// Without SSE, i64->f64 goes through memory.
setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
}
- }
+ } else if (!Subtarget->is64Bit())
+ setOperationAction(ISD::BITCAST , MVT::i64 , Custom);
// Scalar integer divide and remainder are lowered to use operations that
// produce two results, to match the available instructions. This exposes
DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
}
+ const X86RegisterInfo *TRI = Subtarget->getRegisterInfo();
+ const MCPhysReg *I =
+ TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction());
+ if (I) {
+ for (; *I; ++I) {
+ if (X86::GR64RegClass.contains(*I))
+ RetOps.push_back(DAG.getRegister(*I, MVT::i64));
+ else
+ llvm_unreachable("Unexpected register class in CSRsViaCopy!");
+ }
+ }
+
RetOps[0] = Chain; // Update chain.
// Add the flag if we have it.
}
}
+
+bool X86TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
+ const CallInst &I,
+ unsigned Intrinsic) const {
+
+ const IntrinsicData* IntrData = getIntrinsicWithChain(Intrinsic);
+ if (!IntrData)
+ return false;
+
+ switch (IntrData->Type) {
+ case LOADA:
+ case LOADU: {
+ Info.opc = ISD::INTRINSIC_W_CHAIN;
+ Info.memVT = MVT::getVT(I.getType());
+ Info.ptrVal = I.getArgOperand(0);
+ Info.offset = 0;
+ Info.align = (IntrData->Type == LOADA ? Info.memVT.getSizeInBits()/8 : 1);
+ Info.vol = false;
+ Info.readMem = true;
+ Info.writeMem = false;
+ return true;
+ }
+ default:
+ break;
+ }
+
+ return false;
+}
+
/// Returns true if the target can instruction select the
/// specified FP immediate natively. If false, the legalizer will
/// materialize the FP immediate as a load from a constant pool.
DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
return TruncBroadcast;
+ MVT BroadcastVT = VT;
+
+ // Peek through any bitcast (only useful for loads).
+ SDValue BC = V;
+ while (BC.getOpcode() == ISD::BITCAST)
+ BC = BC.getOperand(0);
+
// Also check the simpler case, where we can directly reuse the scalar.
if (V.getOpcode() == ISD::BUILD_VECTOR ||
(V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
// Only AVX2 has register broadcasts.
if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
return SDValue();
- } else if (MayFoldLoad(V) && !cast<LoadSDNode>(V)->isVolatile()) {
+ } else if (MayFoldLoad(BC) && !cast<LoadSDNode>(BC)->isVolatile()) {
+ // 32-bit targets need to load i64 as a f64 and then bitcast the result.
+ if (!Subtarget->is64Bit() && VT.getScalarType() == MVT::i64)
+ BroadcastVT = MVT::getVectorVT(MVT::f64, VT.getVectorNumElements());
+
// If we are broadcasting a load that is only used by the shuffle
// then we can reduce the vector load to the broadcasted scalar load.
- LoadSDNode *Ld = cast<LoadSDNode>(V);
+ LoadSDNode *Ld = cast<LoadSDNode>(BC);
SDValue BaseAddr = Ld->getOperand(1);
EVT AddrVT = BaseAddr.getValueType();
- EVT SVT = VT.getScalarType();
+ EVT SVT = BroadcastVT.getScalarType();
unsigned Offset = BroadcastIdx * SVT.getStoreSize();
SDValue NewAddr = DAG.getNode(
ISD::ADD, DL, AddrVT, BaseAddr,
return SDValue();
}
- return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
+ V = DAG.getNode(X86ISD::VBROADCAST, DL, BroadcastVT, V);
+ return DAG.getBitcast(VT, V);
}
// Check for whether we can use INSERTPS to perform the shuffle. We only use
return Op;
}
+ SDValue ValueToStore = Op.getOperand(0);
+ if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
+ !Subtarget->is64Bit())
+ // Bitcasting to f64 here allows us to do a single 64-bit store from
+ // an SSE register, avoiding the store forwarding penalty that would come
+ // with two 32-bit stores.
+ ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore);
+
unsigned Size = SrcVT.getSizeInBits()/8;
MachineFunction &MF = DAG.getMachineFunction();
auto PtrVT = getPointerTy(MF.getDataLayout());
int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
SDValue Chain = DAG.getStore(
- DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
+ DAG.getEntryNode(), dl, ValueToStore, StackSlot,
MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
false, 0);
return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
}
assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
- SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
+ SDValue ValueToStore = Op.getOperand(0);
+ if (isScalarFPTypeInSSEReg(Op.getValueType()) && !Subtarget->is64Bit())
+ // Bitcasting to f64 here allows us to do a single 64-bit store from
+ // an SSE register, avoiding the store forwarding penalty that would come
+ // with two 32-bit stores.
+ ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore);
+ SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, ValueToStore,
StackSlot, MachinePointerInfo(),
false, false, 0);
// For i64 source, we need to add the appropriate power of 2 if the input
return DAG.getMergeValues(Results, dl);
}
case COMPRESS_TO_MEM: {
- SDLoc dl(Op);
SDValue Mask = Op.getOperand(4);
SDValue DataToCompress = Op.getOperand(3);
SDValue Addr = Op.getOperand(2);
case TRUNCATE_TO_MEM_VI32:
return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
case EXPAND_FROM_MEM: {
- SDLoc dl(Op);
SDValue Mask = Op.getOperand(4);
SDValue PassThru = Op.getOperand(3);
SDValue Addr = Op.getOperand(2);
Mask, PassThru, Subtarget, DAG), Chain};
return DAG.getMergeValues(Results, dl);
}
+ case LOADU:
+ case LOADA: {
+ SDValue Mask = Op.getOperand(4);
+ SDValue PassThru = Op.getOperand(3);
+ SDValue Addr = Op.getOperand(2);
+ SDValue Chain = Op.getOperand(0);
+ MVT VT = Op.getSimpleValueType();
+
+ MemIntrinsicSDNode *MemIntr = dyn_cast<MemIntrinsicSDNode>(Op);
+ assert(MemIntr && "Expected MemIntrinsicSDNode!");
+
+ if (isAllOnesConstant(Mask)) // return just a load
+ return DAG.getLoad(VT, dl, Chain, Addr, MemIntr->getMemOperand());
+
+ MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
+ SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
+ return DAG.getMaskedLoad(VT, dl, Chain, Addr, VMask, PassThru, VT,
+ MemIntr->getMemOperand(), ISD::NON_EXTLOAD);
+ }
}
}
MVT SrcVT = Op.getOperand(0).getSimpleValueType();
MVT DstVT = Op.getSimpleValueType();
- if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
+ if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8 ||
+ SrcVT == MVT::i64) {
assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
if (DstVT != MVT::f64)
// This conversion needs to be expanded.
return SDValue();
- SDValue InVec = Op->getOperand(0);
- SDLoc dl(Op);
- unsigned NumElts = SrcVT.getVectorNumElements();
- MVT SVT = SrcVT.getVectorElementType();
-
- // Widen the vector in input in the case of MVT::v2i32.
- // Example: from MVT::v2i32 to MVT::v4i32.
+ SDValue Op0 = Op->getOperand(0);
SmallVector<SDValue, 16> Elts;
- for (unsigned i = 0, e = NumElts; i != e; ++i)
- Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
- DAG.getIntPtrConstant(i, dl)));
-
+ SDLoc dl(Op);
+ unsigned NumElts;
+ MVT SVT;
+ if (SrcVT.isVector()) {
+ NumElts = SrcVT.getVectorNumElements();
+ SVT = SrcVT.getVectorElementType();
+
+ // Widen the vector in input in the case of MVT::v2i32.
+ // Example: from MVT::v2i32 to MVT::v4i32.
+ for (unsigned i = 0, e = NumElts; i != e; ++i)
+ Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, Op0,
+ DAG.getIntPtrConstant(i, dl)));
+ } else {
+ assert(SrcVT == MVT::i64 && !Subtarget->is64Bit() &&
+ "Unexpected source type in LowerBITCAST");
+ Elts.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op0,
+ DAG.getIntPtrConstant(0, dl)));
+ Elts.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op0,
+ DAG.getIntPtrConstant(1, dl)));
+ NumElts = 2;
+ SVT = MVT::i32;
+ }
// Explicitly mark the extra elements as Undef.
Elts.append(NumElts, DAG.getUNDEF(SVT));
case X86ISD::VSHLI: return "X86ISD::VSHLI";
case X86ISD::VSRLI: return "X86ISD::VSRLI";
case X86ISD::VSRAI: return "X86ISD::VSRAI";
+ case X86ISD::VROTLI: return "X86ISD::VROTLI";
+ case X86ISD::VROTRI: return "X86ISD::VROTRI";
case X86ISD::CMPP: return "X86ISD::CMPP";
case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
Attribute::MinSize);
return OptSize && !VT.isVector();
}
+
+void X86TargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const {
+ if (!Subtarget->is64Bit())
+ return;
+
+ // Update IsSplitCSR in X86MachineFunctionInfo.
+ X86MachineFunctionInfo *AFI =
+ Entry->getParent()->getInfo<X86MachineFunctionInfo>();
+ AFI->setIsSplitCSR(true);
+}
+
+void X86TargetLowering::insertCopiesSplitCSR(
+ MachineBasicBlock *Entry,
+ const SmallVectorImpl<MachineBasicBlock *> &Exits) const {
+ const X86RegisterInfo *TRI = Subtarget->getRegisterInfo();
+ const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent());
+ if (!IStart)
+ return;
+
+ const TargetInstrInfo *TII = Subtarget->getInstrInfo();
+ MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo();
+ for (const MCPhysReg *I = IStart; *I; ++I) {
+ const TargetRegisterClass *RC = nullptr;
+ if (X86::GR64RegClass.contains(*I))
+ RC = &X86::GR64RegClass;
+ else
+ llvm_unreachable("Unexpected register class in CSRsViaCopy!");
+
+ unsigned NewVR = MRI->createVirtualRegister(RC);
+ // Create copy from CSR to a virtual register.
+ // FIXME: this currently does not emit CFI pseudo-instructions, it works
+ // fine for CXX_FAST_TLS since the C++-style TLS access functions should be
+ // nounwind. If we want to generalize this later, we may need to emit
+ // CFI pseudo-instructions.
+ assert(Entry->getParent()->getFunction()->hasFnAttribute(
+ Attribute::NoUnwind) &&
+ "Function should be nounwind in insertCopiesSplitCSR!");
+ Entry->addLiveIn(*I);
+ BuildMI(*Entry, Entry->begin(), DebugLoc(), TII->get(TargetOpcode::COPY),
+ NewVR)
+ .addReg(*I);
+
+ for (auto *Exit : Exits)
+ BuildMI(*Exit, Exit->begin(), DebugLoc(), TII->get(TargetOpcode::COPY),
+ *I)
+ .addReg(NewVR);
+ }
+}
projects / oota-llvm.git / blobdiff