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::FNEG, MVT::v2f32, Expand);
208 setOperationAction(ISD::FNEG, MVT::v4f32, Expand);
210 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
212 setOperationAction(ISD::MUL, MVT::i64, Expand);
213 setOperationAction(ISD::SUB, MVT::i64, Expand);
215 setOperationAction(ISD::UDIV, MVT::i32, Expand);
216 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
217 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
218 setOperationAction(ISD::UREM, MVT::i32, Expand);
219 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
220 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
222 static const MVT::SimpleValueType IntTypes[] = {
223 MVT::v2i32, MVT::v4i32
226 for (MVT VT : IntTypes) {
227 //Expand the following operations for the current type by default
228 setOperationAction(ISD::ADD, VT, Expand);
229 setOperationAction(ISD::AND, VT, Expand);
230 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
231 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
232 setOperationAction(ISD::MUL, VT, Expand);
233 setOperationAction(ISD::OR, VT, Expand);
234 setOperationAction(ISD::SHL, VT, Expand);
235 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
236 setOperationAction(ISD::SRL, VT, Expand);
237 setOperationAction(ISD::SRA, VT, Expand);
238 setOperationAction(ISD::SUB, VT, Expand);
239 setOperationAction(ISD::UDIV, VT, Expand);
240 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
241 setOperationAction(ISD::UREM, VT, Expand);
242 setOperationAction(ISD::SELECT, VT, Expand);
243 setOperationAction(ISD::VSELECT, VT, Expand);
244 setOperationAction(ISD::XOR, VT, Expand);
247 static const MVT::SimpleValueType FloatTypes[] = {
248 MVT::v2f32, MVT::v4f32
251 for (MVT VT : FloatTypes) {
252 setOperationAction(ISD::FABS, VT, Expand);
253 setOperationAction(ISD::FADD, VT, Expand);
254 setOperationAction(ISD::FCOS, VT, Expand);
255 setOperationAction(ISD::FDIV, VT, Expand);
256 setOperationAction(ISD::FPOW, VT, Expand);
257 setOperationAction(ISD::FFLOOR, VT, Expand);
258 setOperationAction(ISD::FTRUNC, VT, Expand);
259 setOperationAction(ISD::FMUL, VT, Expand);
260 setOperationAction(ISD::FRINT, VT, Expand);
261 setOperationAction(ISD::FSQRT, VT, Expand);
262 setOperationAction(ISD::FSIN, VT, Expand);
263 setOperationAction(ISD::FSUB, VT, Expand);
264 setOperationAction(ISD::SELECT, VT, Expand);
267 setTargetDAGCombine(ISD::MUL);
268 setTargetDAGCombine(ISD::SELECT_CC);
271 //===----------------------------------------------------------------------===//
272 // Target Information
273 //===----------------------------------------------------------------------===//
275 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
279 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
281 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
284 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
285 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
287 return ((LScalarSize <= CastScalarSize) ||
288 (CastScalarSize >= 32) ||
292 //===---------------------------------------------------------------------===//
294 //===---------------------------------------------------------------------===//
296 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
297 assert(VT.isFloatingPoint());
298 return VT == MVT::f32;
301 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
302 assert(VT.isFloatingPoint());
303 return VT == MVT::f32;
306 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
307 // Truncate is just accessing a subregister.
308 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
311 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
312 // Truncate is just accessing a subregister.
313 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
314 (Dest->getPrimitiveSizeInBits() % 32 == 0);
317 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
318 const DataLayout *DL = getDataLayout();
319 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
320 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
322 return SrcSize == 32 && DestSize == 64;
325 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
326 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
327 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
328 // this will enable reducing 64-bit operations the 32-bit, which is always
330 return Src == MVT::i32 && Dest == MVT::i64;
333 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
334 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
335 // limited number of native 64-bit operations. Shrinking an operation to fit
336 // in a single 32-bit register should always be helpful. As currently used,
337 // this is much less general than the name suggests, and is only used in
338 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
339 // not profitable, and may actually be harmful.
340 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
343 //===---------------------------------------------------------------------===//
344 // TargetLowering Callbacks
345 //===---------------------------------------------------------------------===//
347 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
348 const SmallVectorImpl<ISD::InputArg> &Ins) const {
350 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
353 SDValue AMDGPUTargetLowering::LowerReturn(
355 CallingConv::ID CallConv,
357 const SmallVectorImpl<ISD::OutputArg> &Outs,
358 const SmallVectorImpl<SDValue> &OutVals,
359 SDLoc DL, SelectionDAG &DAG) const {
360 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
363 //===---------------------------------------------------------------------===//
364 // Target specific lowering
365 //===---------------------------------------------------------------------===//
367 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
368 SmallVectorImpl<SDValue> &InVals) const {
369 SDValue Callee = CLI.Callee;
370 SelectionDAG &DAG = CLI.DAG;
372 const Function &Fn = *DAG.getMachineFunction().getFunction();
374 StringRef FuncName("<unknown>");
376 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
377 FuncName = G->getSymbol();
378 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
379 FuncName = G->getGlobal()->getName();
381 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
382 DAG.getContext()->diagnose(NoCalls);
386 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
388 switch (Op.getOpcode()) {
390 Op.getNode()->dump();
391 llvm_unreachable("Custom lowering code for this"
392 "instruction is not implemented yet!");
394 // AMDIL DAG lowering
395 case ISD::SDIV: return LowerSDIV(Op, DAG);
396 case ISD::SREM: return LowerSREM(Op, DAG);
397 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
398 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
399 // AMDGPU DAG lowering
400 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
401 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
402 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
403 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
404 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
405 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
410 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
411 SmallVectorImpl<SDValue> &Results,
412 SelectionDAG &DAG) const {
413 switch (N->getOpcode()) {
414 case ISD::SIGN_EXTEND_INREG:
415 // Different parts of legalization seem to interpret which type of
416 // sign_extend_inreg is the one to check for custom lowering. The extended
417 // from type is what really matters, but some places check for custom
418 // lowering of the result type. This results in trying to use
419 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
420 // nothing here and let the illegal result integer be handled normally.
423 SDValue Op = SDValue(N, 0);
425 EVT VT = Op.getValueType();
426 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
427 N->getOperand(0), N->getOperand(1));
428 Results.push_back(UDIVREM);
432 SDValue Op = SDValue(N, 0);
434 EVT VT = Op.getValueType();
435 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
436 N->getOperand(0), N->getOperand(1));
437 Results.push_back(UDIVREM.getValue(1));
441 SDValue Op = SDValue(N, 0);
443 EVT VT = Op.getValueType();
444 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
446 SDValue one = DAG.getConstant(1, HalfVT);
447 SDValue zero = DAG.getConstant(0, HalfVT);
450 SDValue LHS = N->getOperand(0);
451 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
452 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
454 SDValue RHS = N->getOperand(1);
455 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
456 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
458 // Get Speculative values
459 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
460 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
462 SDValue REM_Hi = zero;
463 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
465 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
466 SDValue DIV_Lo = zero;
468 const unsigned halfBitWidth = HalfVT.getSizeInBits();
470 for (unsigned i = 0; i < halfBitWidth; ++i) {
471 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
472 // Get Value of high bit
474 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
475 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
477 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
478 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
481 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
482 DAG.getConstant(halfBitWidth - 1, HalfVT));
483 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
484 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
486 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
487 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
490 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
492 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
493 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
495 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
499 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
501 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
502 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
503 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
506 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
507 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
508 Results.push_back(DIV);
509 Results.push_back(REM);
517 // FIXME: This implements accesses to initialized globals in the constant
518 // address space by copying them to private and accessing that. It does not
519 // properly handle illegal types or vectors. The private vector loads are not
520 // scalarized, and the illegal scalars hit an assertion. This technique will not
521 // work well with large initializers, and this should eventually be
522 // removed. Initialized globals should be placed into a data section that the
523 // runtime will load into a buffer before the kernel is executed. Uses of the
524 // global need to be replaced with a pointer loaded from an implicit kernel
525 // argument into this buffer holding the copy of the data, which will remove the
526 // need for any of this.
527 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
528 const GlobalValue *GV,
529 const SDValue &InitPtr,
531 SelectionDAG &DAG) const {
532 const DataLayout *TD = getTargetMachine().getDataLayout();
534 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
535 EVT VT = EVT::getEVT(CI->getType());
536 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
537 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
538 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
539 TD->getPrefTypeAlignment(CI->getType()));
542 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
543 EVT VT = EVT::getEVT(CFP->getType());
544 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
545 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
546 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
547 TD->getPrefTypeAlignment(CFP->getType()));
550 Type *InitTy = Init->getType();
551 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
552 const StructLayout *SL = TD->getStructLayout(ST);
554 EVT PtrVT = InitPtr.getValueType();
555 SmallVector<SDValue, 8> Chains;
557 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
558 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
559 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
561 Constant *Elt = Init->getAggregateElement(I);
562 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
565 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
568 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
569 EVT PtrVT = InitPtr.getValueType();
571 unsigned NumElements;
572 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
573 NumElements = AT->getNumElements();
574 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
575 NumElements = VT->getNumElements();
577 llvm_unreachable("Unexpected type");
579 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
580 SmallVector<SDValue, 8> Chains;
581 for (unsigned i = 0; i < NumElements; ++i) {
582 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
583 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
585 Constant *Elt = Init->getAggregateElement(i);
586 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
589 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
593 llvm_unreachable("Unhandled constant initializer");
596 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
598 SelectionDAG &DAG) const {
600 const DataLayout *TD = getTargetMachine().getDataLayout();
601 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
602 const GlobalValue *GV = G->getGlobal();
604 switch (G->getAddressSpace()) {
605 default: llvm_unreachable("Global Address lowering not implemented for this "
607 case AMDGPUAS::LOCAL_ADDRESS: {
608 // XXX: What does the value of G->getOffset() mean?
609 assert(G->getOffset() == 0 &&
610 "Do not know what to do with an non-zero offset");
613 if (MFI->LocalMemoryObjects.count(GV) == 0) {
614 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
615 Offset = MFI->LDSSize;
616 MFI->LocalMemoryObjects[GV] = Offset;
617 // XXX: Account for alignment?
618 MFI->LDSSize += Size;
620 Offset = MFI->LocalMemoryObjects[GV];
623 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
625 case AMDGPUAS::CONSTANT_ADDRESS: {
626 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
627 Type *EltType = GV->getType()->getElementType();
628 unsigned Size = TD->getTypeAllocSize(EltType);
629 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
631 const GlobalVariable *Var = cast<GlobalVariable>(GV);
632 const Constant *Init = Var->getInitializer();
633 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
634 SDValue InitPtr = DAG.getFrameIndex(FI,
635 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
636 SmallVector<SDNode*, 8> WorkList;
638 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
639 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
640 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
642 WorkList.push_back(*I);
644 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
645 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
646 E = WorkList.end(); I != E; ++I) {
647 SmallVector<SDValue, 8> Ops;
648 Ops.push_back(Chain);
649 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
650 Ops.push_back((*I)->getOperand(i));
652 DAG.UpdateNodeOperands(*I, Ops);
654 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
655 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
660 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
661 SelectionDAG &DAG) const {
662 SmallVector<SDValue, 8> Args;
663 SDValue A = Op.getOperand(0);
664 SDValue B = Op.getOperand(1);
666 DAG.ExtractVectorElements(A, Args);
667 DAG.ExtractVectorElements(B, Args);
669 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
672 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
673 SelectionDAG &DAG) const {
675 SmallVector<SDValue, 8> Args;
676 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
677 EVT VT = Op.getValueType();
678 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
679 VT.getVectorNumElements());
681 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
684 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
685 SelectionDAG &DAG) const {
687 MachineFunction &MF = DAG.getMachineFunction();
688 const AMDGPUFrameLowering *TFL =
689 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
691 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
694 unsigned FrameIndex = FIN->getIndex();
695 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
696 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
700 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
701 SelectionDAG &DAG) const {
702 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
704 EVT VT = Op.getValueType();
706 switch (IntrinsicID) {
708 case AMDGPUIntrinsic::AMDIL_abs:
709 return LowerIntrinsicIABS(Op, DAG);
710 case AMDGPUIntrinsic::AMDIL_exp:
711 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
712 case AMDGPUIntrinsic::AMDGPU_lrp:
713 return LowerIntrinsicLRP(Op, DAG);
714 case AMDGPUIntrinsic::AMDIL_fraction:
715 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
716 case AMDGPUIntrinsic::AMDIL_max:
717 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
719 case AMDGPUIntrinsic::AMDGPU_imax:
720 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
722 case AMDGPUIntrinsic::AMDGPU_umax:
723 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
725 case AMDGPUIntrinsic::AMDIL_min:
726 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
728 case AMDGPUIntrinsic::AMDGPU_imin:
729 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
731 case AMDGPUIntrinsic::AMDGPU_umin:
732 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
735 case AMDGPUIntrinsic::AMDGPU_umul24:
736 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
737 Op.getOperand(1), Op.getOperand(2));
739 case AMDGPUIntrinsic::AMDGPU_imul24:
740 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
741 Op.getOperand(1), Op.getOperand(2));
743 case AMDGPUIntrinsic::AMDGPU_umad24:
744 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
745 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
747 case AMDGPUIntrinsic::AMDGPU_imad24:
748 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
749 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
751 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
752 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
757 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
758 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
763 case AMDGPUIntrinsic::AMDGPU_bfi:
764 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
769 case AMDGPUIntrinsic::AMDGPU_bfm:
770 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
774 case AMDGPUIntrinsic::AMDIL_round_nearest:
775 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
779 ///IABS(a) = SMAX(sub(0, a), a)
780 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
781 SelectionDAG &DAG) const {
783 EVT VT = Op.getValueType();
784 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
787 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
790 /// Linear Interpolation
791 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
792 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
793 SelectionDAG &DAG) const {
795 EVT VT = Op.getValueType();
796 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
797 DAG.getConstantFP(1.0f, MVT::f32),
799 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
801 return DAG.getNode(ISD::FADD, DL, VT,
802 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
806 /// \brief Generate Min/Max node
807 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
808 SelectionDAG &DAG) const {
810 EVT VT = N->getValueType(0);
812 SDValue LHS = N->getOperand(0);
813 SDValue RHS = N->getOperand(1);
814 SDValue True = N->getOperand(2);
815 SDValue False = N->getOperand(3);
816 SDValue CC = N->getOperand(4);
818 if (VT != MVT::f32 ||
819 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
823 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
837 llvm_unreachable("Operation should already be optimised!");
844 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
845 return DAG.getNode(Opc, DL, VT, LHS, RHS);
853 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
854 return DAG.getNode(Opc, DL, VT, LHS, RHS);
856 case ISD::SETCC_INVALID:
857 llvm_unreachable("Invalid setcc condcode!");
862 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
863 SelectionDAG &DAG) const {
864 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
865 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
866 EVT EltVT = Op.getValueType().getVectorElementType();
867 EVT PtrVT = Load->getBasePtr().getValueType();
868 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
869 SmallVector<SDValue, 8> Loads;
872 for (unsigned i = 0, e = NumElts; i != e; ++i) {
873 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
874 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
875 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
876 Load->getChain(), Ptr,
877 MachinePointerInfo(Load->getMemOperand()->getValue()),
878 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
879 Load->getAlignment()));
881 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
884 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
885 SelectionDAG &DAG) const {
886 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
887 EVT MemVT = Store->getMemoryVT();
888 unsigned MemBits = MemVT.getSizeInBits();
890 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
891 // truncating store into an i32 store.
892 // XXX: We could also handle optimize other vector bitwidths.
893 if (!MemVT.isVector() || MemBits > 32) {
898 SDValue Value = Store->getValue();
899 EVT VT = Value.getValueType();
900 EVT ElemVT = VT.getVectorElementType();
901 SDValue Ptr = Store->getBasePtr();
902 EVT MemEltVT = MemVT.getVectorElementType();
903 unsigned MemEltBits = MemEltVT.getSizeInBits();
904 unsigned MemNumElements = MemVT.getVectorNumElements();
905 unsigned PackedSize = MemVT.getStoreSizeInBits();
906 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
908 assert(Value.getValueType().getScalarSizeInBits() >= 32);
911 for (unsigned i = 0; i < MemNumElements; ++i) {
912 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
913 DAG.getConstant(i, MVT::i32));
914 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
915 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
917 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
918 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
923 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
927 if (PackedSize < 32) {
928 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
929 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
930 Store->getMemOperand()->getPointerInfo(),
932 Store->isNonTemporal(), Store->isVolatile(),
933 Store->getAlignment());
936 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
937 Store->getMemOperand()->getPointerInfo(),
938 Store->isVolatile(), Store->isNonTemporal(),
939 Store->getAlignment());
942 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
943 SelectionDAG &DAG) const {
944 StoreSDNode *Store = cast<StoreSDNode>(Op);
945 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
946 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
947 EVT PtrVT = Store->getBasePtr().getValueType();
948 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
951 SmallVector<SDValue, 8> Chains;
953 for (unsigned i = 0, e = NumElts; i != e; ++i) {
954 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
955 Store->getValue(), DAG.getConstant(i, MVT::i32));
956 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
958 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
960 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
961 MachinePointerInfo(Store->getMemOperand()->getValue()),
962 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
963 Store->getAlignment()));
965 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
968 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
970 LoadSDNode *Load = cast<LoadSDNode>(Op);
971 ISD::LoadExtType ExtType = Load->getExtensionType();
972 EVT VT = Op.getValueType();
973 EVT MemVT = Load->getMemoryVT();
975 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
976 // We can do the extload to 32-bits, and then need to separately extend to
979 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
983 Load->getMemOperand());
984 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
987 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
988 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
989 // FIXME: Copied from PPC
990 // First, load into 32 bits, then truncate to 1 bit.
992 SDValue Chain = Load->getChain();
993 SDValue BasePtr = Load->getBasePtr();
994 MachineMemOperand *MMO = Load->getMemOperand();
996 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
997 BasePtr, MVT::i8, MMO);
998 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1001 // Lower loads constant address space global variable loads
1002 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1003 isa<GlobalVariable>(
1004 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1006 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1007 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1008 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1009 DAG.getConstant(2, MVT::i32));
1010 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1011 Load->getChain(), Ptr,
1012 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1015 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1016 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1020 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1021 DAG.getConstant(2, MVT::i32));
1022 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1023 Load->getChain(), Ptr,
1024 DAG.getTargetConstant(0, MVT::i32),
1026 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1028 DAG.getConstant(0x3, MVT::i32));
1029 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1030 DAG.getConstant(3, MVT::i32));
1032 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1034 EVT MemEltVT = MemVT.getScalarType();
1035 if (ExtType == ISD::SEXTLOAD) {
1036 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1037 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1040 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1043 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1045 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1046 if (Result.getNode()) {
1050 StoreSDNode *Store = cast<StoreSDNode>(Op);
1051 SDValue Chain = Store->getChain();
1052 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1053 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1054 Store->getValue().getValueType().isVector()) {
1055 return SplitVectorStore(Op, DAG);
1058 EVT MemVT = Store->getMemoryVT();
1059 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1060 MemVT.bitsLT(MVT::i32)) {
1062 if (Store->getMemoryVT() == MVT::i8) {
1064 } else if (Store->getMemoryVT() == MVT::i16) {
1067 SDValue BasePtr = Store->getBasePtr();
1068 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1069 DAG.getConstant(2, MVT::i32));
1070 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1071 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1073 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1074 DAG.getConstant(0x3, MVT::i32));
1076 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1077 DAG.getConstant(3, MVT::i32));
1079 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1082 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1084 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1085 MaskedValue, ShiftAmt);
1087 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1089 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1090 DAG.getConstant(0xffffffff, MVT::i32));
1091 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1093 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1094 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1095 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1100 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1101 SelectionDAG &DAG) const {
1103 EVT VT = Op.getValueType();
1105 SDValue Num = Op.getOperand(0);
1106 SDValue Den = Op.getOperand(1);
1108 // RCP = URECIP(Den) = 2^32 / Den + e
1109 // e is rounding error.
1110 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1112 // RCP_LO = umulo(RCP, Den) */
1113 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1115 // RCP_HI = mulhu (RCP, Den) */
1116 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1118 // NEG_RCP_LO = -RCP_LO
1119 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1122 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1123 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1126 // Calculate the rounding error from the URECIP instruction
1127 // E = mulhu(ABS_RCP_LO, RCP)
1128 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1130 // RCP_A_E = RCP + E
1131 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1133 // RCP_S_E = RCP - E
1134 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1136 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1137 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1140 // Quotient = mulhu(Tmp0, Num)
1141 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1143 // Num_S_Remainder = Quotient * Den
1144 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1146 // Remainder = Num - Num_S_Remainder
1147 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1149 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1150 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1151 DAG.getConstant(-1, VT),
1152 DAG.getConstant(0, VT),
1154 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1155 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1157 DAG.getConstant(-1, VT),
1158 DAG.getConstant(0, VT),
1160 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1161 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1164 // Calculate Division result:
1166 // Quotient_A_One = Quotient + 1
1167 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1168 DAG.getConstant(1, VT));
1170 // Quotient_S_One = Quotient - 1
1171 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1172 DAG.getConstant(1, VT));
1174 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1175 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1176 Quotient, Quotient_A_One, ISD::SETEQ);
1178 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1179 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1180 Quotient_S_One, Div, ISD::SETEQ);
1182 // Calculate Rem result:
1184 // Remainder_S_Den = Remainder - Den
1185 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1187 // Remainder_A_Den = Remainder + Den
1188 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1190 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1191 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1192 Remainder, Remainder_S_Den, ISD::SETEQ);
1194 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1195 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1196 Remainder_A_Den, Rem, ISD::SETEQ);
1201 return DAG.getMergeValues(Ops, DL);
1204 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1205 SelectionDAG &DAG) const {
1206 SDValue S0 = Op.getOperand(0);
1208 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1211 // f32 uint_to_fp i64
1212 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1213 DAG.getConstant(0, MVT::i32));
1214 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1215 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1216 DAG.getConstant(1, MVT::i32));
1217 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1218 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1219 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1220 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1224 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1226 SelectionDAG &DAG) const {
1227 MVT VT = Op.getSimpleValueType();
1229 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1230 // Shift left by 'Shift' bits.
1231 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1232 // Signed shift Right by 'Shift' bits.
1233 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1236 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1237 SelectionDAG &DAG) const {
1238 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1239 MVT VT = Op.getSimpleValueType();
1240 MVT ScalarVT = VT.getScalarType();
1245 SDValue Src = Op.getOperand(0);
1248 // TODO: Don't scalarize on Evergreen?
1249 unsigned NElts = VT.getVectorNumElements();
1250 SmallVector<SDValue, 8> Args;
1251 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1253 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1254 for (unsigned I = 0; I < NElts; ++I)
1255 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1257 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1260 //===----------------------------------------------------------------------===//
1261 // Custom DAG optimizations
1262 //===----------------------------------------------------------------------===//
1264 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1265 APInt KnownZero, KnownOne;
1266 EVT VT = Op.getValueType();
1267 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1269 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1272 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1273 EVT VT = Op.getValueType();
1275 // In order for this to be a signed 24-bit value, bit 23, must
1277 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1278 // as unsigned 24-bit values.
1279 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1282 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1284 SelectionDAG &DAG = DCI.DAG;
1285 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1286 EVT VT = Op.getValueType();
1288 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1289 APInt KnownZero, KnownOne;
1290 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1291 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1292 DCI.CommitTargetLoweringOpt(TLO);
1295 template <typename IntTy>
1296 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1297 uint32_t Offset, uint32_t Width) {
1298 if (Width + Offset < 32) {
1299 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1300 return DAG.getConstant(Result, MVT::i32);
1303 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1306 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1307 DAGCombinerInfo &DCI) const {
1308 SelectionDAG &DAG = DCI.DAG;
1311 switch(N->getOpcode()) {
1314 EVT VT = N->getValueType(0);
1315 SDValue N0 = N->getOperand(0);
1316 SDValue N1 = N->getOperand(1);
1319 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1320 if (VT.isVector() || VT.getSizeInBits() > 32)
1323 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1324 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1325 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1326 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1327 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1328 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1329 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1330 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1335 // We need to use sext even for MUL_U24, because MUL_U24 is used
1336 // for signed multiply of 8 and 16-bit types.
1337 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1341 case AMDGPUISD::MUL_I24:
1342 case AMDGPUISD::MUL_U24: {
1343 SDValue N0 = N->getOperand(0);
1344 SDValue N1 = N->getOperand(1);
1345 simplifyI24(N0, DCI);
1346 simplifyI24(N1, DCI);
1349 case ISD::SELECT_CC: {
1350 return CombineMinMax(N, DAG);
1352 case AMDGPUISD::BFE_I32:
1353 case AMDGPUISD::BFE_U32: {
1354 assert(!N->getValueType(0).isVector() &&
1355 "Vector handling of BFE not implemented");
1356 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1360 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1362 return DAG.getConstant(0, MVT::i32);
1364 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1368 SDValue BitsFrom = N->getOperand(0);
1369 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1371 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1373 if (OffsetVal == 0) {
1374 // This is already sign / zero extended, so try to fold away extra BFEs.
1375 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1377 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1378 if (OpSignBits >= SignBits)
1381 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1383 // This is a sign_extend_inreg. Replace it to take advantage of existing
1384 // DAG Combines. If not eliminated, we will match back to BFE during
1387 // TODO: The sext_inreg of extended types ends, although we can could
1388 // handle them in a single BFE.
1389 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1390 DAG.getValueType(SmallVT));
1393 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1396 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1398 return constantFoldBFE<int32_t>(DAG,
1399 Val->getSExtValue(),
1404 return constantFoldBFE<uint32_t>(DAG,
1405 Val->getZExtValue(),
1410 APInt Demanded = APInt::getBitsSet(32,
1412 OffsetVal + WidthVal);
1414 if ((OffsetVal + WidthVal) >= 32) {
1415 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1416 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1417 BitsFrom, ShiftVal);
1420 APInt KnownZero, KnownOne;
1421 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1422 !DCI.isBeforeLegalizeOps());
1423 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1424 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1425 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1426 DCI.CommitTargetLoweringOpt(TLO);
1435 //===----------------------------------------------------------------------===//
1437 //===----------------------------------------------------------------------===//
1439 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1442 const SmallVectorImpl<ISD::InputArg> &Ins,
1443 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1445 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1446 if (Ins[i].ArgVT == Ins[i].VT) {
1447 OrigIns.push_back(Ins[i]);
1452 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1453 // Vector has been split into scalars.
1454 VT = Ins[i].ArgVT.getVectorElementType();
1455 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1456 Ins[i].ArgVT.getVectorElementType() !=
1457 Ins[i].VT.getVectorElementType()) {
1458 // Vector elements have been promoted
1461 // Vector has been spilt into smaller vectors.
1465 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1466 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1467 OrigIns.push_back(Arg);
1471 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1472 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1473 return CFP->isExactlyValue(1.0);
1475 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1476 return C->isAllOnesValue();
1481 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1482 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1483 return CFP->getValueAPF().isZero();
1485 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1486 return C->isNullValue();
1491 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1492 const TargetRegisterClass *RC,
1493 unsigned Reg, EVT VT) const {
1494 MachineFunction &MF = DAG.getMachineFunction();
1495 MachineRegisterInfo &MRI = MF.getRegInfo();
1496 unsigned VirtualRegister;
1497 if (!MRI.isLiveIn(Reg)) {
1498 VirtualRegister = MRI.createVirtualRegister(RC);
1499 MRI.addLiveIn(Reg, VirtualRegister);
1501 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1503 return DAG.getRegister(VirtualRegister, VT);
1506 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1508 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1510 default: return nullptr;
1512 NODE_NAME_CASE(CALL);
1513 NODE_NAME_CASE(UMUL);
1514 NODE_NAME_CASE(DIV_INF);
1515 NODE_NAME_CASE(RET_FLAG);
1516 NODE_NAME_CASE(BRANCH_COND);
1519 NODE_NAME_CASE(DWORDADDR)
1520 NODE_NAME_CASE(FRACT)
1521 NODE_NAME_CASE(FMAX)
1522 NODE_NAME_CASE(SMAX)
1523 NODE_NAME_CASE(UMAX)
1524 NODE_NAME_CASE(FMIN)
1525 NODE_NAME_CASE(SMIN)
1526 NODE_NAME_CASE(UMIN)
1527 NODE_NAME_CASE(BFE_U32)
1528 NODE_NAME_CASE(BFE_I32)
1531 NODE_NAME_CASE(MUL_U24)
1532 NODE_NAME_CASE(MUL_I24)
1533 NODE_NAME_CASE(MAD_U24)
1534 NODE_NAME_CASE(MAD_I24)
1535 NODE_NAME_CASE(URECIP)
1536 NODE_NAME_CASE(DOT4)
1537 NODE_NAME_CASE(EXPORT)
1538 NODE_NAME_CASE(CONST_ADDRESS)
1539 NODE_NAME_CASE(REGISTER_LOAD)
1540 NODE_NAME_CASE(REGISTER_STORE)
1541 NODE_NAME_CASE(LOAD_CONSTANT)
1542 NODE_NAME_CASE(LOAD_INPUT)
1543 NODE_NAME_CASE(SAMPLE)
1544 NODE_NAME_CASE(SAMPLEB)
1545 NODE_NAME_CASE(SAMPLED)
1546 NODE_NAME_CASE(SAMPLEL)
1547 NODE_NAME_CASE(STORE_MSKOR)
1548 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1552 static void computeKnownBitsForMinMax(const SDValue Op0,
1556 const SelectionDAG &DAG,
1558 APInt Op0Zero, Op0One;
1559 APInt Op1Zero, Op1One;
1560 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
1561 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
1563 KnownZero = Op0Zero & Op1Zero;
1564 KnownOne = Op0One & Op1One;
1567 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
1571 const SelectionDAG &DAG,
1572 unsigned Depth) const {
1574 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1578 unsigned Opc = Op.getOpcode();
1583 case ISD::INTRINSIC_WO_CHAIN: {
1584 // FIXME: The intrinsic should just use the node.
1585 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1586 case AMDGPUIntrinsic::AMDGPU_imax:
1587 case AMDGPUIntrinsic::AMDGPU_umax:
1588 case AMDGPUIntrinsic::AMDGPU_imin:
1589 case AMDGPUIntrinsic::AMDGPU_umin:
1590 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1591 KnownZero, KnownOne, DAG, Depth);
1599 case AMDGPUISD::SMAX:
1600 case AMDGPUISD::UMAX:
1601 case AMDGPUISD::SMIN:
1602 case AMDGPUISD::UMIN:
1603 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1604 KnownZero, KnownOne, DAG, Depth);
1607 case AMDGPUISD::BFE_I32:
1608 case AMDGPUISD::BFE_U32: {
1609 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1613 unsigned BitWidth = 32;
1614 uint32_t Width = CWidth->getZExtValue() & 0x1f;
1616 KnownZero = APInt::getAllOnesValue(BitWidth);
1617 KnownOne = APInt::getNullValue(BitWidth);
1621 // FIXME: This could do a lot more. If offset is 0, should be the same as
1622 // sign_extend_inreg implementation, but that involves duplicating it.
1623 if (Opc == AMDGPUISD::BFE_I32)
1624 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1626 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1633 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
1635 const SelectionDAG &DAG,
1636 unsigned Depth) const {
1637 switch (Op.getOpcode()) {
1638 case AMDGPUISD::BFE_I32: {
1639 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1643 unsigned SignBits = 32 - Width->getZExtValue() + 1;
1644 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1645 if (!Offset || !Offset->isNullValue())
1648 // TODO: Could probably figure something out with non-0 offsets.
1649 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
1650 return std::max(SignBits, Op0SignBits);
1653 case AMDGPUISD::BFE_U32: {
1654 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1655 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;