#include "llvm/Function.h"
#include "llvm/Intrinsics.h"
#include "llvm/CallingConv.h"
+#include "llvm/Type.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
addRegisterClass(MVT::v4f32, SPU::VECREGRegisterClass);
addRegisterClass(MVT::v2f64, SPU::VECREGRegisterClass);
- // "Odd size" vector classes that we're willing to support:
- addRegisterClass(MVT::v2i32, SPU::VECREGRegisterClass);
- addRegisterClass(MVT::v2f32, SPU::VECREGRegisterClass);
-
for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
// Re-emit as a v16i8 vector load
result = DAG.getLoad(MVT::v16i8, dl, the_chain, basePtr,
- LN->getSrcValue(), LN->getSrcValueOffset(),
+ LN->getPointerInfo(),
LN->isVolatile(), LN->isNonTemporal(), 16);
// Update the chain
if (alignment == 16) {
ConstantSDNode *CN;
-
// Special cases for a known aligned load to simplify the base pointer
// and insertion byte:
if (basePtr.getOpcode() == ISD::ADD
insertEltOffs = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
basePtr,
DAG.getConstant(0, PtrVT));
+ basePtr = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
+ basePtr,
+ DAG.getConstant(0, PtrVT));
}
} else {
// Unaligned load: must be more pessimistic about addressing modes:
DAG.getConstant(0, PtrVT));
}
- // Re-emit as a v16i8 vector load
- alignLoadVec = DAG.getLoad(MVT::v16i8, dl, the_chain, basePtr,
- SN->getSrcValue(), SN->getSrcValueOffset(),
+ // Load the memory to which to store.
+ alignLoadVec = DAG.getLoad(vecVT, dl, the_chain, basePtr,
+ SN->getPointerInfo(),
SN->isVolatile(), SN->isNonTemporal(), 16);
// Update the chain
}
#endif
- SDValue insertEltOp =
- DAG.getNode(SPUISD::SHUFFLE_MASK, dl, vecVT, insertEltOffs);
- SDValue vectorizeOp =
- DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, vecVT, theValue);
+ SDValue insertEltOp = DAG.getNode(SPUISD::SHUFFLE_MASK, dl, vecVT,
+ insertEltOffs);
+ SDValue vectorizeOp = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, vecVT,
+ theValue);
result = DAG.getNode(SPUISD::SHUFB, dl, vecVT,
vectorizeOp, alignLoadVec,
MVT::v4i32, insertEltOp));
result = DAG.getStore(the_chain, dl, result, basePtr,
- LN->getSrcValue(), LN->getSrcValueOffset(),
+ LN->getPointerInfo(),
LN->isVolatile(), LN->isNonTemporal(),
LN->getAlignment());
case MVT::v4i32:
case MVT::v8i16:
case MVT::v16i8:
- case MVT::v2i32:
- case MVT::v2f32:
ArgRegClass = &SPU::VECREGRegClass;
break;
}
// or we're forced to do vararg
int FI = MFI->CreateFixedObject(ObjSize, ArgOffset, true);
SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
- ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, NULL, 0, false, false, 0);
+ ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo(),
+ false, false, 0);
ArgOffset += StackSlotSize;
}
SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
unsigned VReg = MF.addLiveIn(ArgRegs[ArgRegIdx], &SPU::R32CRegClass);
SDValue ArgVal = DAG.getRegister(VReg, MVT::v16i8);
- SDValue Store = DAG.getStore(Chain, dl, ArgVal, FIN, NULL, 0,
+ SDValue Store = DAG.getStore(Chain, dl, ArgVal, FIN, MachinePointerInfo(),
false, false, 0);
Chain = Store.getOperand(0);
MemOps.push_back(Store);
if (ArgRegIdx != NumArgRegs) {
RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
} else {
- MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, NULL, 0,
+ MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
+ MachinePointerInfo(),
false, false, 0));
ArgOffset += StackSlotSize;
}
if (Ins.empty())
return Chain;
+ // Now handle the return value(s)
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCRetInfo(CallConv, isVarArg, getTargetMachine(),
+ RVLocs, *DAG.getContext());
+ CCRetInfo.AnalyzeCallResult(Ins, CCC_SPU);
+
+
// If the call has results, copy the values out of the ret val registers.
- switch (Ins[0].VT.getSimpleVT().SimpleTy) {
- default: llvm_unreachable("Unexpected ret value!");
- case MVT::Other: break;
- case MVT::i32:
- if (Ins.size() > 1 && Ins[1].VT == MVT::i32) {
- Chain = DAG.getCopyFromReg(Chain, dl, SPU::R4,
- MVT::i32, InFlag).getValue(1);
- InVals.push_back(Chain.getValue(0));
- Chain = DAG.getCopyFromReg(Chain, dl, SPU::R3, MVT::i32,
- Chain.getValue(2)).getValue(1);
- InVals.push_back(Chain.getValue(0));
- } else {
- Chain = DAG.getCopyFromReg(Chain, dl, SPU::R3, MVT::i32,
- InFlag).getValue(1);
- InVals.push_back(Chain.getValue(0));
- }
- break;
- case MVT::i8:
- case MVT::i16:
- case MVT::i64:
- case MVT::i128:
- case MVT::f32:
- case MVT::f64:
- case MVT::v2f64:
- case MVT::v2i64:
- case MVT::v4f32:
- case MVT::v4i32:
- case MVT::v8i16:
- case MVT::v16i8:
- Chain = DAG.getCopyFromReg(Chain, dl, SPU::R3, Ins[0].VT,
- InFlag).getValue(1);
- InVals.push_back(Chain.getValue(0));
- break;
- }
+ for (unsigned i = 0; i != RVLocs.size(); ++i) {
+ CCValAssign VA = RVLocs[i];
+
+ SDValue Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
+ InFlag);
+ Chain = Val.getValue(1);
+ InFlag = Val.getValue(2);
+ InVals.push_back(Val);
+ }
return Chain;
}
SDValue T = DAG.getConstant(unsigned(SplatBits), VT.getVectorElementType());
return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, T, T, T, T);
}
- case MVT::v2f32:
- case MVT::v2i32: {
- return SDValue();
- }
case MVT::v2i64: {
return SPU::LowerV2I64Splat(VT, DAG, SplatBits, dl);
}
// If we have a single element being moved from V1 to V2, this can be handled
// using the C*[DX] compute mask instructions, but the vector elements have
- // to be monotonically increasing with one exception element.
+ // to be monotonically increasing with one exception element, and the source
+ // slot of the element to move must be the same as the destination.
EVT VecVT = V1.getValueType();
EVT EltVT = VecVT.getVectorElementType();
unsigned EltsFromV2 = 0;
- unsigned V2Elt = 0;
+ unsigned V2EltOffset = 0;
unsigned V2EltIdx0 = 0;
unsigned CurrElt = 0;
unsigned MaxElts = VecVT.getVectorNumElements();
unsigned PrevElt = 0;
- unsigned V0Elt = 0;
bool monotonic = true;
bool rotate = true;
+ int rotamt=0;
EVT maskVT; // which of the c?d instructions to use
if (EltVT == MVT::i8) {
} else if (EltVT == MVT::i16) {
V2EltIdx0 = 8;
maskVT = MVT::v8i16;
- } else if (VecVT == MVT::v2i32 || VecVT == MVT::v2f32 ) {
- V2EltIdx0 = 2;
- maskVT = MVT::v4i32;
} else if (EltVT == MVT::i32 || EltVT == MVT::f32) {
V2EltIdx0 = 4;
maskVT = MVT::v4i32;
if (monotonic) {
if (SrcElt >= V2EltIdx0) {
- if (1 >= (++EltsFromV2)) {
- V2Elt = (V2EltIdx0 - SrcElt) << 2;
- }
+ // TODO: optimize for the monotonic case when several consecutive
+ // elements are taken form V2. Do we ever get such a case?
+ if (EltsFromV2 == 0 && CurrElt == (SrcElt - V2EltIdx0))
+ V2EltOffset = (SrcElt - V2EltIdx0) * (EltVT.getSizeInBits()/8);
+ else
+ monotonic = false;
+ ++EltsFromV2;
} else if (CurrElt != SrcElt) {
monotonic = false;
}
if (PrevElt > 0 && SrcElt < MaxElts) {
if ((PrevElt == SrcElt - 1)
|| (PrevElt == MaxElts - 1 && SrcElt == 0)) {
+ rotamt = SrcElt-i;
PrevElt = SrcElt;
- if (SrcElt == 0)
- V0Elt = i;
} else {
rotate = false;
}
- } else if (i == 0) {
- // First time through, need to keep track of previous element
+ } else if (i == 0 || (PrevElt==0 && SrcElt==1)) {
+ // First time or after a "wrap around"
PrevElt = SrcElt;
} else {
// This isn't a rotation, takes elements from vector 2
// R1 ($sp) is used here only as it is guaranteed to have last bits zero
SDValue Pointer = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
DAG.getRegister(SPU::R1, PtrVT),
- DAG.getConstant(V2Elt, MVT::i32));
+ DAG.getConstant(V2EltOffset, MVT::i32));
SDValue ShufMaskOp = DAG.getNode(SPUISD::SHUFFLE_MASK, dl,
maskVT, Pointer);
return DAG.getNode(SPUISD::SHUFB, dl, V1.getValueType(), V2, V1,
ShufMaskOp);
} else if (rotate) {
- int rotamt = (MaxElts - V0Elt) * EltVT.getSizeInBits()/8;
-
+ if (rotamt < 0)
+ rotamt +=MaxElts;
+ rotamt *= EltVT.getSizeInBits()/8;
return DAG.getNode(SPUISD::ROTBYTES_LEFT, dl, V1.getValueType(),
V1, DAG.getConstant(rotamt, MVT::i16));
} else {
for (unsigned j = 0; j < BytesPerElement; ++j)
ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,MVT::i8));
}
- // For half vectors padd the mask with zeros for the second half.
- // This is needed because mask is assumed to be full vector elsewhere in
- // the SPU backend.
- if(VecVT == MVT::v2i32 || VecVT == MVT::v2f32)
- for( unsigned i = 0; i < 2; ++i )
- {
- for (unsigned j = 0; j < BytesPerElement; ++j)
- ResultMask.push_back(DAG.getConstant(0,MVT::i8));
- }
-
SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i8,
&ResultMask[0], ResultMask.size());
return DAG.getNode(SPUISD::SHUFB, dl, V1.getValueType(), V1, V2, VPermMask);
case MVT::v4f32: n_copies = 4; VT = MVT::f32; break;
case MVT::v2i64: n_copies = 2; VT = MVT::i64; break;
case MVT::v2f64: n_copies = 2; VT = MVT::f64; break;
- case MVT::v2i32: n_copies = 2; VT = MVT::i32; break;
}
SDValue CValue = DAG.getConstant(CN->getZExtValue(), VT);
SDValue IdxOp = Op.getOperand(2);
DebugLoc dl = Op.getDebugLoc();
EVT VT = Op.getValueType();
+ EVT eltVT = ValOp.getValueType();
// use 0 when the lane to insert to is 'undef'
- int64_t Idx=0;
+ int64_t Offset=0;
if (IdxOp.getOpcode() != ISD::UNDEF) {
ConstantSDNode *CN = cast<ConstantSDNode>(IdxOp);
assert(CN != 0 && "LowerINSERT_VECTOR_ELT: Index is not constant!");
- Idx = (CN->getSExtValue());
+ Offset = (CN->getSExtValue()) * eltVT.getSizeInBits()/8;
}
EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
// Use $sp ($1) because it's always 16-byte aligned and it's available:
SDValue Pointer = DAG.getNode(SPUISD::IndirectAddr, dl, PtrVT,
DAG.getRegister(SPU::R1, PtrVT),
- DAG.getConstant(Idx, PtrVT));
+ DAG.getConstant(Offset, PtrVT));
// widen the mask when dealing with half vectors
EVT maskVT = EVT::getVectorVT(*(DAG.getContext()), VT.getVectorElementType(),
128/ VT.getVectorElementType().getSizeInBits());
SDValue Op0 = Op.getOperand(0);
EVT Op0VT = Op0.getValueType();
- if (Op0VT.getSimpleVT() == MVT::i128 && simpleVT == MVT::i64) {
+ if (Op0VT == MVT::i128 && simpleVT == MVT::i64) {
// Create shuffle mask, least significant doubleword of quadword
unsigned maskHigh = 0x08090a0b;
unsigned maskLow = 0x0c0d0e0f;
DAG.getNode(SPUISD::PREFSLOT2VEC, dl, mvt, Op0, Op0),
DAG.getConstant(31, MVT::i32));
+ // reinterpret as a i128 (SHUFB requires it). This gets lowered away.
+ SDValue extended = SDValue(DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
+ dl, Op0VT, Op0,
+ DAG.getTargetConstant(
+ SPU::GPRCRegClass.getID(),
+ MVT::i32)), 0);
// Shuffle bytes - Copy the sign bits into the upper 64 bits
// and the input value into the lower 64 bits.
SDValue extShuffle = DAG.getNode(SPUISD::SHUFB, dl, mvt,
- DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i128, Op0), sraVal, shufMask);
-
+ extended, sraVal, shufMask);
return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i128, extShuffle);
}
return TargetLowering::getConstraintType(ConstraintLetter);
}
+/// Examine constraint type and operand type and determine a weight value.
+/// This object must already have been set up with the operand type
+/// and the current alternative constraint selected.
+TargetLowering::ConstraintWeight
+SPUTargetLowering::getSingleConstraintMatchWeight(
+ AsmOperandInfo &info, const char *constraint) const {
+ ConstraintWeight weight = CW_Invalid;
+ Value *CallOperandVal = info.CallOperandVal;
+ // If we don't have a value, we can't do a match,
+ // but allow it at the lowest weight.
+ if (CallOperandVal == NULL)
+ return CW_Default;
+ // Look at the constraint type.
+ switch (*constraint) {
+ default:
+ weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
+ break;\r
+ //FIXME: Seems like the supported constraint letters were just copied
+ // from PPC, as the following doesn't correspond to the GCC docs.
+ // I'm leaving it so until someone adds the corresponding lowering support.
+ case 'b':
+ case 'r':
+ case 'f':
+ case 'd':
+ case 'v':
+ case 'y':
+ weight = CW_Register;
+ break;
+ }
+ return weight;
+}
+
std::pair<unsigned, const TargetRegisterClass*>
SPUTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
EVT VT) const
// The SPU target isn't yet aware of offsets.
return false;
}
+
+// can we compare to Imm without writing it into a register?
+bool SPUTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
+ //ceqi, cgti, etc. all take s10 operand
+ return isInt<10>(Imm);
+}
+
+bool
+SPUTargetLowering::isLegalAddressingMode(const AddrMode &AM,
+ const Type * ) const{
+
+ // A-form: 18bit absolute address.
+ if (AM.BaseGV && !AM.HasBaseReg && AM.Scale == 0 && AM.BaseOffs == 0)
+ return true;
+
+ // D-form: reg + 14bit offset
+ if (AM.BaseGV ==0 && AM.HasBaseReg && AM.Scale == 0 && isInt<14>(AM.BaseOffs))
+ return true;
+
+ // X-form: reg+reg
+ if (AM.BaseGV == 0 && AM.HasBaseReg && AM.Scale == 1 && AM.BaseOffs ==0)
+ return true;
+
+ return false;
+}
+
projects / oota-llvm.git / blobdiff