1 //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief This is the parent TargetLowering class for hardware code gen
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUISelLowering.h"
18 #include "AMDGPUFrameLowering.h"
19 #include "AMDGPURegisterInfo.h"
20 #include "AMDGPUSubtarget.h"
21 #include "AMDILIntrinsicInfo.h"
22 #include "R600MachineFunctionInfo.h"
23 #include "SIMachineFunctionInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/DiagnosticPrinter.h"
38 /// Diagnostic information for unimplemented or unsupported feature reporting.
39 class DiagnosticInfoUnsupported : public DiagnosticInfo {
41 const Twine &Description;
46 static int getKindID() {
48 KindID = llvm::getNextAvailablePluginDiagnosticKind();
53 DiagnosticInfoUnsupported(const Function &Fn, const Twine &Desc,
54 DiagnosticSeverity Severity = DS_Error)
55 : DiagnosticInfo(getKindID(), Severity),
59 const Function &getFunction() const { return Fn; }
60 const Twine &getDescription() const { return Description; }
62 void print(DiagnosticPrinter &DP) const override {
63 DP << "unsupported " << getDescription() << " in " << Fn.getName();
66 static bool classof(const DiagnosticInfo *DI) {
67 return DI->getKind() == getKindID();
71 int DiagnosticInfoUnsupported::KindID = 0;
75 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
76 CCValAssign::LocInfo LocInfo,
77 ISD::ArgFlagsTy ArgFlags, CCState &State) {
78 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
79 ArgFlags.getOrigAlign());
80 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
85 #include "AMDGPUGenCallingConv.inc"
87 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
88 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
90 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
92 // Initialize target lowering borrowed from AMDIL
95 // We need to custom lower some of the intrinsics
96 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
98 // Library functions. These default to Expand, but we have instructions
100 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
101 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
102 setOperationAction(ISD::FPOW, MVT::f32, Legal);
103 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
104 setOperationAction(ISD::FABS, MVT::f32, Legal);
105 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
106 setOperationAction(ISD::FRINT, MVT::f32, Legal);
107 setOperationAction(ISD::FROUND, MVT::f32, Legal);
108 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
110 // The hardware supports ROTR, but not ROTL
111 setOperationAction(ISD::ROTL, MVT::i32, Expand);
113 // Lower floating point store/load to integer store/load to reduce the number
114 // of patterns in tablegen.
115 setOperationAction(ISD::STORE, MVT::f32, Promote);
116 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
118 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
119 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
121 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
122 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
124 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
125 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
127 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
128 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
130 setOperationAction(ISD::STORE, MVT::f64, Promote);
131 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
133 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
134 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
136 // Custom lowering of vector stores is required for local address space
138 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
139 // XXX: Native v2i32 local address space stores are possible, but not
140 // currently implemented.
141 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
143 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
144 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
145 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
147 // XXX: This can be change to Custom, once ExpandVectorStores can
148 // handle 64-bit stores.
149 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
151 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
152 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
153 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
154 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
155 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
158 setOperationAction(ISD::LOAD, MVT::f32, Promote);
159 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
161 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
162 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
164 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
165 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
167 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
168 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
170 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
171 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
173 setOperationAction(ISD::LOAD, MVT::f64, Promote);
174 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
176 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
177 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
179 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
180 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
181 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
182 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
183 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
184 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
185 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
186 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
187 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
188 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
190 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
191 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
192 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
193 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
194 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
195 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
196 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
197 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
198 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
199 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
200 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
201 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
203 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
205 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
207 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
209 setOperationAction(ISD::MUL, MVT::i64, Expand);
210 setOperationAction(ISD::SUB, MVT::i64, Expand);
212 setOperationAction(ISD::UDIV, MVT::i32, Expand);
213 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
214 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
215 setOperationAction(ISD::UREM, MVT::i32, Expand);
217 static const MVT::SimpleValueType IntTypes[] = {
218 MVT::v2i32, MVT::v4i32
221 for (MVT VT : IntTypes) {
222 //Expand the following operations for the current type by default
223 setOperationAction(ISD::ADD, VT, Expand);
224 setOperationAction(ISD::AND, VT, Expand);
225 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
226 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
227 setOperationAction(ISD::MUL, VT, Expand);
228 setOperationAction(ISD::OR, VT, Expand);
229 setOperationAction(ISD::SHL, VT, Expand);
230 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
231 setOperationAction(ISD::SRL, VT, Expand);
232 setOperationAction(ISD::SRA, VT, Expand);
233 setOperationAction(ISD::SUB, VT, Expand);
234 setOperationAction(ISD::UDIV, VT, Expand);
235 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
236 setOperationAction(ISD::UREM, VT, Expand);
237 setOperationAction(ISD::SELECT, VT, Expand);
238 setOperationAction(ISD::VSELECT, VT, Expand);
239 setOperationAction(ISD::XOR, VT, Expand);
242 static const MVT::SimpleValueType FloatTypes[] = {
243 MVT::v2f32, MVT::v4f32
246 for (MVT VT : FloatTypes) {
247 setOperationAction(ISD::FABS, VT, Expand);
248 setOperationAction(ISD::FADD, VT, Expand);
249 setOperationAction(ISD::FCOS, VT, Expand);
250 setOperationAction(ISD::FDIV, VT, Expand);
251 setOperationAction(ISD::FPOW, VT, Expand);
252 setOperationAction(ISD::FFLOOR, VT, Expand);
253 setOperationAction(ISD::FTRUNC, VT, Expand);
254 setOperationAction(ISD::FMUL, VT, Expand);
255 setOperationAction(ISD::FRINT, VT, Expand);
256 setOperationAction(ISD::FSQRT, VT, Expand);
257 setOperationAction(ISD::FSIN, VT, Expand);
258 setOperationAction(ISD::FSUB, VT, Expand);
259 setOperationAction(ISD::FNEG, VT, Expand);
260 setOperationAction(ISD::SELECT, VT, Expand);
261 setOperationAction(ISD::VSELECT, VT, Expand);
264 setTargetDAGCombine(ISD::MUL);
265 setTargetDAGCombine(ISD::SELECT_CC);
268 //===----------------------------------------------------------------------===//
269 // Target Information
270 //===----------------------------------------------------------------------===//
272 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
276 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
278 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
281 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
282 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
284 return ((LScalarSize <= CastScalarSize) ||
285 (CastScalarSize >= 32) ||
289 //===---------------------------------------------------------------------===//
291 //===---------------------------------------------------------------------===//
293 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
294 assert(VT.isFloatingPoint());
295 return VT == MVT::f32;
298 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
299 assert(VT.isFloatingPoint());
300 return VT == MVT::f32;
303 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
304 // Truncate is just accessing a subregister.
305 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
308 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
309 // Truncate is just accessing a subregister.
310 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
311 (Dest->getPrimitiveSizeInBits() % 32 == 0);
314 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
315 const DataLayout *DL = getDataLayout();
316 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
317 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
319 return SrcSize == 32 && DestSize == 64;
322 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
323 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
324 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
325 // this will enable reducing 64-bit operations the 32-bit, which is always
327 return Src == MVT::i32 && Dest == MVT::i64;
330 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
331 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
332 // limited number of native 64-bit operations. Shrinking an operation to fit
333 // in a single 32-bit register should always be helpful. As currently used,
334 // this is much less general than the name suggests, and is only used in
335 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
336 // not profitable, and may actually be harmful.
337 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
340 //===---------------------------------------------------------------------===//
341 // TargetLowering Callbacks
342 //===---------------------------------------------------------------------===//
344 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
345 const SmallVectorImpl<ISD::InputArg> &Ins) const {
347 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
350 SDValue AMDGPUTargetLowering::LowerReturn(
352 CallingConv::ID CallConv,
354 const SmallVectorImpl<ISD::OutputArg> &Outs,
355 const SmallVectorImpl<SDValue> &OutVals,
356 SDLoc DL, SelectionDAG &DAG) const {
357 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
360 //===---------------------------------------------------------------------===//
361 // Target specific lowering
362 //===---------------------------------------------------------------------===//
364 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
365 SmallVectorImpl<SDValue> &InVals) const {
366 SDValue Callee = CLI.Callee;
367 SelectionDAG &DAG = CLI.DAG;
369 const Function &Fn = *DAG.getMachineFunction().getFunction();
371 StringRef FuncName("<unknown>");
373 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
374 FuncName = G->getSymbol();
375 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
376 FuncName = G->getGlobal()->getName();
378 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
379 DAG.getContext()->diagnose(NoCalls);
383 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
385 switch (Op.getOpcode()) {
387 Op.getNode()->dump();
388 llvm_unreachable("Custom lowering code for this"
389 "instruction is not implemented yet!");
391 // AMDIL DAG lowering
392 case ISD::SDIV: return LowerSDIV(Op, DAG);
393 case ISD::SREM: return LowerSREM(Op, DAG);
394 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
395 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
396 // AMDGPU DAG lowering
397 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
398 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
399 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
400 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
401 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
402 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
407 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
408 SmallVectorImpl<SDValue> &Results,
409 SelectionDAG &DAG) const {
410 switch (N->getOpcode()) {
411 case ISD::SIGN_EXTEND_INREG:
412 // Different parts of legalization seem to interpret which type of
413 // sign_extend_inreg is the one to check for custom lowering. The extended
414 // from type is what really matters, but some places check for custom
415 // lowering of the result type. This results in trying to use
416 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
417 // nothing here and let the illegal result integer be handled normally.
420 SDValue Op = SDValue(N, 0);
422 EVT VT = Op.getValueType();
423 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
424 N->getOperand(0), N->getOperand(1));
425 Results.push_back(UDIVREM);
429 SDValue Op = SDValue(N, 0);
431 EVT VT = Op.getValueType();
432 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
433 N->getOperand(0), N->getOperand(1));
434 Results.push_back(UDIVREM.getValue(1));
438 SDValue Op = SDValue(N, 0);
440 EVT VT = Op.getValueType();
441 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
443 SDValue one = DAG.getConstant(1, HalfVT);
444 SDValue zero = DAG.getConstant(0, HalfVT);
447 SDValue LHS = N->getOperand(0);
448 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
449 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
451 SDValue RHS = N->getOperand(1);
452 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
453 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
455 // Get Speculative values
456 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
457 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
459 SDValue REM_Hi = zero;
460 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
462 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
463 SDValue DIV_Lo = zero;
465 const unsigned halfBitWidth = HalfVT.getSizeInBits();
467 for (unsigned i = 0; i < halfBitWidth; ++i) {
468 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
469 // Get Value of high bit
471 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
472 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
474 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
475 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
478 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
479 DAG.getConstant(halfBitWidth - 1, HalfVT));
480 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
481 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
483 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
484 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
487 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
489 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
490 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
492 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
496 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
498 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
499 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
500 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
503 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
504 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
505 Results.push_back(DIV);
506 Results.push_back(REM);
514 // FIXME: This implements accesses to initialized globals in the constant
515 // address space by copying them to private and accessing that. It does not
516 // properly handle illegal types or vectors. The private vector loads are not
517 // scalarized, and the illegal scalars hit an assertion. This technique will not
518 // work well with large initializers, and this should eventually be
519 // removed. Initialized globals should be placed into a data section that the
520 // runtime will load into a buffer before the kernel is executed. Uses of the
521 // global need to be replaced with a pointer loaded from an implicit kernel
522 // argument into this buffer holding the copy of the data, which will remove the
523 // need for any of this.
524 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
525 const GlobalValue *GV,
526 const SDValue &InitPtr,
528 SelectionDAG &DAG) const {
529 const DataLayout *TD = getTargetMachine().getDataLayout();
531 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
532 EVT VT = EVT::getEVT(CI->getType());
533 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
534 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
535 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
536 TD->getPrefTypeAlignment(CI->getType()));
539 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
540 EVT VT = EVT::getEVT(CFP->getType());
541 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
542 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
543 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
544 TD->getPrefTypeAlignment(CFP->getType()));
547 Type *InitTy = Init->getType();
548 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
549 const StructLayout *SL = TD->getStructLayout(ST);
551 EVT PtrVT = InitPtr.getValueType();
552 SmallVector<SDValue, 8> Chains;
554 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
555 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
556 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
558 Constant *Elt = Init->getAggregateElement(I);
559 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
562 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
565 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
566 EVT PtrVT = InitPtr.getValueType();
568 unsigned NumElements;
569 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
570 NumElements = AT->getNumElements();
571 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
572 NumElements = VT->getNumElements();
574 llvm_unreachable("Unexpected type");
576 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
577 SmallVector<SDValue, 8> Chains;
578 for (unsigned i = 0; i < NumElements; ++i) {
579 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
580 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
582 Constant *Elt = Init->getAggregateElement(i);
583 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
586 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
590 llvm_unreachable("Unhandled constant initializer");
593 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
595 SelectionDAG &DAG) const {
597 const DataLayout *TD = getTargetMachine().getDataLayout();
598 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
599 const GlobalValue *GV = G->getGlobal();
601 switch (G->getAddressSpace()) {
602 default: llvm_unreachable("Global Address lowering not implemented for this "
604 case AMDGPUAS::LOCAL_ADDRESS: {
605 // XXX: What does the value of G->getOffset() mean?
606 assert(G->getOffset() == 0 &&
607 "Do not know what to do with an non-zero offset");
610 if (MFI->LocalMemoryObjects.count(GV) == 0) {
611 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
612 Offset = MFI->LDSSize;
613 MFI->LocalMemoryObjects[GV] = Offset;
614 // XXX: Account for alignment?
615 MFI->LDSSize += Size;
617 Offset = MFI->LocalMemoryObjects[GV];
620 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
622 case AMDGPUAS::CONSTANT_ADDRESS: {
623 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
624 Type *EltType = GV->getType()->getElementType();
625 unsigned Size = TD->getTypeAllocSize(EltType);
626 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
628 const GlobalVariable *Var = cast<GlobalVariable>(GV);
629 const Constant *Init = Var->getInitializer();
630 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
631 SDValue InitPtr = DAG.getFrameIndex(FI,
632 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
633 SmallVector<SDNode*, 8> WorkList;
635 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
636 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
637 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
639 WorkList.push_back(*I);
641 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
642 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
643 E = WorkList.end(); I != E; ++I) {
644 SmallVector<SDValue, 8> Ops;
645 Ops.push_back(Chain);
646 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
647 Ops.push_back((*I)->getOperand(i));
649 DAG.UpdateNodeOperands(*I, Ops);
651 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
652 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
657 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
658 SelectionDAG &DAG) const {
659 SmallVector<SDValue, 8> Args;
660 SDValue A = Op.getOperand(0);
661 SDValue B = Op.getOperand(1);
663 DAG.ExtractVectorElements(A, Args);
664 DAG.ExtractVectorElements(B, Args);
666 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
669 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
670 SelectionDAG &DAG) const {
672 SmallVector<SDValue, 8> Args;
673 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
674 EVT VT = Op.getValueType();
675 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
676 VT.getVectorNumElements());
678 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
681 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
682 SelectionDAG &DAG) const {
684 MachineFunction &MF = DAG.getMachineFunction();
685 const AMDGPUFrameLowering *TFL =
686 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
688 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
691 unsigned FrameIndex = FIN->getIndex();
692 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
693 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
697 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
698 SelectionDAG &DAG) const {
699 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
701 EVT VT = Op.getValueType();
703 switch (IntrinsicID) {
705 case AMDGPUIntrinsic::AMDIL_abs:
706 return LowerIntrinsicIABS(Op, DAG);
707 case AMDGPUIntrinsic::AMDIL_exp:
708 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
709 case AMDGPUIntrinsic::AMDGPU_lrp:
710 return LowerIntrinsicLRP(Op, DAG);
711 case AMDGPUIntrinsic::AMDIL_fraction:
712 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
713 case AMDGPUIntrinsic::AMDIL_max:
714 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
716 case AMDGPUIntrinsic::AMDGPU_imax:
717 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
719 case AMDGPUIntrinsic::AMDGPU_umax:
720 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
722 case AMDGPUIntrinsic::AMDIL_min:
723 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
725 case AMDGPUIntrinsic::AMDGPU_imin:
726 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
728 case AMDGPUIntrinsic::AMDGPU_umin:
729 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
732 case AMDGPUIntrinsic::AMDGPU_umul24:
733 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
734 Op.getOperand(1), Op.getOperand(2));
736 case AMDGPUIntrinsic::AMDGPU_imul24:
737 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
738 Op.getOperand(1), Op.getOperand(2));
740 case AMDGPUIntrinsic::AMDGPU_umad24:
741 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
742 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
744 case AMDGPUIntrinsic::AMDGPU_imad24:
745 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
746 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
748 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
749 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
754 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
755 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
760 case AMDGPUIntrinsic::AMDGPU_bfi:
761 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
766 case AMDGPUIntrinsic::AMDGPU_bfm:
767 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
771 case AMDGPUIntrinsic::AMDIL_round_nearest:
772 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
776 ///IABS(a) = SMAX(sub(0, a), a)
777 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
778 SelectionDAG &DAG) const {
780 EVT VT = Op.getValueType();
781 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
784 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
787 /// Linear Interpolation
788 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
789 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
790 SelectionDAG &DAG) const {
792 EVT VT = Op.getValueType();
793 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
794 DAG.getConstantFP(1.0f, MVT::f32),
796 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
798 return DAG.getNode(ISD::FADD, DL, VT,
799 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
803 /// \brief Generate Min/Max node
804 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
805 SelectionDAG &DAG) const {
807 EVT VT = N->getValueType(0);
809 SDValue LHS = N->getOperand(0);
810 SDValue RHS = N->getOperand(1);
811 SDValue True = N->getOperand(2);
812 SDValue False = N->getOperand(3);
813 SDValue CC = N->getOperand(4);
815 if (VT != MVT::f32 ||
816 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
820 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
834 llvm_unreachable("Operation should already be optimised!");
841 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
842 return DAG.getNode(Opc, DL, VT, LHS, RHS);
850 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
851 return DAG.getNode(Opc, DL, VT, LHS, RHS);
853 case ISD::SETCC_INVALID:
854 llvm_unreachable("Invalid setcc condcode!");
859 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
860 SelectionDAG &DAG) const {
861 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
862 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
863 EVT EltVT = Op.getValueType().getVectorElementType();
864 EVT PtrVT = Load->getBasePtr().getValueType();
865 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
866 SmallVector<SDValue, 8> Loads;
869 for (unsigned i = 0, e = NumElts; i != e; ++i) {
870 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
871 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
872 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
873 Load->getChain(), Ptr,
874 MachinePointerInfo(Load->getMemOperand()->getValue()),
875 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
876 Load->getAlignment()));
878 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
881 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
882 SelectionDAG &DAG) const {
883 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
884 EVT MemVT = Store->getMemoryVT();
885 unsigned MemBits = MemVT.getSizeInBits();
887 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
888 // truncating store into an i32 store.
889 // XXX: We could also handle optimize other vector bitwidths.
890 if (!MemVT.isVector() || MemBits > 32) {
895 SDValue Value = Store->getValue();
896 EVT VT = Value.getValueType();
897 EVT ElemVT = VT.getVectorElementType();
898 SDValue Ptr = Store->getBasePtr();
899 EVT MemEltVT = MemVT.getVectorElementType();
900 unsigned MemEltBits = MemEltVT.getSizeInBits();
901 unsigned MemNumElements = MemVT.getVectorNumElements();
902 unsigned PackedSize = MemVT.getStoreSizeInBits();
903 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
905 assert(Value.getValueType().getScalarSizeInBits() >= 32);
908 for (unsigned i = 0; i < MemNumElements; ++i) {
909 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
910 DAG.getConstant(i, MVT::i32));
911 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
912 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
914 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
915 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
920 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
924 if (PackedSize < 32) {
925 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
926 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
927 Store->getMemOperand()->getPointerInfo(),
929 Store->isNonTemporal(), Store->isVolatile(),
930 Store->getAlignment());
933 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
934 Store->getMemOperand()->getPointerInfo(),
935 Store->isVolatile(), Store->isNonTemporal(),
936 Store->getAlignment());
939 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
940 SelectionDAG &DAG) const {
941 StoreSDNode *Store = cast<StoreSDNode>(Op);
942 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
943 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
944 EVT PtrVT = Store->getBasePtr().getValueType();
945 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
948 SmallVector<SDValue, 8> Chains;
950 for (unsigned i = 0, e = NumElts; i != e; ++i) {
951 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
952 Store->getValue(), DAG.getConstant(i, MVT::i32));
953 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
955 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
957 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
958 MachinePointerInfo(Store->getMemOperand()->getValue()),
959 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
960 Store->getAlignment()));
962 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
965 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
967 LoadSDNode *Load = cast<LoadSDNode>(Op);
968 ISD::LoadExtType ExtType = Load->getExtensionType();
969 EVT VT = Op.getValueType();
970 EVT MemVT = Load->getMemoryVT();
972 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
973 // We can do the extload to 32-bits, and then need to separately extend to
976 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
980 Load->getMemOperand());
981 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
984 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
985 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
986 // FIXME: Copied from PPC
987 // First, load into 32 bits, then truncate to 1 bit.
989 SDValue Chain = Load->getChain();
990 SDValue BasePtr = Load->getBasePtr();
991 MachineMemOperand *MMO = Load->getMemOperand();
993 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
994 BasePtr, MVT::i8, MMO);
995 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
998 // Lower loads constant address space global variable loads
999 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1000 isa<GlobalVariable>(
1001 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1003 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1004 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1005 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1006 DAG.getConstant(2, MVT::i32));
1007 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1008 Load->getChain(), Ptr,
1009 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1012 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1013 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1017 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1018 DAG.getConstant(2, MVT::i32));
1019 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1020 Load->getChain(), Ptr,
1021 DAG.getTargetConstant(0, MVT::i32),
1023 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1025 DAG.getConstant(0x3, MVT::i32));
1026 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1027 DAG.getConstant(3, MVT::i32));
1029 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1031 EVT MemEltVT = MemVT.getScalarType();
1032 if (ExtType == ISD::SEXTLOAD) {
1033 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1034 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1037 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1040 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1042 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1043 if (Result.getNode()) {
1047 StoreSDNode *Store = cast<StoreSDNode>(Op);
1048 SDValue Chain = Store->getChain();
1049 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1050 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1051 Store->getValue().getValueType().isVector()) {
1052 return SplitVectorStore(Op, DAG);
1055 EVT MemVT = Store->getMemoryVT();
1056 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1057 MemVT.bitsLT(MVT::i32)) {
1059 if (Store->getMemoryVT() == MVT::i8) {
1061 } else if (Store->getMemoryVT() == MVT::i16) {
1064 SDValue BasePtr = Store->getBasePtr();
1065 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1066 DAG.getConstant(2, MVT::i32));
1067 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1068 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1070 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1071 DAG.getConstant(0x3, MVT::i32));
1073 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1074 DAG.getConstant(3, MVT::i32));
1076 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1079 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1081 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1082 MaskedValue, ShiftAmt);
1084 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1086 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1087 DAG.getConstant(0xffffffff, MVT::i32));
1088 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1090 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1091 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1092 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1097 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1098 SelectionDAG &DAG) const {
1100 EVT VT = Op.getValueType();
1102 SDValue Num = Op.getOperand(0);
1103 SDValue Den = Op.getOperand(1);
1105 // RCP = URECIP(Den) = 2^32 / Den + e
1106 // e is rounding error.
1107 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1109 // RCP_LO = umulo(RCP, Den) */
1110 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1112 // RCP_HI = mulhu (RCP, Den) */
1113 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1115 // NEG_RCP_LO = -RCP_LO
1116 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1119 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1120 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1123 // Calculate the rounding error from the URECIP instruction
1124 // E = mulhu(ABS_RCP_LO, RCP)
1125 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1127 // RCP_A_E = RCP + E
1128 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1130 // RCP_S_E = RCP - E
1131 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1133 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1134 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1137 // Quotient = mulhu(Tmp0, Num)
1138 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1140 // Num_S_Remainder = Quotient * Den
1141 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1143 // Remainder = Num - Num_S_Remainder
1144 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1146 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1147 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1148 DAG.getConstant(-1, VT),
1149 DAG.getConstant(0, VT),
1151 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1152 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1154 DAG.getConstant(-1, VT),
1155 DAG.getConstant(0, VT),
1157 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1158 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1161 // Calculate Division result:
1163 // Quotient_A_One = Quotient + 1
1164 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1165 DAG.getConstant(1, VT));
1167 // Quotient_S_One = Quotient - 1
1168 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1169 DAG.getConstant(1, VT));
1171 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1172 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1173 Quotient, Quotient_A_One, ISD::SETEQ);
1175 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1176 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1177 Quotient_S_One, Div, ISD::SETEQ);
1179 // Calculate Rem result:
1181 // Remainder_S_Den = Remainder - Den
1182 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1184 // Remainder_A_Den = Remainder + Den
1185 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1187 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1188 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1189 Remainder, Remainder_S_Den, ISD::SETEQ);
1191 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1192 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1193 Remainder_A_Den, Rem, ISD::SETEQ);
1198 return DAG.getMergeValues(Ops, DL);
1201 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1202 SelectionDAG &DAG) const {
1203 SDValue S0 = Op.getOperand(0);
1205 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1208 // f32 uint_to_fp i64
1209 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1210 DAG.getConstant(0, MVT::i32));
1211 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1212 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1213 DAG.getConstant(1, MVT::i32));
1214 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1215 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1216 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1217 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1221 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1223 SelectionDAG &DAG) const {
1224 MVT VT = Op.getSimpleValueType();
1226 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1227 // Shift left by 'Shift' bits.
1228 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1229 // Signed shift Right by 'Shift' bits.
1230 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1233 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1234 SelectionDAG &DAG) const {
1235 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1236 MVT VT = Op.getSimpleValueType();
1237 MVT ScalarVT = VT.getScalarType();
1242 SDValue Src = Op.getOperand(0);
1245 // TODO: Don't scalarize on Evergreen?
1246 unsigned NElts = VT.getVectorNumElements();
1247 SmallVector<SDValue, 8> Args;
1248 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1250 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1251 for (unsigned I = 0; I < NElts; ++I)
1252 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1254 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1257 //===----------------------------------------------------------------------===//
1258 // Custom DAG optimizations
1259 //===----------------------------------------------------------------------===//
1261 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1262 APInt KnownZero, KnownOne;
1263 EVT VT = Op.getValueType();
1264 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1266 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1269 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1270 EVT VT = Op.getValueType();
1272 // In order for this to be a signed 24-bit value, bit 23, must
1274 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1275 // as unsigned 24-bit values.
1276 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1279 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1281 SelectionDAG &DAG = DCI.DAG;
1282 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1283 EVT VT = Op.getValueType();
1285 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1286 APInt KnownZero, KnownOne;
1287 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1288 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1289 DCI.CommitTargetLoweringOpt(TLO);
1292 template <typename IntTy>
1293 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1294 uint32_t Offset, uint32_t Width) {
1295 if (Width + Offset < 32) {
1296 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1297 return DAG.getConstant(Result, MVT::i32);
1300 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1303 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1304 DAGCombinerInfo &DCI) const {
1305 SelectionDAG &DAG = DCI.DAG;
1308 switch(N->getOpcode()) {
1311 EVT VT = N->getValueType(0);
1312 SDValue N0 = N->getOperand(0);
1313 SDValue N1 = N->getOperand(1);
1316 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1317 if (VT.isVector() || VT.getSizeInBits() > 32)
1320 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1321 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1322 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1323 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1324 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1325 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1326 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1327 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1332 // We need to use sext even for MUL_U24, because MUL_U24 is used
1333 // for signed multiply of 8 and 16-bit types.
1334 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1338 case AMDGPUISD::MUL_I24:
1339 case AMDGPUISD::MUL_U24: {
1340 SDValue N0 = N->getOperand(0);
1341 SDValue N1 = N->getOperand(1);
1342 simplifyI24(N0, DCI);
1343 simplifyI24(N1, DCI);
1346 case ISD::SELECT_CC: {
1347 return CombineMinMax(N, DAG);
1349 case AMDGPUISD::BFE_I32:
1350 case AMDGPUISD::BFE_U32: {
1351 assert(!N->getValueType(0).isVector() &&
1352 "Vector handling of BFE not implemented");
1353 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1357 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1359 return DAG.getConstant(0, MVT::i32);
1361 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1365 SDValue BitsFrom = N->getOperand(0);
1366 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1368 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1370 if (OffsetVal == 0) {
1371 // This is already sign / zero extended, so try to fold away extra BFEs.
1372 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1374 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1375 if (OpSignBits >= SignBits)
1378 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1380 // This is a sign_extend_inreg. Replace it to take advantage of existing
1381 // DAG Combines. If not eliminated, we will match back to BFE during
1384 // TODO: The sext_inreg of extended types ends, although we can could
1385 // handle them in a single BFE.
1386 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1387 DAG.getValueType(SmallVT));
1390 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1393 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1395 return constantFoldBFE<int32_t>(DAG,
1396 Val->getSExtValue(),
1401 return constantFoldBFE<uint32_t>(DAG,
1402 Val->getZExtValue(),
1407 APInt Demanded = APInt::getBitsSet(32,
1409 OffsetVal + WidthVal);
1411 if ((OffsetVal + WidthVal) >= 32) {
1412 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1413 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1414 BitsFrom, ShiftVal);
1417 APInt KnownZero, KnownOne;
1418 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1419 !DCI.isBeforeLegalizeOps());
1420 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1421 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1422 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1423 DCI.CommitTargetLoweringOpt(TLO);
1432 //===----------------------------------------------------------------------===//
1434 //===----------------------------------------------------------------------===//
1436 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1439 const SmallVectorImpl<ISD::InputArg> &Ins,
1440 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1442 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1443 if (Ins[i].ArgVT == Ins[i].VT) {
1444 OrigIns.push_back(Ins[i]);
1449 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1450 // Vector has been split into scalars.
1451 VT = Ins[i].ArgVT.getVectorElementType();
1452 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1453 Ins[i].ArgVT.getVectorElementType() !=
1454 Ins[i].VT.getVectorElementType()) {
1455 // Vector elements have been promoted
1458 // Vector has been spilt into smaller vectors.
1462 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1463 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1464 OrigIns.push_back(Arg);
1468 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1469 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1470 return CFP->isExactlyValue(1.0);
1472 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1473 return C->isAllOnesValue();
1478 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1479 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1480 return CFP->getValueAPF().isZero();
1482 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1483 return C->isNullValue();
1488 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1489 const TargetRegisterClass *RC,
1490 unsigned Reg, EVT VT) const {
1491 MachineFunction &MF = DAG.getMachineFunction();
1492 MachineRegisterInfo &MRI = MF.getRegInfo();
1493 unsigned VirtualRegister;
1494 if (!MRI.isLiveIn(Reg)) {
1495 VirtualRegister = MRI.createVirtualRegister(RC);
1496 MRI.addLiveIn(Reg, VirtualRegister);
1498 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1500 return DAG.getRegister(VirtualRegister, VT);
1503 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1505 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1507 default: return nullptr;
1509 NODE_NAME_CASE(CALL);
1510 NODE_NAME_CASE(UMUL);
1511 NODE_NAME_CASE(DIV_INF);
1512 NODE_NAME_CASE(RET_FLAG);
1513 NODE_NAME_CASE(BRANCH_COND);
1516 NODE_NAME_CASE(DWORDADDR)
1517 NODE_NAME_CASE(FRACT)
1518 NODE_NAME_CASE(FMAX)
1519 NODE_NAME_CASE(SMAX)
1520 NODE_NAME_CASE(UMAX)
1521 NODE_NAME_CASE(FMIN)
1522 NODE_NAME_CASE(SMIN)
1523 NODE_NAME_CASE(UMIN)
1524 NODE_NAME_CASE(BFE_U32)
1525 NODE_NAME_CASE(BFE_I32)
1528 NODE_NAME_CASE(MUL_U24)
1529 NODE_NAME_CASE(MUL_I24)
1530 NODE_NAME_CASE(MAD_U24)
1531 NODE_NAME_CASE(MAD_I24)
1532 NODE_NAME_CASE(URECIP)
1533 NODE_NAME_CASE(DOT4)
1534 NODE_NAME_CASE(EXPORT)
1535 NODE_NAME_CASE(CONST_ADDRESS)
1536 NODE_NAME_CASE(REGISTER_LOAD)
1537 NODE_NAME_CASE(REGISTER_STORE)
1538 NODE_NAME_CASE(LOAD_CONSTANT)
1539 NODE_NAME_CASE(LOAD_INPUT)
1540 NODE_NAME_CASE(SAMPLE)
1541 NODE_NAME_CASE(SAMPLEB)
1542 NODE_NAME_CASE(SAMPLED)
1543 NODE_NAME_CASE(SAMPLEL)
1544 NODE_NAME_CASE(STORE_MSKOR)
1545 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1549 static void computeKnownBitsForMinMax(const SDValue Op0,
1553 const SelectionDAG &DAG,
1555 APInt Op0Zero, Op0One;
1556 APInt Op1Zero, Op1One;
1557 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
1558 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
1560 KnownZero = Op0Zero & Op1Zero;
1561 KnownOne = Op0One & Op1One;
1564 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
1568 const SelectionDAG &DAG,
1569 unsigned Depth) const {
1571 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1575 unsigned Opc = Op.getOpcode();
1580 case ISD::INTRINSIC_WO_CHAIN: {
1581 // FIXME: The intrinsic should just use the node.
1582 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1583 case AMDGPUIntrinsic::AMDGPU_imax:
1584 case AMDGPUIntrinsic::AMDGPU_umax:
1585 case AMDGPUIntrinsic::AMDGPU_imin:
1586 case AMDGPUIntrinsic::AMDGPU_umin:
1587 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1588 KnownZero, KnownOne, DAG, Depth);
1596 case AMDGPUISD::SMAX:
1597 case AMDGPUISD::UMAX:
1598 case AMDGPUISD::SMIN:
1599 case AMDGPUISD::UMIN:
1600 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1601 KnownZero, KnownOne, DAG, Depth);
1604 case AMDGPUISD::BFE_I32:
1605 case AMDGPUISD::BFE_U32: {
1606 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1610 unsigned BitWidth = 32;
1611 uint32_t Width = CWidth->getZExtValue() & 0x1f;
1613 KnownZero = APInt::getAllOnesValue(BitWidth);
1614 KnownOne = APInt::getNullValue(BitWidth);
1618 // FIXME: This could do a lot more. If offset is 0, should be the same as
1619 // sign_extend_inreg implementation, but that involves duplicating it.
1620 if (Opc == AMDGPUISD::BFE_I32)
1621 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1623 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1630 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
1632 const SelectionDAG &DAG,
1633 unsigned Depth) const {
1634 switch (Op.getOpcode()) {
1635 case AMDGPUISD::BFE_I32: {
1636 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1640 unsigned SignBits = 32 - Width->getZExtValue() + 1;
1641 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1642 if (!Offset || !Offset->isNullValue())
1645 // TODO: Could probably figure something out with non-0 offsets.
1646 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
1647 return std::max(SignBits, Op0SignBits);
1650 case AMDGPUISD::BFE_U32: {
1651 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1652 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;