setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand);
setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
setOperationAction(ISD::BUILD_VECTOR, VT, Expand);
+ setOperationAction(ISD::MULHU, VT, Expand);
+ setOperationAction(ISD::MULHS, VT, Expand);
setOperationAction(ISD::UMUL_LOHI, VT, Expand);
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
setOperationAction(ISD::UDIVREM, VT, Expand);
// Altivec does not contain unordered floating-point compare instructions
setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand);
setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand);
- setCondCodeAction(ISD::SETUGT, MVT::v4f32, Expand);
- setCondCodeAction(ISD::SETUGE, MVT::v4f32, Expand);
- setCondCodeAction(ISD::SETULT, MVT::v4f32, Expand);
- setCondCodeAction(ISD::SETULE, MVT::v4f32, Expand);
-
setCondCodeAction(ISD::SETO, MVT::v4f32, Expand);
setCondCodeAction(ISD::SETONE, MVT::v4f32, Expand);
// Share the Altivec comparison restrictions.
setCondCodeAction(ISD::SETUO, MVT::v2f64, Expand);
setCondCodeAction(ISD::SETUEQ, MVT::v2f64, Expand);
- setCondCodeAction(ISD::SETUGT, MVT::v2f64, Expand);
- setCondCodeAction(ISD::SETUGE, MVT::v2f64, Expand);
- setCondCodeAction(ISD::SETULT, MVT::v2f64, Expand);
- setCondCodeAction(ISD::SETULE, MVT::v2f64, Expand);
-
setCondCodeAction(ISD::SETO, MVT::v2f64, Expand);
setCondCodeAction(ISD::SETONE, MVT::v2f64, Expand);
// Altivec instructions set fields to all zeros or all ones.
setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
+ if (!isPPC64) {
+ // These libcalls are not available in 32-bit.
+ setLibcallName(RTLIB::SHL_I128, nullptr);
+ setLibcallName(RTLIB::SRL_I128, nullptr);
+ setLibcallName(RTLIB::SRA_I128, nullptr);
+ }
+
if (isPPC64) {
setStackPointerRegisterToSaveRestore(PPC::X1);
setExceptionPointerRegister(PPC::X3);
/// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a
/// VPKUHUM instruction.
-bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary,
+/// The ShuffleKind distinguishes between big-endian operations with
+/// two different inputs (0), either-endian operations with two identical
+/// inputs (1), and little-endian operantion with two different inputs (2).
+/// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
+bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
SelectionDAG &DAG) {
- unsigned j = DAG.getTarget().getDataLayout()->isLittleEndian() ? 0 : 1;
- if (!isUnary) {
+ bool IsLE = DAG.getSubtarget().getDataLayout()->isLittleEndian();
+ if (ShuffleKind == 0) {
+ if (IsLE)
+ return false;
for (unsigned i = 0; i != 16; ++i)
- if (!isConstantOrUndef(N->getMaskElt(i), i*2+j))
+ if (!isConstantOrUndef(N->getMaskElt(i), i*2+1))
return false;
- } else {
+ } else if (ShuffleKind == 2) {
+ if (!IsLE)
+ return false;
+ for (unsigned i = 0; i != 16; ++i)
+ if (!isConstantOrUndef(N->getMaskElt(i), i*2))
+ return false;
+ } else if (ShuffleKind == 1) {
+ unsigned j = IsLE ? 0 : 1;
for (unsigned i = 0; i != 8; ++i)
if (!isConstantOrUndef(N->getMaskElt(i), i*2+j) ||
!isConstantOrUndef(N->getMaskElt(i+8), i*2+j))
/// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a
/// VPKUWUM instruction.
-bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, bool isUnary,
+/// The ShuffleKind distinguishes between big-endian operations with
+/// two different inputs (0), either-endian operations with two identical
+/// inputs (1), and little-endian operantion with two different inputs (2).
+/// For the latter, the input operands are swapped (see PPCInstrAltivec.td).
+bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind,
SelectionDAG &DAG) {
- unsigned j, k;
- if (DAG.getTarget().getDataLayout()->isLittleEndian()) {
- j = 0;
- k = 1;
- } else {
- j = 2;
- k = 3;
- }
- if (!isUnary) {
+ bool IsLE = DAG.getSubtarget().getDataLayout()->isLittleEndian();
+ if (ShuffleKind == 0) {
+ if (IsLE)
+ return false;
for (unsigned i = 0; i != 16; i += 2)
- if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) ||
- !isConstantOrUndef(N->getMaskElt(i+1), i*2+k))
+ if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) ||
+ !isConstantOrUndef(N->getMaskElt(i+1), i*2+3))
return false;
- } else {
+ } else if (ShuffleKind == 2) {
+ if (!IsLE)
+ return false;
+ for (unsigned i = 0; i != 16; i += 2)
+ if (!isConstantOrUndef(N->getMaskElt(i ), i*2) ||
+ !isConstantOrUndef(N->getMaskElt(i+1), i*2+1))
+ return false;
+ } else if (ShuffleKind == 1) {
+ unsigned j = IsLE ? 0 : 2;
for (unsigned i = 0; i != 8; i += 2)
- if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) ||
- !isConstantOrUndef(N->getMaskElt(i+1), i*2+k) ||
- !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) ||
- !isConstantOrUndef(N->getMaskElt(i+9), i*2+k))
+ if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) ||
+ !isConstantOrUndef(N->getMaskElt(i+1), i*2+j+1) ||
+ !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) ||
+ !isConstantOrUndef(N->getMaskElt(i+9), i*2+j+1))
return false;
}
return true;
/// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for
/// a VMRGL* instruction with the specified unit size (1,2 or 4 bytes).
+/// The ShuffleKind distinguishes between big-endian merges with two
+/// different inputs (0), either-endian merges with two identical inputs (1),
+/// and little-endian merges with two different inputs (2). For the latter,
+/// the input operands are swapped (see PPCInstrAltivec.td).
bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
- bool isUnary, SelectionDAG &DAG) {
- if (DAG.getTarget().getDataLayout()->isLittleEndian()) {
- if (!isUnary)
+ unsigned ShuffleKind, SelectionDAG &DAG) {
+ if (DAG.getSubtarget().getDataLayout()->isLittleEndian()) {
+ if (ShuffleKind == 1) // unary
+ return isVMerge(N, UnitSize, 0, 0);
+ else if (ShuffleKind == 2) // swapped
return isVMerge(N, UnitSize, 0, 16);
- return isVMerge(N, UnitSize, 0, 0);
+ else
+ return false;
} else {
- if (!isUnary)
+ if (ShuffleKind == 1) // unary
+ return isVMerge(N, UnitSize, 8, 8);
+ else if (ShuffleKind == 0) // normal
return isVMerge(N, UnitSize, 8, 24);
- return isVMerge(N, UnitSize, 8, 8);
+ else
+ return false;
}
}
/// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for
/// a VMRGH* instruction with the specified unit size (1,2 or 4 bytes).
+/// The ShuffleKind distinguishes between big-endian merges with two
+/// different inputs (0), either-endian merges with two identical inputs (1),
+/// and little-endian merges with two different inputs (2). For the latter,
+/// the input operands are swapped (see PPCInstrAltivec.td).
bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize,
- bool isUnary, SelectionDAG &DAG) {
- if (DAG.getTarget().getDataLayout()->isLittleEndian()) {
- if (!isUnary)
+ unsigned ShuffleKind, SelectionDAG &DAG) {
+ if (DAG.getSubtarget().getDataLayout()->isLittleEndian()) {
+ if (ShuffleKind == 1) // unary
+ return isVMerge(N, UnitSize, 8, 8);
+ else if (ShuffleKind == 2) // swapped
return isVMerge(N, UnitSize, 8, 24);
- return isVMerge(N, UnitSize, 8, 8);
+ else
+ return false;
} else {
- if (!isUnary)
+ if (ShuffleKind == 1) // unary
+ return isVMerge(N, UnitSize, 0, 0);
+ else if (ShuffleKind == 0) // normal
return isVMerge(N, UnitSize, 0, 16);
- return isVMerge(N, UnitSize, 0, 0);
+ else
+ return false;
}
}
/// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift
/// amount, otherwise return -1.
-int PPC::isVSLDOIShuffleMask(SDNode *N, bool isUnary, SelectionDAG &DAG) {
+/// The ShuffleKind distinguishes between big-endian operations with two
+/// different inputs (0), either-endian operations with two identical inputs
+/// (1), and little-endian operations with two different inputs (2). For the
+/// latter, the input operands are swapped (see PPCInstrAltivec.td).
+int PPC::isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind,
+ SelectionDAG &DAG) {
if (N->getValueType(0) != MVT::v16i8)
return -1;
unsigned ShiftAmt = SVOp->getMaskElt(i);
if (ShiftAmt < i) return -1;
- if (DAG.getTarget().getDataLayout()->isLittleEndian()) {
-
- ShiftAmt += i;
-
- if (!isUnary) {
- // Check the rest of the elements to see if they are consecutive.
- for (++i; i != 16; ++i)
- if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt - i))
- return -1;
- } else {
- // Check the rest of the elements to see if they are consecutive.
- for (++i; i != 16; ++i)
- if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt - i) & 15))
- return -1;
- }
-
- } else { // Big Endian
+ ShiftAmt -= i;
+ bool isLE = DAG.getTarget().getSubtargetImpl()->getDataLayout()->
+ isLittleEndian();
+
+ if ((ShuffleKind == 0 && !isLE) || (ShuffleKind == 2 && isLE)) {
+ // Check the rest of the elements to see if they are consecutive.
+ for (++i; i != 16; ++i)
+ if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
+ return -1;
+ } else if (ShuffleKind == 1) {
+ // Check the rest of the elements to see if they are consecutive.
+ for (++i; i != 16; ++i)
+ if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
+ return -1;
+ } else
+ return -1;
- ShiftAmt -= i;
+ if (ShuffleKind == 2 && isLE)
+ ShiftAmt = 16 - ShiftAmt;
- if (!isUnary) {
- // Check the rest of the elements to see if they are consecutive.
- for (++i; i != 16; ++i)
- if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i))
- return -1;
- } else {
- // Check the rest of the elements to see if they are consecutive.
- for (++i; i != 16; ++i)
- if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15))
- return -1;
- }
- }
return ShiftAmt;
}
SelectionDAG &DAG) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
assert(isSplatShuffleMask(SVOp, EltSize));
- if (DAG.getTarget().getDataLayout()->isLittleEndian())
+ if (DAG.getSubtarget().getDataLayout()->isLittleEndian())
return (16 / EltSize) - 1 - (SVOp->getMaskElt(0) / EltSize);
else
return SVOp->getMaskElt(0) / EltSize;
if (Model == TLSModel::GeneralDynamic) {
SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
- SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
- SDValue GOTEntryHi = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT,
- GOTReg, TGA);
+ SDValue GOTPtr;
+ if (is64bit) {
+ SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
+ GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT,
+ GOTReg, TGA);
+ } else {
+ GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
+ }
SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSGD_L, dl, PtrVT,
- GOTEntryHi, TGA);
+ GOTPtr, TGA);
// We need a chain node, and don't have one handy. The underlying
// call has no side effects, so using the function entry node
// suffices.
SDValue Chain = DAG.getEntryNode();
- Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, GOTEntry);
- SDValue ParmReg = DAG.getRegister(PPC::X3, MVT::i64);
+ Chain = DAG.getCopyToReg(Chain, dl,
+ is64bit ? PPC::X3 : PPC::R3, GOTEntry);
+ SDValue ParmReg = DAG.getRegister(is64bit ? PPC::X3 : PPC::R3,
+ is64bit ? MVT::i64 : MVT::i32);
SDValue TLSAddr = DAG.getNode(PPCISD::GET_TLS_ADDR, dl,
PtrVT, ParmReg, TGA);
// The return value from GET_TLS_ADDR really is in X3 already, but
// some hacks are needed here to tie everything together. The extra
// copies dissolve during subsequent transforms.
- Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, TLSAddr);
- return DAG.getCopyFromReg(Chain, dl, PPC::X3, PtrVT);
+ Chain = DAG.getCopyToReg(Chain, dl, is64bit ? PPC::X3 : PPC::R3, TLSAddr);
+ return DAG.getCopyFromReg(Chain, dl, is64bit ? PPC::X3 : PPC::R3, PtrVT);
}
if (Model == TLSModel::LocalDynamic) {
SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0);
- SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
- SDValue GOTEntryHi = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT,
- GOTReg, TGA);
+ SDValue GOTPtr;
+ if (is64bit) {
+ SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64);
+ GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT,
+ GOTReg, TGA);
+ } else {
+ GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT);
+ }
SDValue GOTEntry = DAG.getNode(PPCISD::ADDI_TLSLD_L, dl, PtrVT,
- GOTEntryHi, TGA);
+ GOTPtr, TGA);
// We need a chain node, and don't have one handy. The underlying
// call has no side effects, so using the function entry node
// suffices.
SDValue Chain = DAG.getEntryNode();
- Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, GOTEntry);
- SDValue ParmReg = DAG.getRegister(PPC::X3, MVT::i64);
+ Chain = DAG.getCopyToReg(Chain, dl,
+ is64bit ? PPC::X3 : PPC::R3, GOTEntry);
+ SDValue ParmReg = DAG.getRegister(is64bit ? PPC::X3 : PPC::R3,
+ is64bit ? MVT::i64 : MVT::i32);
SDValue TLSAddr = DAG.getNode(PPCISD::GET_TLSLD_ADDR, dl,
PtrVT, ParmReg, TGA);
// The return value from GET_TLSLD_ADDR really is in X3 already, but
// some hacks are needed here to tie everything together. The extra
// copies dissolve during subsequent transforms.
- Chain = DAG.getCopyToReg(Chain, dl, PPC::X3, TLSAddr);
+ Chain = DAG.getCopyToReg(Chain, dl, is64bit ? PPC::X3 : PPC::R3, TLSAddr);
SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl, PtrVT,
Chain, ParmReg, TGA);
return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA);
// gpr_index
SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
VAListPtr, MachinePointerInfo(SV), MVT::i8,
- false, false, 0);
+ false, false, false, 0);
InChain = GprIndex.getValue(1);
if (VT == MVT::i64) {
// fpr
SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain,
FprPtr, MachinePointerInfo(SV), MVT::i8,
- false, false, 0);
+ false, false, false, 0);
InChain = FprIndex.getValue(1);
SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr,
unsigned ArgSize = ArgVT.getStoreSize();
if (Flags.isByVal())
ArgSize = Flags.getByValSize();
- ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
+
+ // Round up to multiples of the pointer size, except for array members,
+ // which are always packed.
+ if (!Flags.isInConsecutiveRegs())
+ ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
return ArgSize;
}
/// CalculateStackSlotAlignment - Calculates the alignment of this argument
/// on the stack.
-static unsigned CalculateStackSlotAlignment(EVT ArgVT, ISD::ArgFlagsTy Flags,
+static unsigned CalculateStackSlotAlignment(EVT ArgVT, EVT OrigVT,
+ ISD::ArgFlagsTy Flags,
unsigned PtrByteSize) {
unsigned Align = PtrByteSize;
}
}
+ // Array members are always packed to their original alignment.
+ if (Flags.isInConsecutiveRegs()) {
+ // If the array member was split into multiple registers, the first
+ // needs to be aligned to the size of the full type. (Except for
+ // ppcf128, which is only aligned as its f64 components.)
+ if (Flags.isSplit() && OrigVT != MVT::ppcf128)
+ Align = OrigVT.getStoreSize();
+ else
+ Align = ArgVT.getStoreSize();
+ }
+
return Align;
}
/// stack slot (instead of being passed in registers). ArgOffset,
/// AvailableFPRs, and AvailableVRs must hold the current argument
/// position, and will be updated to account for this argument.
-static bool CalculateStackSlotUsed(EVT ArgVT, ISD::ArgFlagsTy Flags,
+static bool CalculateStackSlotUsed(EVT ArgVT, EVT OrigVT,
+ ISD::ArgFlagsTy Flags,
unsigned PtrByteSize,
unsigned LinkageSize,
unsigned ParamAreaSize,
bool UseMemory = false;
// Respect alignment of argument on the stack.
- unsigned Align = CalculateStackSlotAlignment(ArgVT, Flags, PtrByteSize);
+ unsigned Align =
+ CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
// If there's no space left in the argument save area, we must
// use memory (this check also catches zero-sized arguments).
// Allocate argument on the stack.
ArgOffset += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
+ if (Flags.isInConsecutiveRegsLast())
+ ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
// If we overran the argument save area, we must use memory
// (this check catches arguments passed partially in memory)
if (ArgOffset > LinkageSize + ParamAreaSize)
/// ensure minimum alignment required for target.
static unsigned EnsureStackAlignment(const TargetMachine &Target,
unsigned NumBytes) {
- unsigned TargetAlign = Target.getFrameLowering()->getStackAlignment();
+ unsigned TargetAlign =
+ Target.getSubtargetImpl()->getFrameLowering()->getStackAlignment();
unsigned AlignMask = TargetAlign - 1;
NumBytes = (NumBytes + AlignMask) & ~AlignMask;
return NumBytes;
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- getTargetMachine(), ArgLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
+ *DAG.getContext());
// Reserve space for the linkage area on the stack.
unsigned LinkageSize = PPCFrameLowering::getLinkageSize(false, false, false);
// caller's stack frame, right above the parameter list area.
SmallVector<CCValAssign, 16> ByValArgLocs;
CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- getTargetMachine(), ByValArgLocs, *DAG.getContext());
+ ByValArgLocs, *DAG.getContext());
// Reserve stack space for the allocations in CCInfo.
CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
unsigned AvailableFPRs = Num_FPR_Regs;
unsigned AvailableVRs = Num_VR_Regs;
for (unsigned i = 0, e = Ins.size(); i != e; ++i)
- if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].Flags,
+ if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].ArgVT, Ins[i].Flags,
PtrByteSize, LinkageSize, ParamAreaSize,
NumBytes, AvailableFPRs, AvailableVRs))
HasParameterArea = true;
SDValue ArgVal;
bool needsLoad = false;
EVT ObjectVT = Ins[ArgNo].VT;
+ EVT OrigVT = Ins[ArgNo].ArgVT;
unsigned ObjSize = ObjectVT.getStoreSize();
unsigned ArgSize = ObjSize;
ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags;
/* Respect alignment of argument on the stack. */
unsigned Align =
- CalculateStackSlotAlignment(ObjectVT, Flags, PtrByteSize);
+ CalculateStackSlotAlignment(ObjectVT, OrigVT, Flags, PtrByteSize);
ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
unsigned CurArgOffset = ArgOffset;
int FI;
if (HasParameterArea ||
ArgSize + ArgOffset > LinkageSize + Num_GPR_Regs * PtrByteSize)
- FI = MFI->CreateFixedObject(ArgSize, ArgOffset, true);
+ FI = MFI->CreateFixedObject(ArgSize, ArgOffset, false);
else
FI = MFI->CreateStackObject(ArgSize, Align, false);
SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
case MVT::i1:
case MVT::i32:
case MVT::i64:
+ // These can be scalar arguments or elements of an integer array type
+ // passed directly. Clang may use those instead of "byval" aggregate
+ // types to avoid forcing arguments to memory unnecessarily.
if (GPR_idx != Num_GPR_Regs) {
unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
case MVT::f32:
case MVT::f64:
+ // These can be scalar arguments or elements of a float array type
+ // passed directly. The latter are used to implement ELFv2 homogenous
+ // float aggregates.
if (FPR_idx != Num_FPR_Regs) {
unsigned VReg;
ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT);
++FPR_idx;
+ } else if (GPR_idx != Num_GPR_Regs) {
+ // This can only ever happen in the presence of f32 array types,
+ // since otherwise we never run out of FPRs before running out
+ // of GPRs.
+ unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass);
+ ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
+
+ if (ObjectVT == MVT::f32) {
+ if ((ArgOffset % PtrByteSize) == (isLittleEndian ? 4 : 0))
+ ArgVal = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgVal,
+ DAG.getConstant(32, MVT::i32));
+ ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal);
+ }
+
+ ArgVal = DAG.getNode(ISD::BITCAST, dl, ObjectVT, ArgVal);
} else {
needsLoad = true;
- ArgSize = PtrByteSize;
}
- ArgOffset += 8;
+ // When passing an array of floats, the array occupies consecutive
+ // space in the argument area; only round up to the next doubleword
+ // at the end of the array. Otherwise, each float takes 8 bytes.
+ ArgSize = Flags.isInConsecutiveRegs() ? ObjSize : PtrByteSize;
+ ArgOffset += ArgSize;
+ if (Flags.isInConsecutiveRegsLast())
+ ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
break;
case MVT::v4f32:
case MVT::v4i32:
case MVT::v16i8:
case MVT::v2f64:
case MVT::v2i64:
+ // These can be scalar arguments or elements of a vector array type
+ // passed directly. The latter are used to implement ELFv2 homogenous
+ // vector aggregates.
if (VR_idx != Num_VR_Regs) {
unsigned VReg = (ObjectVT == MVT::v2f64 || ObjectVT == MVT::v2i64) ?
MF.addLiveIn(VSRH[VR_idx], &PPC::VSHRCRegClass) :
SmallVectorImpl<SDValue> &InVals) const {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- getTargetMachine(), RVLocs, *DAG.getContext());
+ CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
+ *DAG.getContext());
CCRetInfo.AnalyzeCallResult(Ins, RetCC_PPC);
// Copy all of the result registers out of their specified physreg.
getTargetMachine().Options.GuaranteedTailCallOpt) ? NumBytes : 0;
// Add a register mask operand representing the call-preserved registers.
- const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
+ const TargetRegisterInfo *TRI =
+ getTargetMachine().getSubtargetImpl()->getRegisterInfo();
const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
assert(Mask && "Missing call preserved mask for calling convention");
Ops.push_back(DAG.getRegisterMask(Mask));
// Assign locations to all of the outgoing arguments.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- getTargetMachine(), ArgLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
+ *DAG.getContext());
// Reserve space for the linkage area on the stack.
CCInfo.AllocateStack(PPCFrameLowering::getLinkageSize(false, false, false),
// Assign locations to all of the outgoing aggregate by value arguments.
SmallVector<CCValAssign, 16> ByValArgLocs;
CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- getTargetMachine(), ByValArgLocs, *DAG.getContext());
+ ByValArgLocs, *DAG.getContext());
// Reserve stack space for the allocations in CCInfo.
CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrByteSize);
for (unsigned i = 0; i != NumOps; ++i) {
ISD::ArgFlagsTy Flags = Outs[i].Flags;
EVT ArgVT = Outs[i].VT;
+ EVT OrigVT = Outs[i].ArgVT;
/* Respect alignment of argument on the stack. */
- unsigned Align = CalculateStackSlotAlignment(ArgVT, Flags, PtrByteSize);
+ unsigned Align =
+ CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
NumBytes = ((NumBytes + Align - 1) / Align) * Align;
NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize);
+ if (Flags.isInConsecutiveRegsLast())
+ NumBytes = ((NumBytes + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
}
unsigned NumBytesActuallyUsed = NumBytes;
for (unsigned i = 0; i != NumOps; ++i) {
SDValue Arg = OutVals[i];
ISD::ArgFlagsTy Flags = Outs[i].Flags;
+ EVT ArgVT = Outs[i].VT;
+ EVT OrigVT = Outs[i].ArgVT;
/* Respect alignment of argument on the stack. */
unsigned Align =
- CalculateStackSlotAlignment(Outs[i].VT, Flags, PtrByteSize);
+ CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize);
ArgOffset = ((ArgOffset + Align - 1) / Align) * Align;
/* Compute GPR index associated with argument offset. */
if (GPR_idx != NumGPRs) {
SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
MachinePointerInfo(), VT,
- false, false, 0);
+ false, false, false, 0);
MemOpChains.push_back(Load.getValue(1));
RegsToPass.push_back(std::make_pair(GPR[GPR_idx], Load));
case MVT::i1:
case MVT::i32:
case MVT::i64:
+ // These can be scalar arguments or elements of an integer array type
+ // passed directly. Clang may use those instead of "byval" aggregate
+ // types to avoid forcing arguments to memory unnecessarily.
if (GPR_idx != NumGPRs) {
RegsToPass.push_back(std::make_pair(GPR[GPR_idx], Arg));
} else {
ArgOffset += PtrByteSize;
break;
case MVT::f32:
- case MVT::f64:
- if (FPR_idx != NumFPRs) {
+ case MVT::f64: {
+ // These can be scalar arguments or elements of a float array type
+ // passed directly. The latter are used to implement ELFv2 homogenous
+ // float aggregates.
+
+ // Named arguments go into FPRs first, and once they overflow, the
+ // remaining arguments go into GPRs and then the parameter save area.
+ // Unnamed arguments for vararg functions always go to GPRs and
+ // then the parameter save area. For now, put all arguments to vararg
+ // routines always in both locations (FPR *and* GPR or stack slot).
+ bool NeedGPROrStack = isVarArg || FPR_idx == NumFPRs;
+
+ // First load the argument into the next available FPR.
+ if (FPR_idx != NumFPRs)
RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
- if (isVarArg) {
- // A single float or an aggregate containing only a single float
- // must be passed right-justified in the stack doubleword, and
- // in the GPR, if one is available.
- SDValue StoreOff;
- if (Arg.getSimpleValueType().SimpleTy == MVT::f32 &&
- !isLittleEndian) {
- SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
- StoreOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
- } else
- StoreOff = PtrOff;
-
- SDValue Store = DAG.getStore(Chain, dl, Arg, StoreOff,
- MachinePointerInfo(), false, false, 0);
- MemOpChains.push_back(Store);
-
- // Float varargs are always shadowed in available integer registers
- if (GPR_idx != NumGPRs) {
- SDValue Load = DAG.getLoad(PtrVT, dl, Store, PtrOff,
- MachinePointerInfo(), false, false,
- false, 0);
- MemOpChains.push_back(Load.getValue(1));
- RegsToPass.push_back(std::make_pair(GPR[GPR_idx], Load));
- }
- }
+ // Next, load the argument into GPR or stack slot if needed.
+ if (!NeedGPROrStack)
+ ;
+ else if (GPR_idx != NumGPRs) {
+ // In the non-vararg case, this can only ever happen in the
+ // presence of f32 array types, since otherwise we never run
+ // out of FPRs before running out of GPRs.
+ SDValue ArgVal;
+
+ // Double values are always passed in a single GPR.
+ if (Arg.getValueType() != MVT::f32) {
+ ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
+
+ // Non-array float values are extended and passed in a GPR.
+ } else if (!Flags.isInConsecutiveRegs()) {
+ ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
+ ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
+
+ // If we have an array of floats, we collect every odd element
+ // together with its predecessor into one GPR.
+ } else if (ArgOffset % PtrByteSize != 0) {
+ SDValue Lo, Hi;
+ Lo = DAG.getNode(ISD::BITCAST, dl, MVT::i32, OutVals[i - 1]);
+ Hi = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
+ if (!isLittleEndian)
+ std::swap(Lo, Hi);
+ ArgVal = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
+
+ // The final element, if even, goes into the first half of a GPR.
+ } else if (Flags.isInConsecutiveRegsLast()) {
+ ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg);
+ ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal);
+ if (!isLittleEndian)
+ ArgVal = DAG.getNode(ISD::SHL, dl, MVT::i64, ArgVal,
+ DAG.getConstant(32, MVT::i32));
+
+ // Non-final even elements are skipped; they will be handled
+ // together the with subsequent argument on the next go-around.
+ } else
+ ArgVal = SDValue();
+
+ if (ArgVal.getNode())
+ RegsToPass.push_back(std::make_pair(GPR[GPR_idx], ArgVal));
} else {
// Single-precision floating-point values are mapped to the
// second (rightmost) word of the stack doubleword.
- if (Arg.getValueType() == MVT::f32 && !isLittleEndian) {
+ if (Arg.getValueType() == MVT::f32 &&
+ !isLittleEndian && !Flags.isInConsecutiveRegs()) {
SDValue ConstFour = DAG.getConstant(4, PtrOff.getValueType());
PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour);
}
true, isTailCall, false, MemOpChains,
TailCallArguments, dl);
}
- ArgOffset += 8;
+ // When passing an array of floats, the array occupies consecutive
+ // space in the argument area; only round up to the next doubleword
+ // at the end of the array. Otherwise, each float takes 8 bytes.
+ ArgOffset += (Arg.getValueType() == MVT::f32 &&
+ Flags.isInConsecutiveRegs()) ? 4 : 8;
+ if (Flags.isInConsecutiveRegsLast())
+ ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize;
break;
+ }
case MVT::v4f32:
case MVT::v4i32:
case MVT::v8i16:
case MVT::v16i8:
case MVT::v2f64:
case MVT::v2i64:
+ // These can be scalar arguments or elements of a vector array type
+ // passed directly. The latter are used to implement ELFv2 homogenous
+ // vector aggregates.
+
// For a varargs call, named arguments go into VRs or on the stack as
// usual; unnamed arguments always go to the stack or the corresponding
// GPRs when within range. For now, we always put the value in both
if (GPR_idx != NumGPRs) {
SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg,
MachinePointerInfo(), VT,
- false, false, 0);
+ false, false, false, 0);
MemOpChains.push_back(Load.getValue(1));
RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
const SmallVectorImpl<ISD::OutputArg> &Outs,
LLVMContext &Context) const {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
- RVLocs, Context);
+ CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
return CCInfo.CheckReturn(Outs, RetCC_PPC);
}
SDLoc dl, SelectionDAG &DAG) const {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- getTargetMachine(), RVLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
+ *DAG.getContext());
CCInfo.AnalyzeReturn(Outs, RetCC_PPC);
SDValue Flag;
if (PPC::isSplatShuffleMask(SVOp, 1) ||
PPC::isSplatShuffleMask(SVOp, 2) ||
PPC::isSplatShuffleMask(SVOp, 4) ||
- PPC::isVPKUWUMShuffleMask(SVOp, true, DAG) ||
- PPC::isVPKUHUMShuffleMask(SVOp, true, DAG) ||
- PPC::isVSLDOIShuffleMask(SVOp, true, DAG) != -1 ||
- PPC::isVMRGLShuffleMask(SVOp, 1, true, DAG) ||
- PPC::isVMRGLShuffleMask(SVOp, 2, true, DAG) ||
- PPC::isVMRGLShuffleMask(SVOp, 4, true, DAG) ||
- PPC::isVMRGHShuffleMask(SVOp, 1, true, DAG) ||
- PPC::isVMRGHShuffleMask(SVOp, 2, true, DAG) ||
- PPC::isVMRGHShuffleMask(SVOp, 4, true, DAG)) {
+ PPC::isVPKUWUMShuffleMask(SVOp, 1, DAG) ||
+ PPC::isVPKUHUMShuffleMask(SVOp, 1, DAG) ||
+ PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) != -1 ||
+ PPC::isVMRGLShuffleMask(SVOp, 1, 1, DAG) ||
+ PPC::isVMRGLShuffleMask(SVOp, 2, 1, DAG) ||
+ PPC::isVMRGLShuffleMask(SVOp, 4, 1, DAG) ||
+ PPC::isVMRGHShuffleMask(SVOp, 1, 1, DAG) ||
+ PPC::isVMRGHShuffleMask(SVOp, 2, 1, DAG) ||
+ PPC::isVMRGHShuffleMask(SVOp, 4, 1, DAG)) {
return Op;
}
}
// Altivec has a variety of "shuffle immediates" that take two vector inputs
// and produce a fixed permutation. If any of these match, do not lower to
// VPERM.
- if (PPC::isVPKUWUMShuffleMask(SVOp, false, DAG) ||
- PPC::isVPKUHUMShuffleMask(SVOp, false, DAG) ||
- PPC::isVSLDOIShuffleMask(SVOp, false, DAG) != -1 ||
- PPC::isVMRGLShuffleMask(SVOp, 1, false, DAG) ||
- PPC::isVMRGLShuffleMask(SVOp, 2, false, DAG) ||
- PPC::isVMRGLShuffleMask(SVOp, 4, false, DAG) ||
- PPC::isVMRGHShuffleMask(SVOp, 1, false, DAG) ||
- PPC::isVMRGHShuffleMask(SVOp, 2, false, DAG) ||
- PPC::isVMRGHShuffleMask(SVOp, 4, false, DAG))
+ unsigned int ShuffleKind = isLittleEndian ? 2 : 0;
+ if (PPC::isVPKUWUMShuffleMask(SVOp, ShuffleKind, DAG) ||
+ PPC::isVPKUHUMShuffleMask(SVOp, ShuffleKind, DAG) ||
+ PPC::isVSLDOIShuffleMask(SVOp, ShuffleKind, DAG) != -1 ||
+ PPC::isVMRGLShuffleMask(SVOp, 1, ShuffleKind, DAG) ||
+ PPC::isVMRGLShuffleMask(SVOp, 2, ShuffleKind, DAG) ||
+ PPC::isVMRGLShuffleMask(SVOp, 4, ShuffleKind, DAG) ||
+ PPC::isVMRGHShuffleMask(SVOp, 1, ShuffleKind, DAG) ||
+ PPC::isVMRGHShuffleMask(SVOp, 2, ShuffleKind, DAG) ||
+ PPC::isVMRGHShuffleMask(SVOp, 4, ShuffleKind, DAG))
return Op;
// Check to see if this is a shuffle of 4-byte values. If so, we can use our
PPCTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
bool is64bit, unsigned BinOpcode) const {
// This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
- const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ const TargetInstrInfo *TII =
+ getTargetMachine().getSubtargetImpl()->getInstrInfo();
const BasicBlock *LLVM_BB = BB->getBasicBlock();
MachineFunction *F = BB->getParent();
bool is8bit, // operation
unsigned BinOpcode) const {
// This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
- const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ const TargetInstrInfo *TII =
+ getTargetMachine().getSubtargetImpl()->getInstrInfo();
// In 64 bit mode we have to use 64 bits for addresses, even though the
// lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address
// registers without caring whether they're 32 or 64, but here we're
PPCTargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
MachineBasicBlock *MBB) const {
DebugLoc DL = MI->getDebugLoc();
- const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ const TargetInstrInfo *TII =
+ getTargetMachine().getSubtargetImpl()->getInstrInfo();
MachineFunction *MF = MBB->getParent();
MachineRegisterInfo &MRI = MF->getRegInfo();
// Setup
MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::BCLalways)).addMBB(mainMBB);
const PPCRegisterInfo *TRI =
- static_cast<const PPCRegisterInfo*>(getTargetMachine().getRegisterInfo());
+ getTargetMachine().getSubtarget<PPCSubtarget>().getRegisterInfo();
MIB.addRegMask(TRI->getNoPreservedMask());
BuildMI(*thisMBB, MI, DL, TII->get(PPC::LI), restoreDstReg).addImm(1);
PPCTargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
MachineBasicBlock *MBB) const {
DebugLoc DL = MI->getDebugLoc();
- const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ const TargetInstrInfo *TII =
+ getTargetMachine().getSubtargetImpl()->getInstrInfo();
MachineFunction *MF = MBB->getParent();
MachineRegisterInfo &MRI = MF->getRegInfo();
return emitEHSjLjLongJmp(MI, BB);
}
- const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ const TargetInstrInfo *TII =
+ getTargetMachine().getSubtargetImpl()->getInstrInfo();
// To "insert" these instructions we actually have to insert their
// control-flow patterns.
Cond.push_back(MI->getOperand(1));
DebugLoc dl = MI->getDebugLoc();
- const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ const TargetInstrInfo *TII =
+ getTargetMachine().getSubtargetImpl()->getInstrInfo();
TII->insertSelect(*BB, MI, dl, MI->getOperand(0).getReg(),
Cond, MI->getOperand(2).getReg(),
MI->getOperand(3).getReg());
return SDValue();
}
-// Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does
-// not enforce equality of the chain operands.
-static bool isConsecutiveLS(LSBaseSDNode *LS, LSBaseSDNode *Base,
+static bool isConsecutiveLSLoc(SDValue Loc, EVT VT, LSBaseSDNode *Base,
unsigned Bytes, int Dist,
SelectionDAG &DAG) {
- EVT VT = LS->getMemoryVT();
if (VT.getSizeInBits() / 8 != Bytes)
return false;
- SDValue Loc = LS->getBasePtr();
SDValue BaseLoc = Base->getBasePtr();
if (Loc.getOpcode() == ISD::FrameIndex) {
if (BaseLoc.getOpcode() != ISD::FrameIndex)
return false;
}
+// Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does
+// not enforce equality of the chain operands.
+static bool isConsecutiveLS(SDNode *N, LSBaseSDNode *Base,
+ unsigned Bytes, int Dist,
+ SelectionDAG &DAG) {
+ if (LSBaseSDNode *LS = dyn_cast<LSBaseSDNode>(N)) {
+ EVT VT = LS->getMemoryVT();
+ SDValue Loc = LS->getBasePtr();
+ return isConsecutiveLSLoc(Loc, VT, Base, Bytes, Dist, DAG);
+ }
+
+ if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) {
+ EVT VT;
+ switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
+ default: return false;
+ case Intrinsic::ppc_altivec_lvx:
+ case Intrinsic::ppc_altivec_lvxl:
+ VT = MVT::v4i32;
+ break;
+ case Intrinsic::ppc_altivec_lvebx:
+ VT = MVT::i8;
+ break;
+ case Intrinsic::ppc_altivec_lvehx:
+ VT = MVT::i16;
+ break;
+ case Intrinsic::ppc_altivec_lvewx:
+ VT = MVT::i32;
+ break;
+ }
+
+ return isConsecutiveLSLoc(N->getOperand(2), VT, Base, Bytes, Dist, DAG);
+ }
+
+ if (N->getOpcode() == ISD::INTRINSIC_VOID) {
+ EVT VT;
+ switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
+ default: return false;
+ case Intrinsic::ppc_altivec_stvx:
+ case Intrinsic::ppc_altivec_stvxl:
+ VT = MVT::v4i32;
+ break;
+ case Intrinsic::ppc_altivec_stvebx:
+ VT = MVT::i8;
+ break;
+ case Intrinsic::ppc_altivec_stvehx:
+ VT = MVT::i16;
+ break;
+ case Intrinsic::ppc_altivec_stvewx:
+ VT = MVT::i32;
+ break;
+ }
+
+ return isConsecutiveLSLoc(N->getOperand(3), VT, Base, Bytes, Dist, DAG);
+ }
+
+ return false;
+}
+
// Return true is there is a nearyby consecutive load to the one provided
// (regardless of alignment). We search up and down the chain, looking though
-// token factors and other loads (but nothing else). As a result, a true
-// results indicates that it is safe to create a new consecutive load adjacent
-// to the load provided.
+// token factors and other loads (but nothing else). As a result, a true result
+// indicates that it is safe to create a new consecutive load adjacent to the
+// load provided.
static bool findConsecutiveLoad(LoadSDNode *LD, SelectionDAG &DAG) {
SDValue Chain = LD->getChain();
EVT VT = LD->getMemoryVT();
if (!Visited.insert(ChainNext))
continue;
- if (LoadSDNode *ChainLD = dyn_cast<LoadSDNode>(ChainNext)) {
+ if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(ChainNext)) {
if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
return true;
if (!Visited.insert(LoadRoot))
continue;
- if (LoadSDNode *ChainLD = dyn_cast<LoadSDNode>(LoadRoot))
+ if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(LoadRoot))
if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG))
return true;
for (SDNode::use_iterator UI = LoadRoot->use_begin(),
UE = LoadRoot->use_end(); UI != UE; ++UI)
- if (((isa<LoadSDNode>(*UI) &&
- cast<LoadSDNode>(*UI)->getChain().getNode() == LoadRoot) ||
+ if (((isa<MemSDNode>(*UI) &&
+ cast<MemSDNode>(*UI)->getChain().getNode() == LoadRoot) ||
UI->getOpcode() == ISD::TokenFactor) && !Visited.count(*UI))
Queue.push_back(*UI);
}
Intrinsic::ppc_altivec_lvsl);
SDValue PermCntl = BuildIntrinsicOp(Intr, Ptr, DAG, dl, MVT::v16i8);
- // Refine the alignment of the original load (a "new" load created here
- // which was identical to the first except for the alignment would be
- // merged with the existing node regardless).
+ // Create the new MMO for the new base load. It is like the original MMO,
+ // but represents an area in memory almost twice the vector size centered
+ // on the original address. If the address is unaligned, we might start
+ // reading up to (sizeof(vector)-1) bytes below the address of the
+ // original unaligned load.
MachineFunction &MF = DAG.getMachineFunction();
- MachineMemOperand *MMO =
- MF.getMachineMemOperand(LD->getPointerInfo(),
- LD->getMemOperand()->getFlags(),
- LD->getMemoryVT().getStoreSize(),
- ABIAlignment);
- LD->refineAlignment(MMO);
- SDValue BaseLoad = SDValue(LD, 0);
+ MachineMemOperand *BaseMMO =
+ MF.getMachineMemOperand(LD->getMemOperand(),
+ -LD->getMemoryVT().getStoreSize()+1,
+ 2*LD->getMemoryVT().getStoreSize()-1);
+
+ // Create the new base load.
+ SDValue LDXIntID = DAG.getTargetConstant(Intrinsic::ppc_altivec_lvx,
+ getPointerTy());
+ SDValue BaseLoadOps[] = { Chain, LDXIntID, Ptr };
+ SDValue BaseLoad =
+ DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
+ DAG.getVTList(MVT::v4i32, MVT::Other),
+ BaseLoadOps, MVT::v4i32, BaseMMO);
// Note that the value of IncOffset (which is provided to the next
// load's pointer info offset value, and thus used to calculate the
SDValue Increment = DAG.getConstant(IncValue, getPointerTy());
Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
+ MachineMemOperand *ExtraMMO =
+ MF.getMachineMemOperand(LD->getMemOperand(),
+ 1, 2*LD->getMemoryVT().getStoreSize()-1);
+ SDValue ExtraLoadOps[] = { Chain, LDXIntID, Ptr };
SDValue ExtraLoad =
- DAG.getLoad(VT, dl, Chain, Ptr,
- LD->getPointerInfo().getWithOffset(IncOffset),
- LD->isVolatile(), LD->isNonTemporal(),
- LD->isInvariant(), ABIAlignment);
+ DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl,
+ DAG.getVTList(MVT::v4i32, MVT::Other),
+ ExtraLoadOps, MVT::v4i32, ExtraMMO);
SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
BaseLoad.getValue(1), ExtraLoad.getValue(1));
- if (BaseLoad.getValueType() != MVT::v4i32)
- BaseLoad = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, BaseLoad);
-
- if (ExtraLoad.getValueType() != MVT::v4i32)
- ExtraLoad = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, ExtraLoad);
-
// Because vperm has a big-endian bias, we must reverse the order
// of the input vectors and complement the permute control vector
// when generating little endian code. We have already handled the
if (VT != MVT::v4i32)
Perm = DAG.getNode(ISD::BITCAST, dl, VT, Perm);
- // Now we need to be really careful about how we update the users of the
- // original load. We cannot just call DCI.CombineTo (or
- // DAG.ReplaceAllUsesWith for that matter), because the load still has
- // uses created here (the permutation for example) that need to stay.
- SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
- while (UI != UE) {
- SDUse &Use = UI.getUse();
- SDNode *User = *UI;
- // Note: BaseLoad is checked here because it might not be N, but a
- // bitcast of N.
- if (User == Perm.getNode() || User == BaseLoad.getNode() ||
- User == TF.getNode() || Use.getResNo() > 1) {
- ++UI;
- continue;
- }
-
- SDValue To = Use.getResNo() ? TF : Perm;
- ++UI;
-
- SmallVector<SDValue, 8> Ops;
- for (const SDUse &O : User->ops()) {
- if (O == Use)
- Ops.push_back(To);
- else
- Ops.push_back(O);
- }
-
- DAG.UpdateNodeOperands(User, Ops);
- }
-
+ // The output of the permutation is our loaded result, the TokenFactor is
+ // our new chain.
+ DCI.CombineTo(N, Perm, TF);
return SDValue(N, 0);
}
}
// the AsmName field from *RegisterInfo.td, then this would not be necessary.
if (R.first && VT == MVT::i64 && Subtarget.isPPC64() &&
PPC::GPRCRegClass.contains(R.first)) {
- const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
+ const TargetRegisterInfo *TRI =
+ getTargetMachine().getSubtargetImpl()->getRegisterInfo();
return std::make_pair(TRI->getMatchingSuperReg(R.first,
PPC::sub_32, &PPC::G8RCRegClass),
&PPC::G8RCRegClass);
return isInt<16>(Imm) || isUInt<16>(Imm);
}
-bool PPCTargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
- unsigned,
- bool *Fast) const {
+bool PPCTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
+ unsigned,
+ unsigned,
+ bool *Fast) const {
if (DisablePPCUnaligned)
return false;
projects / oota-llvm.git / blobdiff