SDValue
AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const
{
SDValue Data = Op.getOperand(0);
VTSDNode *BaseType = cast<VTSDNode>(Op.getOperand(1));
DebugLoc DL = Op.getDebugLoc();
EVT DVT = Data.getValueType();
EVT BVT = BaseType->getVT();
unsigned baseBits = BVT.getScalarType().getSizeInBits();
unsigned srcBits = DVT.isSimple() ? DVT.getScalarType().getSizeInBits() : 1;
unsigned shiftBits = srcBits - baseBits;
if (srcBits < 32) {
// If the op is less than 32 bits, then it needs to extend to 32bits
// so it can properly keep the upper bits valid.
EVT IVT = genIntType(32, DVT.isVector() ? DVT.getVectorNumElements() : 1);
Data = DAG.getNode(ISD::ZERO_EXTEND, DL, IVT, Data);
shiftBits = 32 - baseBits;
DVT = IVT;
}
SDValue Shift = DAG.getConstant(shiftBits, DVT);
// Shift left by 'Shift' bits.
Data = DAG.getNode(ISD::SHL, DL, DVT, Data, Shift);
// Signed shift Right by 'Shift' bits.
Data = DAG.getNode(ISD::SRA, DL, DVT, Data, Shift);
if (srcBits < 32) {
// Once the sign extension is done, the op needs to be converted to
// its original type.
Data = DAG.getSExtOrTrunc(Data, DL, Op.getOperand(0).getValueType());
}
return Data;
}
SDValue
AMDGPUTargetLowering::LowerSREM16(SDValue Op, SelectionDAG &DAG) const {
DebugLoc DL = Op.getDebugLoc();
EVT OVT = Op.getValueType();
MVT INTTY = MVT::i32;
if (OVT == MVT::v2i16) {
INTTY = MVT::v2i32;
} else if (OVT == MVT::v4i16) {
INTTY = MVT::v4i32;
}
SDValue LHS = DAG.getSExtOrTrunc(Op.getOperand(0), DL, INTTY);
SDValue RHS = DAG.getSExtOrTrunc(Op.getOperand(1), DL, INTTY);
LHS = DAG.getNode(ISD::SREM, DL, INTTY, LHS, RHS);
LHS = DAG.getSExtOrTrunc(LHS, DL, OVT);
return LHS;
}
SDValue
AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const
{
DebugLoc DL = Op.getDebugLoc();
EVT OVT = Op.getValueType();
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
MVT INTTY;
MVT FLTTY;
if (!OVT.isVector()) {
INTTY = MVT::i32;
FLTTY = MVT::f32;
} else if (OVT.getVectorNumElements() == 2) {
INTTY = MVT::v2i32;
FLTTY = MVT::v2f32;
} else if (OVT.getVectorNumElements() == 4) {
INTTY = MVT::v4i32;
FLTTY = MVT::v4f32;
}
unsigned bitsize = OVT.getScalarType().getSizeInBits();
// char|short jq = ia ^ ib;
SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
// jq = jq >> (bitsize - 2)
jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
// jq = jq | 0x1
jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
// jq = (int)jq
jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
// int ia = (int)LHS;
SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
// int ib, (int)RHS;
SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
// float fa = (float)ia;
SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
// float fb = (float)ib;
SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
// float fq = native_divide(fa, fb);
SDValue fq = DAG.getNode(AMDGPUISD::DIV_INF, DL, FLTTY, fa, fb);
// fq = trunc(fq);
fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
// float fqneg = -fq;
SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
// float fr = mad(fqneg, fb, fa);
SDValue fr = DAG.getNode(AMDGPUISD::MAD, DL, FLTTY, fqneg, fb, fa);
// int iq = (int)fq;
SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
// fr = fabs(fr);
fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
// fb = fabs(fb);
fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
// int cv = fr >= fb;
SDValue cv;
if (INTTY == MVT::i32) {
cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
} else {
cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
}
// jq = (cv ? jq : 0);
jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
DAG.getConstant(0, OVT));
// dst = iq + jq;
iq = DAG.getSExtOrTrunc(iq, DL, OVT);
iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
return iq;
}