658 MachinePointerInfo(Store->getMemOperand()->getValue()),
659 Store->isVolatile(), Store->isNonTemporal(),
660 Store->getAlignment());
663 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
664 SelectionDAG &DAG) const {
665 StoreSDNode *Store = cast<StoreSDNode>(Op);
666 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
667 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
668 EVT PtrVT = Store->getBasePtr().getValueType();
669 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
672 SmallVector<SDValue, 8> Chains;
674 for (unsigned i = 0, e = NumElts; i != e; ++i) {
675 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
676 Store->getValue(), DAG.getConstant(i, MVT::i32));
677 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
679 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
681 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
682 MachinePointerInfo(Store->getMemOperand()->getValue()),
683 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
684 Store->getAlignment()));
686 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, &Chains[0], NumElts);
689 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
691 LoadSDNode *Load = cast<LoadSDNode>(Op);
692 ISD::LoadExtType ExtType = Load->getExtensionType();
694 // Lower loads constant address space global variable loads
695 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
696 isa<GlobalVariable>(GetUnderlyingObject(Load->getPointerInfo().V))) {
698 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
699 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
700 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
701 DAG.getConstant(2, MVT::i32));
702 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
703 Load->getChain(), Ptr,
704 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
707 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
708 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
712 EVT VT = Op.getValueType();
713 EVT MemVT = Load->getMemoryVT();
715 if (Load->getMemoryVT() == MVT::i8) {
717 } else if (Load->getMemoryVT() == MVT::i16) {
720 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
721 DAG.getConstant(2, MVT::i32));
722 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
723 Load->getChain(), Ptr,
724 DAG.getTargetConstant(0, MVT::i32),
726 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
728 DAG.getConstant(0x3, MVT::i32));
729 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
730 DAG.getConstant(3, MVT::i32));
731 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
732 Ret = DAG.getNode(ISD::AND, DL, MVT::i32, Ret,
733 DAG.getConstant(Mask, MVT::i32));
734 if (ExtType == ISD::SEXTLOAD) {
735 SDValue SExtShift = DAG.getConstant(
736 VT.getSizeInBits() - MemVT.getSizeInBits(), MVT::i32);
737 Ret = DAG.getNode(ISD::SHL, DL, MVT::i32, Ret, SExtShift);
738 Ret = DAG.getNode(ISD::SRA, DL, MVT::i32, Ret, SExtShift);
744 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
746 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
747 if (Result.getNode()) {
751 StoreSDNode *Store = cast<StoreSDNode>(Op);
752 SDValue Chain = Store->getChain();
753 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
754 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
755 Store->getValue().getValueType().isVector()) {
756 return SplitVectorStore(Op, DAG);
759 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
760 Store->getMemoryVT().bitsLT(MVT::i32)) {
762 if (Store->getMemoryVT() == MVT::i8) {
764 } else if (Store->getMemoryVT() == MVT::i16) {
767 SDValue TruncPtr = DAG.getZExtOrTrunc(Store->getBasePtr(), DL, MVT::i32);
768 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, TruncPtr,
769 DAG.getConstant(2, MVT::i32));
770 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
771 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
772 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, TruncPtr,
773 DAG.getConstant(0x3, MVT::i32));
774 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
775 DAG.getConstant(3, MVT::i32));
776 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
778 SDValue MaskedValue = DAG.getNode(ISD::AND, DL, MVT::i32, SExtValue,
779 DAG.getConstant(Mask, MVT::i32));
780 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
781 MaskedValue, ShiftAmt);
782 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
784 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
785 DAG.getConstant(0xffffffff, MVT::i32));
786 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
788 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
789 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
790 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
795 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
796 SelectionDAG &DAG) const {
798 EVT VT = Op.getValueType();
800 SDValue Num = Op.getOperand(0);
801 SDValue Den = Op.getOperand(1);
803 SmallVector<SDValue, 8> Results;
805 // RCP = URECIP(Den) = 2^32 / Den + e
806 // e is rounding error.
807 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
809 // RCP_LO = umulo(RCP, Den) */
810 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
812 // RCP_HI = mulhu (RCP, Den) */
813 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
815 // NEG_RCP_LO = -RCP_LO
816 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
819 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
820 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
823 // Calculate the rounding error from the URECIP instruction
824 // E = mulhu(ABS_RCP_LO, RCP)
825 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
828 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
831 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
833 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
834 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
837 // Quotient = mulhu(Tmp0, Num)
838 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
840 // Num_S_Remainder = Quotient * Den
841 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
843 // Remainder = Num - Num_S_Remainder
844 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
846 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
847 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
848 DAG.getConstant(-1, VT),
849 DAG.getConstant(0, VT),
851 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
852 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
854 DAG.getConstant(-1, VT),
855 DAG.getConstant(0, VT),
857 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
858 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
861 // Calculate Division result:
863 // Quotient_A_One = Quotient + 1
864 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
865 DAG.getConstant(1, VT));
867 // Quotient_S_One = Quotient - 1
868 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
869 DAG.getConstant(1, VT));
871 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
872 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
873 Quotient, Quotient_A_One, ISD::SETEQ);
875 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
876 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
877 Quotient_S_One, Div, ISD::SETEQ);
879 // Calculate Rem result:
881 // Remainder_S_Den = Remainder - Den
882 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
884 // Remainder_A_Den = Remainder + Den
885 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
887 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
888 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
889 Remainder, Remainder_S_Den, ISD::SETEQ);
891 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
892 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
893 Remainder_A_Den, Rem, ISD::SETEQ);
897 return DAG.getMergeValues(Ops, 2, DL);
900 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
901 SelectionDAG &DAG) const {
902 SDValue S0 = Op.getOperand(0);
904 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
907 // f32 uint_to_fp i64
908 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
909 DAG.getConstant(0, MVT::i32));
910 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
911 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
912 DAG.getConstant(1, MVT::i32));
913 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
914 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
915 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
916 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
920 //===----------------------------------------------------------------------===//
922 //===----------------------------------------------------------------------===//
924 void AMDGPUTargetLowering::getOriginalFunctionArgs(
927 const SmallVectorImpl<ISD::InputArg> &Ins,
928 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
930 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
931 if (Ins[i].ArgVT == Ins[i].VT) {
932 OrigIns.push_back(Ins[i]);
937 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
938 // Vector has been split into scalars.
939 VT = Ins[i].ArgVT.getVectorElementType();
940 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
941 Ins[i].ArgVT.getVectorElementType() !=
942 Ins[i].VT.getVectorElementType()) {
943 // Vector elements have been promoted
946 // Vector has been spilt into smaller vectors.
950 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
951 Ins[i].OrigArgIndex, Ins[i].PartOffset);
952 OrigIns.push_back(Arg);
956 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
957 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
958 return CFP->isExactlyValue(1.0);
960 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
961 return C->isAllOnesValue();
966 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
967 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
968 return CFP->getValueAPF().isZero();
970 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
971 return C->isNullValue();
976 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
977 const TargetRegisterClass *RC,
978 unsigned Reg, EVT VT) const {
979 MachineFunction &MF = DAG.getMachineFunction();
980 MachineRegisterInfo &MRI = MF.getRegInfo();
981 unsigned VirtualRegister;
982 if (!MRI.isLiveIn(Reg)) {
983 VirtualRegister = MRI.createVirtualRegister(RC);
984 MRI.addLiveIn(Reg, VirtualRegister);
986 VirtualRegister = MRI.getLiveInVirtReg(Reg);
988 return DAG.getRegister(VirtualRegister, VT);
991 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
993 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
997 NODE_NAME_CASE(CALL);
998 NODE_NAME_CASE(UMUL);
999 NODE_NAME_CASE(DIV_INF);
1000 NODE_NAME_CASE(RET_FLAG);
1001 NODE_NAME_CASE(BRANCH_COND);
1004 NODE_NAME_CASE(DWORDADDR)
1005 NODE_NAME_CASE(FRACT)
1006 NODE_NAME_CASE(FMAX)
1007 NODE_NAME_CASE(SMAX)
1008 NODE_NAME_CASE(UMAX)
1009 NODE_NAME_CASE(FMIN)
1010 NODE_NAME_CASE(SMIN)
1011 NODE_NAME_CASE(UMIN)
1012 NODE_NAME_CASE(URECIP)
1013 NODE_NAME_CASE(DOT4)
1014 NODE_NAME_CASE(EXPORT)
1015 NODE_NAME_CASE(CONST_ADDRESS)
1016 NODE_NAME_CASE(REGISTER_LOAD)
1017 NODE_NAME_CASE(REGISTER_STORE)
1018 NODE_NAME_CASE(LOAD_CONSTANT)
1019 NODE_NAME_CASE(LOAD_INPUT)
1020 NODE_NAME_CASE(SAMPLE)
1021 NODE_NAME_CASE(SAMPLEB)
1022 NODE_NAME_CASE(SAMPLED)
1023 NODE_NAME_CASE(SAMPLEL)
1024 NODE_NAME_CASE(STORE_MSKOR)
1025 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)