// These are the current register classes that are
// supported
- addRegisterClass(MVT::i32, AMDIL::GPRI32RegisterClass);
- addRegisterClass(MVT::f32, AMDIL::GPRF32RegisterClass);
-
- if (STM.device()->isSupported(AMDILDeviceInfo::DoubleOps)) {
- addRegisterClass(MVT::f64, AMDIL::GPRF64RegisterClass);
- addRegisterClass(MVT::v2f64, AMDIL::GPRV2F64RegisterClass);
- }
- if (STM.device()->isSupported(AMDILDeviceInfo::ByteOps)) {
- addRegisterClass(MVT::i8, AMDIL::GPRI8RegisterClass);
- addRegisterClass(MVT::v2i8, AMDIL::GPRV2I8RegisterClass);
- addRegisterClass(MVT::v4i8, AMDIL::GPRV4I8RegisterClass);
- setOperationAction(ISD::Constant , MVT::i8 , Legal);
- }
- if (STM.device()->isSupported(AMDILDeviceInfo::ShortOps)) {
- addRegisterClass(MVT::i16, AMDIL::GPRI16RegisterClass);
- addRegisterClass(MVT::v2i16, AMDIL::GPRV2I16RegisterClass);
- addRegisterClass(MVT::v4i16, AMDIL::GPRV4I16RegisterClass);
- setOperationAction(ISD::Constant , MVT::i16 , Legal);
- }
- addRegisterClass(MVT::v2f32, AMDIL::GPRV2F32RegisterClass);
- addRegisterClass(MVT::v4f32, AMDIL::GPRV4F32RegisterClass);
- addRegisterClass(MVT::v2i32, AMDIL::GPRV2I32RegisterClass);
- addRegisterClass(MVT::v4i32, AMDIL::GPRV4I32RegisterClass);
- if (STM.device()->isSupported(AMDILDeviceInfo::LongOps)) {
- addRegisterClass(MVT::i64, AMDIL::GPRI64RegisterClass);
- addRegisterClass(MVT::v2i64, AMDIL::GPRV2I64RegisterClass);
- }
-
for (unsigned int x = 0; x < numTypes; ++x) {
MVT::SimpleValueType VT = (MVT::SimpleValueType)types[x];
//FIXME: SIGN_EXTEND_INREG is not meaningful for floating point types
// We cannot sextinreg, expand to shifts
setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom);
- setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
- setOperationAction(ISD::FP_ROUND, VT, Expand);
setOperationAction(ISD::SUBE, VT, Expand);
setOperationAction(ISD::SUBC, VT, Expand);
setOperationAction(ISD::ADDE, VT, Expand);
setOperationAction(ISD::BR_JT, VT, Expand);
setOperationAction(ISD::BRIND, VT, Expand);
// TODO: Implement custom UREM/SREM routines
- setOperationAction(ISD::UREM, VT, Expand);
setOperationAction(ISD::SREM, VT, Expand);
- setOperationAction(ISD::UINT_TO_FP, VT, Custom);
- setOperationAction(ISD::FP_TO_UINT, VT, Custom);
setOperationAction(ISD::GlobalAddress, VT, Custom);
setOperationAction(ISD::JumpTable, VT, Custom);
setOperationAction(ISD::ConstantPool, VT, Custom);
- setOperationAction(ISD::SELECT_CC, VT, Custom);
setOperationAction(ISD::SELECT, VT, Custom);
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
setOperationAction(ISD::UMUL_LOHI, VT, Expand);
if (VT != MVT::i64 && VT != MVT::v2i64) {
setOperationAction(ISD::SDIV, VT, Custom);
}
- setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
- setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
}
for (unsigned int x = 0; x < numFloatTypes; ++x) {
MVT::SimpleValueType VT = (MVT::SimpleValueType)FloatTypes[x];
// IL does not have these operations for floating point types
setOperationAction(ISD::FP_ROUND_INREG, VT, Expand);
- setOperationAction(ISD::FP_ROUND, VT, Custom);
setOperationAction(ISD::SETOLT, VT, Expand);
setOperationAction(ISD::SETOGE, VT, Expand);
setOperationAction(ISD::SETOGT, VT, Expand);
// GPU also does not have divrem function for signed or unsigned
setOperationAction(ISD::SDIVREM, VT, Expand);
- setOperationAction(ISD::UDIVREM, VT, Expand);
- setOperationAction(ISD::FP_ROUND, VT, Expand);
// GPU does not have [S|U]MUL_LOHI functions as a single instruction
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
MVT::SimpleValueType VT = (MVT::SimpleValueType)VectorTypes[ii];
setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
- setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
- setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
- setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
- setOperationAction(ISD::FP_ROUND, VT, Expand);
setOperationAction(ISD::SDIVREM, VT, Expand);
- setOperationAction(ISD::UDIVREM, VT, Expand);
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
// setOperationAction(ISD::VSETCC, VT, Expand);
setOperationAction(ISD::SETCC, VT, Expand);
setOperationAction(ISD::SELECT, VT, Expand);
}
- setOperationAction(ISD::FP_ROUND, MVT::Other, Expand);
if (STM.device()->isSupported(AMDILDeviceInfo::LongOps)) {
- if (STM.calVersion() < CAL_VERSION_SC_139
- || STM.device()->getGeneration() == AMDILDeviceInfo::HD4XXX) {
- setOperationAction(ISD::MUL, MVT::i64, Custom);
- }
- setOperationAction(ISD::SUB, MVT::i64, Custom);
- setOperationAction(ISD::ADD, MVT::i64, Custom);
setOperationAction(ISD::MULHU, MVT::i64, Expand);
setOperationAction(ISD::MULHU, MVT::v2i64, Expand);
setOperationAction(ISD::MULHS, MVT::i64, Expand);
setOperationAction(ISD::MULHS, MVT::v2i64, Expand);
- setOperationAction(ISD::MUL, MVT::v2i64, Expand);
- setOperationAction(ISD::SUB, MVT::v2i64, Expand);
setOperationAction(ISD::ADD, MVT::v2i64, Expand);
setOperationAction(ISD::SREM, MVT::v2i64, Expand);
setOperationAction(ISD::Constant , MVT::i64 , Legal);
setOperationAction(ISD::SDIV, MVT::v2i64, Expand);
- setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Expand);
- setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Expand);
setOperationAction(ISD::TRUNCATE, MVT::v2i64, Expand);
setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Expand);
setOperationAction(ISD::ZERO_EXTEND, MVT::v2i64, Expand);
setOperationAction(ISD::FADD, MVT::v2f64, Expand);
setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
- setOperationAction(ISD::FP_ROUND, MVT::v2f64, Expand);
setOperationAction(ISD::FP_ROUND_INREG, MVT::v2f64, Expand);
setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
setOperationAction(ISD::ConstantFP , MVT::f64 , Legal);
- setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
// We want to expand vector conversions into their scalar
// counterparts.
- setOperationAction(ISD::UINT_TO_FP, MVT::v2f64, Expand);
- setOperationAction(ISD::FP_TO_UINT, MVT::v2f64, Expand);
setOperationAction(ISD::TRUNCATE, MVT::v2f64, Expand);
setOperationAction(ISD::SIGN_EXTEND, MVT::v2f64, Expand);
setOperationAction(ISD::ZERO_EXTEND, MVT::v2f64, Expand);
setOperationAction(ISD::BRIND, MVT::Other, Expand);
setOperationAction(ISD::SETCC, MVT::Other, Custom);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Expand);
- setOperationAction(ISD::FDIV, MVT::f32, Custom);
- setOperationAction(ISD::FDIV, MVT::v2f32, Custom);
- setOperationAction(ISD::FDIV, MVT::v4f32, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::Other, Custom);
// Use the default implementation.
LOWER(JumpTable);
LOWER(ConstantPool);
LOWER(ExternalSymbol);
- LOWER(FP_TO_UINT);
- LOWER(UINT_TO_FP);
- LOWER(MUL);
- LOWER(SUB);
- LOWER(FDIV);
LOWER(SDIV);
LOWER(SREM);
- LOWER(UREM);
LOWER(BUILD_VECTOR);
- LOWER(INSERT_VECTOR_ELT);
- LOWER(EXTRACT_VECTOR_ELT);
- LOWER(EXTRACT_SUBVECTOR);
- LOWER(SCALAR_TO_VECTOR);
- LOWER(CONCAT_VECTORS);
LOWER(SELECT);
LOWER(SETCC);
LOWER(SIGN_EXTEND_INREG);
LOWER(DYNAMIC_STACKALLOC);
LOWER(BRCOND);
LOWER(BR_CC);
- LOWER(FP_ROUND);
}
return Op;
}
InVals);
}
-SDValue
-AMDILTargetLowering::genCLZuN(SDValue Op, SelectionDAG &DAG,
- uint32_t bits) const
-{
- DebugLoc DL = Op.getDebugLoc();
- EVT INTTY = Op.getValueType();
- EVT FPTY;
- if (INTTY.isVector()) {
- FPTY = EVT(MVT::getVectorVT(MVT::f32,
- INTTY.getVectorNumElements()));
- } else {
- FPTY = EVT(MVT::f32);
- }
- /* static inline uint
- __clz_Nbit(uint x)
- {
- int xor = 0x3f800000U | x;
- float tp = as_float(xor);
- float t = tp + -1.0f;
- uint tint = as_uint(t);
- int cmp = (x != 0);
- uint tsrc = tint >> 23;
- uint tmask = tsrc & 0xffU;
- uint cst = (103 + N)U - tmask;
- return cmp ? cst : N;
- }
- */
- assert(INTTY.getScalarType().getSimpleVT().SimpleTy == MVT::i32
- && "genCLZu16 only works on 32bit types");
- // uint x = Op
- SDValue x = Op;
- // xornode = 0x3f800000 | x
- SDValue xornode = DAG.getNode(ISD::OR, DL, INTTY,
- DAG.getConstant(0x3f800000, INTTY), x);
- // float tp = as_float(xornode)
- SDValue tp = DAG.getNode(ISDBITCAST, DL, FPTY, xornode);
- // float t = tp + -1.0f
- SDValue t = DAG.getNode(ISD::FADD, DL, FPTY, tp,
- DAG.getConstantFP(-1.0f, FPTY));
- // uint tint = as_uint(t)
- SDValue tint = DAG.getNode(ISDBITCAST, DL, INTTY, t);
- // int cmp = (x != 0)
- SDValue cmp = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETNE, MVT::i32), MVT::i32), x,
- DAG.getConstant(0, INTTY));
- // uint tsrc = tint >> 23
- SDValue tsrc = DAG.getNode(ISD::SRL, DL, INTTY, tint,
- DAG.getConstant(23, INTTY));
- // uint tmask = tsrc & 0xFF
- SDValue tmask = DAG.getNode(ISD::AND, DL, INTTY, tsrc,
- DAG.getConstant(0xFFU, INTTY));
- // uint cst = (103 + bits) - tmask
- SDValue cst = DAG.getNode(ISD::SUB, DL, INTTY,
- DAG.getConstant((103U + bits), INTTY), tmask);
- // return cmp ? cst : N
- cst = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, cmp, cst,
- DAG.getConstant(bits, INTTY));
- return cst;
-}
-
-SDValue
-AMDILTargetLowering::genCLZu32(SDValue Op, SelectionDAG &DAG) const
-{
- SDValue DST = SDValue();
- DebugLoc DL = Op.getDebugLoc();
- EVT INTTY = Op.getValueType();
- const AMDILSubtarget &STM = getTargetMachine().getSubtarget<AMDILSubtarget>();
- if (STM.device()->getGeneration() >= AMDILDeviceInfo::HD5XXX) {
- //__clz_32bit(uint u)
- //{
- // int z = __amdil_ffb_hi(u) ;
- // return z < 0 ? 32 : z;
- // }
- // uint u = op
- SDValue u = Op;
- // int z = __amdil_ffb_hi(u)
- SDValue z = DAG.getNode(AMDILISD::IFFB_HI, DL, INTTY, u);
- // int cmp = z < 0
- SDValue cmp = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETLT, MVT::i32), MVT::i32),
- z, DAG.getConstant(0, INTTY));
- // return cmp ? 32 : z
- DST = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, cmp,
- DAG.getConstant(32, INTTY), z);
- } else if (STM.device()->getGeneration() == AMDILDeviceInfo::HD4XXX) {
- // static inline uint
- //__clz_32bit(uint x)
- //{
- // uint zh = __clz_16bit(x >> 16);
- // uint zl = __clz_16bit(x & 0xffffU);
- // return zh == 16U ? 16U + zl : zh;
- //}
- // uint x = Op
- SDValue x = Op;
- // uint xs16 = x >> 16
- SDValue xs16 = DAG.getNode(ISD::SRL, DL, INTTY, x,
- DAG.getConstant(16, INTTY));
- // uint zh = __clz_16bit(xs16)
- SDValue zh = genCLZuN(xs16, DAG, 16);
- // uint xa16 = x & 0xFFFF
- SDValue xa16 = DAG.getNode(ISD::AND, DL, INTTY, x,
- DAG.getConstant(0xFFFFU, INTTY));
- // uint zl = __clz_16bit(xa16)
- SDValue zl = genCLZuN(xa16, DAG, 16);
- // uint cmp = zh == 16U
- SDValue cmp = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::i32), MVT::i32),
- zh, DAG.getConstant(16U, INTTY));
- // uint zl16 = zl + 16
- SDValue zl16 = DAG.getNode(ISD::ADD, DL, INTTY,
- DAG.getConstant(16, INTTY), zl);
- // return cmp ? zl16 : zh
- DST = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY,
- cmp, zl16, zh);
- } else {
- assert(0 && "Attempting to generate a CLZ function with an"
- " unknown graphics card");
- }
- return DST;
-}
-SDValue
-AMDILTargetLowering::genCLZu64(SDValue Op, SelectionDAG &DAG) const
-{
- SDValue DST = SDValue();
- DebugLoc DL = Op.getDebugLoc();
- EVT INTTY;
- EVT LONGTY = Op.getValueType();
- bool isVec = LONGTY.isVector();
- if (isVec) {
- INTTY = EVT(MVT::getVectorVT(MVT::i32, Op.getValueType()
- .getVectorNumElements()));
- } else {
- INTTY = EVT(MVT::i32);
- }
- const AMDILSubtarget &STM = getTargetMachine().getSubtarget<AMDILSubtarget>();
- if (STM.device()->getGeneration() >= AMDILDeviceInfo::HD5XXX) {
- // Evergreen:
- // static inline uint
- // __clz_u64(ulong x)
- // {
- //uint zhi = __clz_32bit((uint)(x >> 32));
- //uint zlo = __clz_32bit((uint)(x & 0xffffffffUL));
- //return zhi == 32U ? 32U + zlo : zhi;
- //}
- //ulong x = op
- SDValue x = Op;
- // uint xhi = x >> 32
- SDValue xlo = DAG.getNode((isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTTY, x);
- // uint xlo = x & 0xFFFFFFFF
- SDValue xhi = DAG.getNode((isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI, DL, INTTY, x);
- // uint zhi = __clz_32bit(xhi)
- SDValue zhi = genCLZu32(xhi, DAG);
- // uint zlo = __clz_32bit(xlo)
- SDValue zlo = genCLZu32(xlo, DAG);
- // uint cmp = zhi == 32
- SDValue cmp = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::i32), MVT::i32),
- zhi, DAG.getConstant(32U, INTTY));
- // uint zlop32 = 32 + zlo
- SDValue zlop32 = DAG.getNode(AMDILISD::ADD, DL, INTTY,
- DAG.getConstant(32U, INTTY), zlo);
- // return cmp ? zlop32: zhi
- DST = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, cmp, zlop32, zhi);
- } else if (STM.device()->getGeneration() == AMDILDeviceInfo::HD4XXX) {
- // HD4XXX:
- // static inline uint
- //__clz_64bit(ulong x)
- //{
- //uint zh = __clz_23bit((uint)(x >> 46)) - 5U;
- //uint zm = __clz_23bit((uint)(x >> 23) & 0x7fffffU);
- //uint zl = __clz_23bit((uint)x & 0x7fffffU);
- //uint r = zh == 18U ? 18U + zm : zh;
- //return zh + zm == 41U ? 41U + zl : r;
- //}
- //ulong x = Op
- SDValue x = Op;
- // ulong xs46 = x >> 46
- SDValue xs46 = DAG.getNode(ISD::SRL, DL, LONGTY, x,
- DAG.getConstant(46, LONGTY));
- // uint ixs46 = (uint)xs46
- SDValue ixs46 = DAG.getNode(ISD::TRUNCATE, DL, INTTY, xs46);
- // ulong xs23 = x >> 23
- SDValue xs23 = DAG.getNode(ISD::SRL, DL, LONGTY, x,
- DAG.getConstant(23, LONGTY));
- // uint ixs23 = (uint)xs23
- SDValue ixs23 = DAG.getNode(ISD::TRUNCATE, DL, INTTY, xs23);
- // uint xs23m23 = ixs23 & 0x7FFFFF
- SDValue xs23m23 = DAG.getNode(ISD::AND, DL, INTTY, ixs23,
- DAG.getConstant(0x7fffffU, INTTY));
- // uint ix = (uint)x
- SDValue ix = DAG.getNode((isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTTY, x);
- // uint xm23 = ix & 0x7FFFFF
- SDValue xm23 = DAG.getNode(ISD::AND, DL, INTTY, ix,
- DAG.getConstant(0x7fffffU, INTTY));
- // uint zh = __clz_23bit(ixs46)
- SDValue zh = genCLZuN(ixs46, DAG, 23);
- // uint zm = __clz_23bit(xs23m23)
- SDValue zm = genCLZuN(xs23m23, DAG, 23);
- // uint zl = __clz_23bit(xm23)
- SDValue zl = genCLZuN(xm23, DAG, 23);
- // uint zhm5 = zh - 5
- SDValue zhm5 = DAG.getNode(ISD::ADD, DL, INTTY, zh,
- DAG.getConstant(-5U, INTTY));
- SDValue const18 = DAG.getConstant(18, INTTY);
- SDValue const41 = DAG.getConstant(41, INTTY);
- // uint cmp1 = zh = 18
- SDValue cmp1 = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::i32), MVT::i32),
- zhm5, const18);
- // uint zhm5zm = zhm5 + zh
- SDValue zhm5zm = DAG.getNode(ISD::ADD, DL, INTTY, zhm5, zm);
- // uint cmp2 = zhm5zm == 41
- SDValue cmp2 = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::i32), MVT::i32),
- zhm5zm, const41);
- // uint zmp18 = zhm5 + 18
- SDValue zmp18 = DAG.getNode(ISD::ADD, DL, INTTY, zm, const18);
- // uint zlp41 = zl + 41
- SDValue zlp41 = DAG.getNode(ISD::ADD, DL, INTTY, zl, const41);
- // uint r = cmp1 ? zmp18 : zh
- SDValue r = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY,
- cmp1, zmp18, zhm5);
- // return cmp2 ? zlp41 : r
- DST = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, cmp2, zlp41, r);
- } else {
- assert(0 && "Attempting to generate a CLZ function with an"
- " unknown graphics card");
- }
- return DST;
-}
-SDValue
-AMDILTargetLowering::genf64toi64(SDValue RHS, SelectionDAG &DAG,
- bool includeSign) const
-{
- EVT INTVT;
- EVT LONGVT;
- SDValue DST;
- DebugLoc DL = RHS.getDebugLoc();
- EVT RHSVT = RHS.getValueType();
- bool isVec = RHSVT.isVector();
- if (isVec) {
- LONGVT = EVT(MVT::getVectorVT(MVT::i64, RHSVT
- .getVectorNumElements()));
- INTVT = EVT(MVT::getVectorVT(MVT::i32, RHSVT
- .getVectorNumElements()));
- } else {
- LONGVT = EVT(MVT::i64);
- INTVT = EVT(MVT::i32);
- }
- const AMDILSubtarget &STM = getTargetMachine().getSubtarget<AMDILSubtarget>();
- if (STM.device()->getGeneration() > AMDILDeviceInfo::HD6XXX) {
- // unsigned version:
- // uint uhi = (uint)(d * 0x1.0p-32);
- // uint ulo = (uint)(mad((double)uhi, -0x1.0p+32, d));
- // return as_ulong2((uint2)(ulo, uhi));
- //
- // signed version:
- // double ad = fabs(d);
- // long l = unsigned_version(ad);
- // long nl = -l;
- // return d == ad ? l : nl;
- SDValue d = RHS;
- if (includeSign) {
- d = DAG.getNode(ISD::FABS, DL, RHSVT, d);
- }
- SDValue uhid = DAG.getNode(ISD::FMUL, DL, RHSVT, d,
- DAG.getConstantFP(0x2f800000, RHSVT));
- SDValue uhi = DAG.getNode(ISD::FP_TO_UINT, DL, INTVT, uhid);
- SDValue ulod = DAG.getNode(ISD::UINT_TO_FP, DL, RHSVT, uhi);
- ulod = DAG.getNode(AMDILISD::MAD, DL, RHSVT, ulod,
- DAG.getConstantFP(0xcf800000, RHSVT), d);
- SDValue ulo = DAG.getNode(ISD::FP_TO_UINT, DL, INTVT, ulod);
- SDValue l = DAG.getNode((isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT, ulo, uhi);
- if (includeSign) {
- SDValue nl = DAG.getNode(AMDILISD::INEGATE, DL, LONGVT, l);
- SDValue c = DAG.getNode(AMDILISD::CMP, DL, RHSVT,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::f64), MVT::i32),
- RHS, d);
- l = DAG.getNode(AMDILISD::CMOVLOG, DL, LONGVT, c, l, nl);
- }
- DST = l;
- } else {
- /*
- __attribute__((always_inline)) long
- cast_f64_to_i64(double d)
- {
- // Convert d in to 32-bit components
- long x = as_long(d);
- xhi = LCOMPHI(x);
- xlo = LCOMPLO(x);
-
- // Generate 'normalized' mantissa
- mhi = xhi | 0x00100000; // hidden bit
- mhi <<= 11;
- temp = xlo >> (32 - 11);
- mhi |= temp
- mlo = xlo << 11;
-
- // Compute shift right count from exponent
- e = (xhi >> (52-32)) & 0x7ff;
- sr = 1023 + 63 - e;
- srge64 = sr >= 64;
- srge32 = sr >= 32;
-
- // Compute result for 0 <= sr < 32
- rhi0 = mhi >> (sr &31);
- rlo0 = mlo >> (sr &31);
- temp = mhi << (32 - sr);
- temp |= rlo0;
- rlo0 = sr ? temp : rlo0;
-
- // Compute result for 32 <= sr
- rhi1 = 0;
- rlo1 = srge64 ? 0 : rhi0;
-
- // Pick between the 2 results
- rhi = srge32 ? rhi1 : rhi0;
- rlo = srge32 ? rlo1 : rlo0;
-
- // Optional saturate on overflow
- srlt0 = sr < 0;
- rhi = srlt0 ? MAXVALUE : rhi;
- rlo = srlt0 ? MAXVALUE : rlo;
-
- // Create long
- res = LCREATE( rlo, rhi );
-
- // Deal with sign bit (ignoring whether result is signed or unsigned value)
- if (includeSign) {
- sign = ((signed int) xhi) >> 31; fill with sign bit
- sign = LCREATE( sign, sign );
- res += sign;
- res ^= sign;
- }
-
- return res;
- }
- */
- SDValue c11 = DAG.getConstant( 63 - 52, INTVT );
- SDValue c32 = DAG.getConstant( 32, INTVT );
-
- // Convert d in to 32-bit components
- SDValue d = RHS;
- SDValue x = DAG.getNode(ISDBITCAST, DL, LONGVT, d);
- SDValue xhi = DAG.getNode( (isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI, DL, INTVT, x );
- SDValue xlo = DAG.getNode( (isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTVT, x );
-
- // Generate 'normalized' mantissa
- SDValue mhi = DAG.getNode( ISD::OR, DL, INTVT,
- xhi, DAG.getConstant( 0x00100000, INTVT ) );
- mhi = DAG.getNode( ISD::SHL, DL, INTVT, mhi, c11 );
- SDValue temp = DAG.getNode( ISD::SRL, DL, INTVT,
- xlo, DAG.getConstant( 32 - (63 - 52), INTVT ) );
- mhi = DAG.getNode( ISD::OR, DL, INTVT, mhi, temp );
- SDValue mlo = DAG.getNode( ISD::SHL, DL, INTVT, xlo, c11 );
-
- // Compute shift right count from exponent
- SDValue e = DAG.getNode( ISD::SRL, DL, INTVT,
- xhi, DAG.getConstant( 52-32, INTVT ) );
- e = DAG.getNode( ISD::AND, DL, INTVT,
- e, DAG.getConstant( 0x7ff, INTVT ) );
- SDValue sr = DAG.getNode( ISD::SUB, DL, INTVT,
- DAG.getConstant( 1023 + 63, INTVT ), e );
- SDValue srge64 = DAG.getNode( AMDILISD::CMP, DL, INTVT,
- DAG.getConstant(CondCCodeToCC(ISD::SETGE, MVT::i32), MVT::i32),
- sr, DAG.getConstant(64, INTVT));
- SDValue srge32 = DAG.getNode( AMDILISD::CMP, DL, INTVT,
- DAG.getConstant(CondCCodeToCC(ISD::SETGE, MVT::i32), MVT::i32),
- sr, DAG.getConstant(32, INTVT));
-
- // Compute result for 0 <= sr < 32
- SDValue rhi0 = DAG.getNode( ISD::SRL, DL, INTVT, mhi, sr );
- SDValue rlo0 = DAG.getNode( ISD::SRL, DL, INTVT, mlo, sr );
- temp = DAG.getNode( ISD::SUB, DL, INTVT, c32, sr );
- temp = DAG.getNode( ISD::SHL, DL, INTVT, mhi, temp );
- temp = DAG.getNode( ISD::OR, DL, INTVT, rlo0, temp );
- rlo0 = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT, sr, temp, rlo0 );
-
- // Compute result for 32 <= sr
- SDValue rhi1 = DAG.getConstant( 0, INTVT );
- SDValue rlo1 = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT,
- srge64, rhi1, rhi0 );
-
- // Pick between the 2 results
- SDValue rhi = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT,
- srge32, rhi1, rhi0 );
- SDValue rlo = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT,
- srge32, rlo1, rlo0 );
-
- // Create long
- SDValue res = DAG.getNode( (isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT, rlo, rhi );
-
- // Deal with sign bit
- if (includeSign) {
- SDValue sign = DAG.getNode( ISD::SRA, DL, INTVT,
- xhi, DAG.getConstant( 31, INTVT ) );
- sign = DAG.getNode( (isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT, sign, sign );
- res = DAG.getNode( ISD::ADD, DL, LONGVT, res, sign );
- res = DAG.getNode( ISD::XOR, DL, LONGVT, res, sign );
- }
- DST = res;
- }
- return DST;
-}
-SDValue
-AMDILTargetLowering::genf64toi32(SDValue RHS, SelectionDAG &DAG,
- bool includeSign) const
-{
- EVT INTVT;
- EVT LONGVT;
- DebugLoc DL = RHS.getDebugLoc();
- EVT RHSVT = RHS.getValueType();
- bool isVec = RHSVT.isVector();
- if (isVec) {
- LONGVT = EVT(MVT::getVectorVT(MVT::i64,
- RHSVT.getVectorNumElements()));
- INTVT = EVT(MVT::getVectorVT(MVT::i32,
- RHSVT.getVectorNumElements()));
- } else {
- LONGVT = EVT(MVT::i64);
- INTVT = EVT(MVT::i32);
- }
- /*
- __attribute__((always_inline)) int
- cast_f64_to_[u|i]32(double d)
- {
- // Convert d in to 32-bit components
- long x = as_long(d);
- xhi = LCOMPHI(x);
- xlo = LCOMPLO(x);
-
- // Generate 'normalized' mantissa
- mhi = xhi | 0x00100000; // hidden bit
- mhi <<= 11;
- temp = xlo >> (32 - 11);
- mhi |= temp
-
- // Compute shift right count from exponent
- e = (xhi >> (52-32)) & 0x7ff;
- sr = 1023 + 31 - e;
- srge32 = sr >= 32;
-
- // Compute result for 0 <= sr < 32
- res = mhi >> (sr &31);
- res = srge32 ? 0 : res;
-
- // Optional saturate on overflow
- srlt0 = sr < 0;
- res = srlt0 ? MAXVALUE : res;
-
- // Deal with sign bit (ignoring whether result is signed or unsigned value)
- if (includeSign) {
- sign = ((signed int) xhi) >> 31; fill with sign bit
- res += sign;
- res ^= sign;
- }
-
- return res;
- }
- */
- SDValue c11 = DAG.getConstant( 63 - 52, INTVT );
-
- // Convert d in to 32-bit components
- SDValue d = RHS;
- SDValue x = DAG.getNode(ISDBITCAST, DL, LONGVT, d);
- SDValue xhi = DAG.getNode( (isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI, DL, INTVT, x );
- SDValue xlo = DAG.getNode( (isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTVT, x );
-
- // Generate 'normalized' mantissa
- SDValue mhi = DAG.getNode( ISD::OR, DL, INTVT,
- xhi, DAG.getConstant( 0x00100000, INTVT ) );
- mhi = DAG.getNode( ISD::SHL, DL, INTVT, mhi, c11 );
- SDValue temp = DAG.getNode( ISD::SRL, DL, INTVT,
- xlo, DAG.getConstant( 32 - (63 - 52), INTVT ) );
- mhi = DAG.getNode( ISD::OR, DL, INTVT, mhi, temp );
-
- // Compute shift right count from exponent
- SDValue e = DAG.getNode( ISD::SRL, DL, INTVT,
- xhi, DAG.getConstant( 52-32, INTVT ) );
- e = DAG.getNode( ISD::AND, DL, INTVT,
- e, DAG.getConstant( 0x7ff, INTVT ) );
- SDValue sr = DAG.getNode( ISD::SUB, DL, INTVT,
- DAG.getConstant( 1023 + 31, INTVT ), e );
- SDValue srge32 = DAG.getNode( AMDILISD::CMP, DL, INTVT,
- DAG.getConstant(CondCCodeToCC(ISD::SETGE, MVT::i32), MVT::i32),
- sr, DAG.getConstant(32, INTVT));
-
- // Compute result for 0 <= sr < 32
- SDValue res = DAG.getNode( ISD::SRL, DL, INTVT, mhi, sr );
- res = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT,
- srge32, DAG.getConstant(0,INTVT), res );
-
- // Deal with sign bit
- if (includeSign) {
- SDValue sign = DAG.getNode( ISD::SRA, DL, INTVT,
- xhi, DAG.getConstant( 31, INTVT ) );
- res = DAG.getNode( ISD::ADD, DL, INTVT, res, sign );
- res = DAG.getNode( ISD::XOR, DL, INTVT, res, sign );
- }
- return res;
-}
-
-SDValue
-AMDILTargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const
-{
- SDValue DST;
- SDValue RHS = Op.getOperand(0);
- EVT RHSVT = RHS.getValueType();
- MVT RST = RHSVT.getScalarType().getSimpleVT();
- EVT LHSVT = Op.getValueType();
- MVT LST = LHSVT.getScalarType().getSimpleVT();
- DebugLoc DL = Op.getDebugLoc();
- const AMDILSubtarget &STM = getTargetMachine().getSubtarget<AMDILSubtarget>();
- if (RST == MVT::f64 && RHSVT.isVector()
- && STM.device()->getGeneration() > AMDILDeviceInfo::HD6XXX) {
- // We dont support vector 64bit floating point convertions.
- for (unsigned x = 0, y = RHSVT.getVectorNumElements(); x < y; ++x) {
- SDValue op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
- DL, RST, RHS, DAG.getTargetConstant(x, MVT::i32));
- op = DAG.getNode(ISD::FP_TO_SINT, DL, LST, op);
- if (!x) {
- DST = DAG.getNode(AMDILISD::VBUILD, DL, LHSVT, op);
- } else {
- DST = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, LHSVT,
- DST, op, DAG.getTargetConstant(x, MVT::i32));
- }
-
- }
- } else {
- if (RST == MVT::f64
- && LST == MVT::i32) {
- if (STM.device()->getGeneration() > AMDILDeviceInfo::HD6XXX) {
- DST = SDValue(Op.getNode(), 0);
- } else {
- DST = genf64toi32(RHS, DAG, false);
- }
- } else if (RST == MVT::f64
- && LST == MVT::i64) {
- DST = genf64toi64(RHS, DAG, false);
- } else if (RST == MVT::f64
- && (LST == MVT::i8 || LST == MVT::i16)) {
- if (STM.device()->getGeneration() > AMDILDeviceInfo::HD6XXX) {
- DST = DAG.getNode(ISD::TRUNCATE, DL, LHSVT, SDValue(Op.getNode(), 0));
- } else {
- SDValue ToInt = genf64toi32(RHS, DAG, false);
- DST = DAG.getNode(ISD::TRUNCATE, DL, LHSVT, ToInt);
- }
-
- } else {
- DST = SDValue(Op.getNode(), 0);
- }
- }
- return DST;
-}
-SDValue
-AMDILTargetLowering::genu32tof64(SDValue RHS, EVT LHSVT,
- SelectionDAG &DAG) const
-{
- EVT RHSVT = RHS.getValueType();
- DebugLoc DL = RHS.getDebugLoc();
- EVT INTVT;
- EVT LONGVT;
- bool isVec = RHSVT.isVector();
- if (isVec) {
- LONGVT = EVT(MVT::getVectorVT(MVT::i64,
- RHSVT.getVectorNumElements()));
- INTVT = EVT(MVT::getVectorVT(MVT::i32,
- RHSVT.getVectorNumElements()));
- } else {
- LONGVT = EVT(MVT::i64);
- INTVT = EVT(MVT::i32);
- }
- SDValue x = RHS;
- const AMDILSubtarget &STM = getTargetMachine().getSubtarget<AMDILSubtarget>();
- if (STM.calVersion() >= CAL_VERSION_SC_135) {
- // unsigned x = RHS;
- // ulong xd = (ulong)(0x4330_0000 << 32) | x;
- // double d = as_double( xd );
- // return d - 0x1.0p+52; // 0x1.0p+52 == 0x4330_0000_0000_0000
- SDValue xd = DAG.getNode( (isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT, x,
- DAG.getConstant( 0x43300000, INTVT ) );
- SDValue d = DAG.getNode( ISDBITCAST, DL, LHSVT, xd );
- SDValue offsetd = DAG.getNode( ISDBITCAST, DL, LHSVT,
- DAG.getConstant( 0x4330000000000000ULL, LONGVT ) );
- return DAG.getNode( ISD::FSUB, DL, LHSVT, d, offsetd );
- } else {
- SDValue clz = genCLZu32(x, DAG);
-
- // Compute the exponent. 1023 is the bias, 31-clz the actual power of 2
- // Except for an input 0... which requires a 0 exponent
- SDValue exp = DAG.getNode( ISD::SUB, DL, INTVT,
- DAG.getConstant( (1023+31), INTVT), clz );
- exp = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT, x, exp, x );
-
- // Normalize frac
- SDValue rhi = DAG.getNode( ISD::SHL, DL, INTVT, x, clz );
-
- // Eliminate hidden bit
- rhi = DAG.getNode( ISD::AND, DL, INTVT,
- rhi, DAG.getConstant( 0x7fffffff, INTVT ) );
-
- // Pack exponent and frac
- SDValue rlo = DAG.getNode( ISD::SHL, DL, INTVT,
- rhi, DAG.getConstant( (32 - 11), INTVT ) );
- rhi = DAG.getNode( ISD::SRL, DL, INTVT,
- rhi, DAG.getConstant( 11, INTVT ) );
- exp = DAG.getNode( ISD::SHL, DL, INTVT,
- exp, DAG.getConstant( 20, INTVT ) );
- rhi = DAG.getNode( ISD::OR, DL, INTVT, rhi, exp );
-
- // Convert 2 x 32 in to 1 x 64, then to double precision float type
- SDValue res = DAG.getNode( (isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT, rlo, rhi );
- return DAG.getNode(ISDBITCAST, DL, LHSVT, res);
- }
-}
-SDValue
-AMDILTargetLowering::genu64tof64(SDValue RHS, EVT LHSVT,
- SelectionDAG &DAG) const
-{
- EVT RHSVT = RHS.getValueType();
- DebugLoc DL = RHS.getDebugLoc();
- EVT INTVT;
- EVT LONGVT;
- bool isVec = RHSVT.isVector();
- if (isVec) {
- INTVT = EVT(MVT::getVectorVT(MVT::i32,
- RHSVT.getVectorNumElements()));
- } else {
- INTVT = EVT(MVT::i32);
- }
- LONGVT = RHSVT;
- SDValue x = RHS;
- const AMDILSubtarget &STM = getTargetMachine().getSubtarget<AMDILSubtarget>();
- if (STM.device()->getGeneration() > AMDILDeviceInfo::HD6XXX) {
- // double dhi = (double)(as_uint2(x).y);
- // double dlo = (double)(as_uint2(x).x);
- // return mad(dhi, 0x1.0p+32, dlo)
- SDValue dhi = DAG.getNode((isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI, DL, INTVT, x);
- dhi = DAG.getNode(ISD::UINT_TO_FP, DL, LHSVT, dhi);
- SDValue dlo = DAG.getNode((isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTVT, x);
- dlo = DAG.getNode(ISD::UINT_TO_FP, DL, LHSVT, dlo);
- return DAG.getNode(AMDILISD::MAD, DL, LHSVT, dhi,
- DAG.getConstantFP(0x4f800000, LHSVT), dlo);
- } else if (STM.calVersion() >= CAL_VERSION_SC_135) {
- // double lo = as_double( as_ulong( 0x1.0p+52) | (u & 0xffff_ffffUL));
- // double hi = as_double( as_ulong( 0x1.0p+84) | (u >> 32));
- // return (hi - (0x1.0p+84 + 0x1.0p+52)) + lo;
- SDValue xlo = DAG.getNode( (isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTVT, x ); // x & 0xffff_ffffUL
- SDValue xd = DAG.getNode( (isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT, xlo, DAG.getConstant( 0x43300000, INTVT ) );
- SDValue lo = DAG.getNode( ISDBITCAST, DL, LHSVT, xd );
- SDValue xhi = DAG.getNode((isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI, DL, INTVT, x ); // x >> 32
- SDValue xe = DAG.getNode( (isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT, xhi, DAG.getConstant( 0x45300000, INTVT ) );
- SDValue hi = DAG.getNode( ISDBITCAST, DL, LHSVT, xe );
- SDValue c = DAG.getNode( ISDBITCAST, DL, LHSVT,
- DAG.getConstant( 0x4530000000100000ULL, LONGVT ) );
- hi = DAG.getNode( ISD::FSUB, DL, LHSVT, hi, c );
- return DAG.getNode( ISD::FADD, DL, LHSVT, hi, lo );
-
- } else {
- SDValue clz = genCLZu64(x, DAG);
- SDValue xhi = DAG.getNode( (isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI, DL, INTVT, x );
- SDValue xlo = DAG.getNode( (isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTVT, x );
-
- // Compute the exponent. 1023 is the bias, 63-clz the actual power of 2
- SDValue exp = DAG.getNode( ISD::SUB, DL, INTVT,
- DAG.getConstant( (1023+63), INTVT), clz );
- SDValue mash = DAG.getNode( ISD::OR, DL, INTVT, xhi, xlo );
- exp = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT,
- mash, exp, mash ); // exp = exp, or 0 if input was 0
-
- // Normalize frac
- SDValue clz31 = DAG.getNode( ISD::AND, DL, INTVT,
- clz, DAG.getConstant( 31, INTVT ) );
- SDValue rshift = DAG.getNode( ISD::SUB, DL, INTVT,
- DAG.getConstant( 32, INTVT ), clz31 );
- SDValue t1 = DAG.getNode( ISD::SHL, DL, INTVT, xhi, clz31 );
- SDValue t2 = DAG.getNode( ISD::SRL, DL, INTVT, xlo, rshift );
- t2 = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT, clz31, t2, t1 );
- SDValue rhi1 = DAG.getNode( ISD::OR, DL, INTVT, t1, t2 );
- SDValue rlo1 = DAG.getNode( ISD::SHL, DL, INTVT, xlo, clz31 );
- SDValue rhi2 = DAG.getNode( ISD::SHL, DL, INTVT, xlo, clz31 );
- SDValue rlo2 = DAG.getConstant( 0, INTVT );
- SDValue clz32 = DAG.getNode( ISD::AND, DL, INTVT,
- clz, DAG.getConstant( 32, INTVT ) );
- SDValue rhi = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT,
- clz32, rhi2, rhi1 );
- SDValue rlo = DAG.getNode( AMDILISD::CMOVLOG, DL, INTVT,
- clz32, rlo2, rlo1 );
-
- // Eliminate hidden bit
- rhi = DAG.getNode( ISD::AND, DL, INTVT,
- rhi, DAG.getConstant( 0x7fffffff, INTVT ) );
-
- // Save bits needed to round properly
- SDValue round = DAG.getNode( ISD::AND, DL, INTVT,
- rlo, DAG.getConstant( 0x7ff, INTVT ) );
-
- // Pack exponent and frac
- rlo = DAG.getNode( ISD::SRL, DL, INTVT,
- rlo, DAG.getConstant( 11, INTVT ) );
- SDValue temp = DAG.getNode( ISD::SHL, DL, INTVT,
- rhi, DAG.getConstant( (32 - 11), INTVT ) );
- rlo = DAG.getNode( ISD::OR, DL, INTVT, rlo, temp );
- rhi = DAG.getNode( ISD::SRL, DL, INTVT,
- rhi, DAG.getConstant( 11, INTVT ) );
- exp = DAG.getNode( ISD::SHL, DL, INTVT,
- exp, DAG.getConstant( 20, INTVT ) );
- rhi = DAG.getNode( ISD::OR, DL, INTVT, rhi, exp );
-
- // Compute rounding bit
- SDValue even = DAG.getNode( ISD::AND, DL, INTVT,
- rlo, DAG.getConstant( 1, INTVT ) );
- SDValue grs = DAG.getNode( ISD::AND, DL, INTVT,
- round, DAG.getConstant( 0x3ff, INTVT ) );
- grs = DAG.getNode( AMDILISD::CMP, DL, INTVT,
- DAG.getConstant( CondCCodeToCC( ISD::SETNE, MVT::i32), MVT::i32),
- grs, DAG.getConstant( 0, INTVT ) ); // -1 if any GRS set, 0 if none
- grs = DAG.getNode( ISD::OR, DL, INTVT, grs, even );
- round = DAG.getNode( ISD::SRL, DL, INTVT,
- round, DAG.getConstant( 10, INTVT ) );
- round = DAG.getNode( ISD::AND, DL, INTVT, round, grs ); // 0 or 1
-
- // Add rounding bit
- SDValue lround = DAG.getNode( (isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT,
- round, DAG.getConstant( 0, INTVT ) );
- SDValue res = DAG.getNode( (isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, LONGVT, rlo, rhi );
- res = DAG.getNode( ISD::ADD, DL, LONGVT, res, lround );
- return DAG.getNode(ISDBITCAST, DL, LHSVT, res);
- }
-}
-SDValue
-AMDILTargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const
-{
- SDValue RHS = Op.getOperand(0);
- EVT RHSVT = RHS.getValueType();
- MVT RST = RHSVT.getScalarType().getSimpleVT();
- EVT LHSVT = Op.getValueType();
- MVT LST = LHSVT.getScalarType().getSimpleVT();
- DebugLoc DL = Op.getDebugLoc();
- SDValue DST;
- EVT INTVT;
- EVT LONGVT;
- const AMDILSubtarget &STM = getTargetMachine().getSubtarget<AMDILSubtarget>();
- if (LST == MVT::f64 && LHSVT.isVector()
- && STM.device()->getGeneration() > AMDILDeviceInfo::HD6XXX) {
- // We dont support vector 64bit floating point convertions.
- DST = Op;
- for (unsigned x = 0, y = LHSVT.getVectorNumElements(); x < y; ++x) {
- SDValue op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
- DL, RST, RHS, DAG.getTargetConstant(x, MVT::i32));
- op = DAG.getNode(ISD::UINT_TO_FP, DL, LST, op);
- if (!x) {
- DST = DAG.getNode(AMDILISD::VBUILD, DL, LHSVT, op);
- } else {
- DST = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, LHSVT, DST,
- op, DAG.getTargetConstant(x, MVT::i32));
- }
-
- }
- } else {
-
- if (RST == MVT::i32
- && LST == MVT::f64) {
- if (STM.device()->getGeneration() > AMDILDeviceInfo::HD6XXX) {
- DST = SDValue(Op.getNode(), 0);
- } else {
- DST = genu32tof64(RHS, LHSVT, DAG);
- }
- } else if (RST == MVT::i64
- && LST == MVT::f64) {
- DST = genu64tof64(RHS, LHSVT, DAG);
- } else {
- DST = SDValue(Op.getNode(), 0);
- }
- }
- return DST;
-}
-
-SDValue
-AMDILTargetLowering::LowerSUB(SDValue Op, SelectionDAG &DAG) const
-{
- SDValue LHS = Op.getOperand(0);
- SDValue RHS = Op.getOperand(1);
- DebugLoc DL = Op.getDebugLoc();
- EVT OVT = Op.getValueType();
- SDValue DST;
- bool isVec = RHS.getValueType().isVector();
- if (OVT.getScalarType() == MVT::i64) {
- MVT INTTY = MVT::i32;
- if (OVT == MVT::v2i64) {
- INTTY = MVT::v2i32;
- }
- SDValue LHSLO, LHSHI, RHSLO, RHSHI, INTLO, INTHI;
- // TODO: need to turn this into a bitcast of i64/v2i64 to v2i32/v4i32
- LHSLO = DAG.getNode((isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTTY, LHS);
- RHSLO = DAG.getNode((isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO, DL, INTTY, RHS);
- LHSHI = DAG.getNode((isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI, DL, INTTY, LHS);
- RHSHI = DAG.getNode((isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI, DL, INTTY, RHS);
- INTLO = DAG.getNode(ISD::SUB, DL, INTTY, LHSLO, RHSLO);
- INTHI = DAG.getNode(ISD::SUB, DL, INTTY, LHSHI, RHSHI);
- //TODO: need to use IBORROW on HD5XXX and later hardware
- SDValue cmp;
- if (OVT == MVT::i64) {
- cmp = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETULT, MVT::i32), MVT::i32),
- LHSLO, RHSLO);
- } else {
- SDValue cmplo;
- SDValue cmphi;
- SDValue LHSRLO = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
- DL, MVT::i32, LHSLO, DAG.getTargetConstant(0, MVT::i32));
- SDValue LHSRHI = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
- DL, MVT::i32, LHSLO, DAG.getTargetConstant(1, MVT::i32));
- SDValue RHSRLO = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
- DL, MVT::i32, RHSLO, DAG.getTargetConstant(0, MVT::i32));
- SDValue RHSRHI = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
- DL, MVT::i32, RHSLO, DAG.getTargetConstant(1, MVT::i32));
- cmplo = DAG.getNode(AMDILISD::CMP, DL, MVT::i32,
- DAG.getConstant(CondCCodeToCC(ISD::SETULT, MVT::i32), MVT::i32),
- LHSRLO, RHSRLO);
- cmphi = DAG.getNode(AMDILISD::CMP, DL, MVT::i32,
- DAG.getConstant(CondCCodeToCC(ISD::SETULT, MVT::i32), MVT::i32),
- LHSRHI, RHSRHI);
- cmp = DAG.getNode(AMDILISD::VBUILD, DL, MVT::v2i32, cmplo);
- cmp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, MVT::v2i32,
- cmp, cmphi, DAG.getTargetConstant(1, MVT::i32));
- }
- INTHI = DAG.getNode(ISD::ADD, DL, INTTY, INTHI, cmp);
- DST = DAG.getNode((isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE, DL, OVT,
- INTLO, INTHI);
- } else {
- DST = SDValue(Op.getNode(), 0);
- }
- return DST;
-}
-SDValue
-AMDILTargetLowering::LowerFDIV(SDValue Op, SelectionDAG &DAG) const
-{
- EVT OVT = Op.getValueType();
- SDValue DST;
- if (OVT.getScalarType() == MVT::f64) {
- DST = LowerFDIV64(Op, DAG);
- } else if (OVT.getScalarType() == MVT::f32) {
- DST = LowerFDIV32(Op, DAG);
- } else {
- DST = SDValue(Op.getNode(), 0);
- }
- return DST;
-}
-
SDValue
AMDILTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const
{
return DST;
}
-SDValue
-AMDILTargetLowering::LowerUREM(SDValue Op, SelectionDAG &DAG) const
-{
- EVT OVT = Op.getValueType();
- SDValue DST;
- if (OVT.getScalarType() == MVT::i64) {
- DST = LowerUREM64(Op, DAG);
- } else if (OVT.getScalarType() == MVT::i32) {
- DST = LowerUREM32(Op, DAG);
- } else if (OVT.getScalarType() == MVT::i16) {
- DST = LowerUREM16(Op, DAG);
- } else if (OVT.getScalarType() == MVT::i8) {
- DST = LowerUREM8(Op, DAG);
- } else {
- DST = SDValue(Op.getNode(), 0);
- }
- return DST;
-}
-
-SDValue
-AMDILTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const
-{
- DebugLoc DL = Op.getDebugLoc();
- EVT OVT = Op.getValueType();
- SDValue DST;
- bool isVec = OVT.isVector();
- if (OVT.getScalarType() != MVT::i64)
- {
- DST = SDValue(Op.getNode(), 0);
- } else {
- assert(OVT.getScalarType() == MVT::i64 && "Only 64 bit mul should be lowered!");
- // TODO: This needs to be turned into a tablegen pattern
- SDValue LHS = Op.getOperand(0);
- SDValue RHS = Op.getOperand(1);
-
- MVT INTTY = MVT::i32;
- if (OVT == MVT::v2i64) {
- INTTY = MVT::v2i32;
- }
- // mul64(h1, l1, h0, l0)
- SDValue LHSLO = DAG.getNode((isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO,
- DL,
- INTTY, LHS);
- SDValue LHSHI = DAG.getNode((isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI,
- DL,
- INTTY, LHS);
- SDValue RHSLO = DAG.getNode((isVec) ? AMDILISD::LCOMPLO2 : AMDILISD::LCOMPLO,
- DL,
- INTTY, RHS);
- SDValue RHSHI = DAG.getNode((isVec) ? AMDILISD::LCOMPHI2 : AMDILISD::LCOMPHI,
- DL,
- INTTY, RHS);
- // MULLO_UINT_1 r1, h0, l1
- SDValue RHILLO = DAG.getNode(AMDILISD::UMUL,
- DL,
- INTTY, RHSHI, LHSLO);
- // MULLO_UINT_1 r2, h1, l0
- SDValue RLOHHI = DAG.getNode(AMDILISD::UMUL,
- DL,
- INTTY, RHSLO, LHSHI);
- // ADD_INT hr, r1, r2
- SDValue ADDHI = DAG.getNode(ISD::ADD,
- DL,
- INTTY, RHILLO, RLOHHI);
- // MULHI_UINT_1 r3, l1, l0
- SDValue RLOLLO = DAG.getNode(ISD::MULHU,
- DL,
- INTTY, RHSLO, LHSLO);
- // ADD_INT hr, hr, r3
- SDValue HIGH = DAG.getNode(ISD::ADD,
- DL,
- INTTY, ADDHI, RLOLLO);
- // MULLO_UINT_1 l3, l1, l0
- SDValue LOW = DAG.getNode(AMDILISD::UMUL,
- DL,
- INTTY, LHSLO, RHSLO);
- DST = DAG.getNode((isVec) ? AMDILISD::LCREATE2 : AMDILISD::LCREATE,
- DL,
- OVT, LOW, HIGH);
- }
- return DST;
-}
SDValue
AMDILTargetLowering::LowerBUILD_VECTOR( SDValue Op, SelectionDAG &DAG ) const
{
return Nodes1;
}
-SDValue
-AMDILTargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
- SelectionDAG &DAG) const
-{
- DebugLoc DL = Op.getDebugLoc();
- EVT VT = Op.getValueType();
- const SDValue *ptr = NULL;
- const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(Op.getOperand(2));
- uint32_t swizzleNum = 0;
- SDValue DST;
- if (!VT.isVector()) {
- SDValue Res = Op.getOperand(0);
- return Res;
- }
-
- if (Op.getOperand(1).getOpcode() != ISD::UNDEF) {
- ptr = &Op.getOperand(1);
- } else {
- ptr = &Op.getOperand(0);
- }
- if (CSDN) {
- swizzleNum = (uint32_t)CSDN->getZExtValue();
- uint32_t mask2 = 0x04030201 & ~(0xFF << (swizzleNum * 8));
- uint32_t mask3 = 0x01010101 & (0xFF << (swizzleNum * 8));
- DST = DAG.getNode(AMDILISD::VINSERT,
- DL,
- VT,
- Op.getOperand(0),
- *ptr,
- DAG.getTargetConstant(mask2, MVT::i32),
- DAG.getTargetConstant(mask3, MVT::i32));
- } else {
- uint32_t mask2 = 0x04030201 & ~(0xFF << (swizzleNum * 8));
- uint32_t mask3 = 0x01010101 & (0xFF << (swizzleNum * 8));
- SDValue res = DAG.getNode(AMDILISD::VINSERT,
- DL, VT, Op.getOperand(0), *ptr,
- DAG.getTargetConstant(mask2, MVT::i32),
- DAG.getTargetConstant(mask3, MVT::i32));
- for (uint32_t x = 1; x < VT.getVectorNumElements(); ++x) {
- mask2 = 0x04030201 & ~(0xFF << (x * 8));
- mask3 = 0x01010101 & (0xFF << (x * 8));
- SDValue t = DAG.getNode(AMDILISD::VINSERT,
- DL, VT, Op.getOperand(0), *ptr,
- DAG.getTargetConstant(mask2, MVT::i32),
- DAG.getTargetConstant(mask3, MVT::i32));
- SDValue c = DAG.getNode(AMDILISD::CMP, DL, ptr->getValueType(),
- DAG.getConstant(AMDILCC::IL_CC_I_EQ, MVT::i32),
- Op.getOperand(2), DAG.getConstant(x, MVT::i32));
- c = DAG.getNode(AMDILISD::VBUILD, DL, Op.getValueType(), c);
- res = DAG.getNode(AMDILISD::CMOVLOG, DL, VT, c, t, res);
- }
- DST = res;
- }
- return DST;
-}
-
-SDValue
-AMDILTargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
- SelectionDAG &DAG) const
-{
- EVT VT = Op.getValueType();
- const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(Op.getOperand(1));
- uint64_t swizzleNum = 0;
- DebugLoc DL = Op.getDebugLoc();
- SDValue Res;
- if (!Op.getOperand(0).getValueType().isVector()) {
- Res = Op.getOperand(0);
- return Res;
- }
- if (CSDN) {
- // Static vector extraction
- swizzleNum = CSDN->getZExtValue() + 1;
- Res = DAG.getNode(AMDILISD::VEXTRACT,
- DL, VT,
- Op.getOperand(0),
- DAG.getTargetConstant(swizzleNum, MVT::i32));
- } else {
- SDValue Op1 = Op.getOperand(1);
- uint32_t vecSize = 4;
- SDValue Op0 = Op.getOperand(0);
- SDValue res = DAG.getNode(AMDILISD::VEXTRACT,
- DL, VT, Op0,
- DAG.getTargetConstant(1, MVT::i32));
- if (Op0.getValueType().isVector()) {
- vecSize = Op0.getValueType().getVectorNumElements();
- }
- for (uint32_t x = 2; x <= vecSize; ++x) {
- SDValue t = DAG.getNode(AMDILISD::VEXTRACT,
- DL, VT, Op0,
- DAG.getTargetConstant(x, MVT::i32));
- SDValue c = DAG.getNode(AMDILISD::CMP,
- DL, Op1.getValueType(),
- DAG.getConstant(AMDILCC::IL_CC_I_EQ, MVT::i32),
- Op1, DAG.getConstant(x, MVT::i32));
- res = DAG.getNode(AMDILISD::CMOVLOG, DL,
- VT, c, t, res);
-
- }
- Res = res;
- }
- return Res;
-}
-
-SDValue
-AMDILTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
- SelectionDAG &DAG) const
-{
- uint32_t vecSize = Op.getValueType().getVectorNumElements();
- SDValue src = Op.getOperand(0);
- const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(Op.getOperand(1));
- uint64_t offset = 0;
- EVT vecType = Op.getValueType().getVectorElementType();
- DebugLoc DL = Op.getDebugLoc();
- SDValue Result;
- if (CSDN) {
- offset = CSDN->getZExtValue();
- Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
- DL,vecType, src, DAG.getConstant(offset, MVT::i32));
- Result = DAG.getNode(AMDILISD::VBUILD, DL,
- Op.getValueType(), Result);
- for (uint32_t x = 1; x < vecSize; ++x) {
- SDValue elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, vecType,
- src, DAG.getConstant(offset + x, MVT::i32));
- if (elt.getOpcode() != ISD::UNDEF) {
- Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL,
- Op.getValueType(), Result, elt,
- DAG.getConstant(x, MVT::i32));
- }
- }
- } else {
- SDValue idx = Op.getOperand(1);
- Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
- DL, vecType, src, idx);
- Result = DAG.getNode(AMDILISD::VBUILD, DL,
- Op.getValueType(), Result);
- for (uint32_t x = 1; x < vecSize; ++x) {
- idx = DAG.getNode(ISD::ADD, DL, vecType,
- idx, DAG.getConstant(1, MVT::i32));
- SDValue elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, vecType,
- src, idx);
- if (elt.getOpcode() != ISD::UNDEF) {
- Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL,
- Op.getValueType(), Result, elt, idx);
- }
- }
- }
- return Result;
-}
-SDValue
-AMDILTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op,
- SelectionDAG &DAG) const
-{
- SDValue Res = DAG.getNode(AMDILISD::VBUILD,
- Op.getDebugLoc(),
- Op.getValueType(),
- Op.getOperand(0));
- return Res;
-}
SDValue
AMDILTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const
{
return Result;
}
-SDValue
-AMDILTargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const
-{
- SDValue Result = DAG.getNode(
- AMDILISD::DP_TO_FP,
- Op.getDebugLoc(),
- Op.getValueType(),
- Op.getOperand(0),
- Op.getOperand(1));
- return Result;
-}
-
-SDValue
-AMDILTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const
-{
- SDValue Result = DAG.getNode(
- AMDILISD::VCONCAT,
- Op.getDebugLoc(),
- Op.getValueType(),
- Op.getOperand(0),
- Op.getOperand(1));
- return Result;
-}
// LowerRET - Lower an ISD::RET node.
SDValue
AMDILTargetLowering::LowerReturn(SDValue Chain,
{
return SDValue(Op.getNode(), 0);
}
-
-SDValue
-AMDILTargetLowering::LowerUREM8(SDValue Op, SelectionDAG &DAG) const
-{
- DebugLoc DL = Op.getDebugLoc();
- EVT OVT = Op.getValueType();
- MVT INTTY = MVT::i32;
- if (OVT == MVT::v2i8) {
- INTTY = MVT::v2i32;
- } else if (OVT == MVT::v4i8) {
- INTTY = MVT::v4i32;
- }
- SDValue LHS = Op.getOperand(0);
- SDValue RHS = Op.getOperand(1);
- // The LowerUREM8 function generates equivalent to the following IL.
- // mov r0, as_u32(LHS)
- // mov r1, as_u32(RHS)
- // and r10, r0, 0xFF
- // and r11, r1, 0xFF
- // cmov_logical r3, r11, r11, 0x1
- // udiv r3, r10, r3
- // cmov_logical r3, r11, r3, 0
- // umul r3, r3, r11
- // sub r3, r10, r3
- // and as_u8(DST), r3, 0xFF
-
- // mov r0, as_u32(LHS)
- SDValue r0 = DAG.getSExtOrTrunc(LHS, DL, INTTY);
-
- // mov r1, as_u32(RHS)
- SDValue r1 = DAG.getSExtOrTrunc(RHS, DL, INTTY);
-
- // and r10, r0, 0xFF
- SDValue r10 = DAG.getNode(ISD::AND, DL, INTTY, r0,
- DAG.getConstant(0xFF, INTTY));
-
- // and r11, r1, 0xFF
- SDValue r11 = DAG.getNode(ISD::AND, DL, INTTY, r1,
- DAG.getConstant(0xFF, INTTY));
-
- // cmov_logical r3, r11, r11, 0x1
- SDValue r3 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, r11, r11,
- DAG.getConstant(0x01, INTTY));
-
- // udiv r3, r10, r3
- r3 = DAG.getNode(ISD::UREM, DL, INTTY, r10, r3);
-
- // cmov_logical r3, r11, r3, 0
- r3 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, r11, r3,
- DAG.getConstant(0, INTTY));
-
- // umul r3, r3, r11
- r3 = DAG.getNode(AMDILISD::UMUL, DL, INTTY, r3, r11);
-
- // sub r3, r10, r3
- r3 = DAG.getNode(ISD::SUB, DL, INTTY, r10, r3);
-
- // and as_u8(DST), r3, 0xFF
- SDValue DST = DAG.getNode(ISD::AND, DL, INTTY, r3,
- DAG.getConstant(0xFF, INTTY));
- DST = DAG.getZExtOrTrunc(DST, DL, OVT);
- return DST;
-}
-
-SDValue
-AMDILTargetLowering::LowerUREM16(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 = Op.getOperand(0);
- SDValue RHS = Op.getOperand(1);
- // The LowerUREM16 function generatest equivalent to the following IL.
- // mov r0, LHS
- // mov r1, RHS
- // DIV = LowerUDIV16(LHS, RHS)
- // and r10, r0, 0xFFFF
- // and r11, r1, 0xFFFF
- // cmov_logical r3, r11, r11, 0x1
- // udiv as_u16(r3), as_u32(r10), as_u32(r3)
- // and r3, r3, 0xFFFF
- // cmov_logical r3, r11, r3, 0
- // umul r3, r3, r11
- // sub r3, r10, r3
- // and DST, r3, 0xFFFF
-
- // mov r0, LHS
- SDValue r0 = LHS;
-
- // mov r1, RHS
- SDValue r1 = RHS;
-
- // and r10, r0, 0xFFFF
- SDValue r10 = DAG.getNode(ISD::AND, DL, OVT, r0,
- DAG.getConstant(0xFFFF, OVT));
-
- // and r11, r1, 0xFFFF
- SDValue r11 = DAG.getNode(ISD::AND, DL, OVT, r1,
- DAG.getConstant(0xFFFF, OVT));
-
- // cmov_logical r3, r11, r11, 0x1
- SDValue r3 = DAG.getNode(AMDILISD::CMOVLOG, DL, OVT, r11, r11,
- DAG.getConstant(0x01, OVT));
-
- // udiv as_u16(r3), as_u32(r10), as_u32(r3)
- r10 = DAG.getZExtOrTrunc(r10, DL, INTTY);
- r3 = DAG.getZExtOrTrunc(r3, DL, INTTY);
- r3 = DAG.getNode(ISD::UREM, DL, INTTY, r10, r3);
- r3 = DAG.getZExtOrTrunc(r3, DL, OVT);
- r10 = DAG.getZExtOrTrunc(r10, DL, OVT);
-
- // and r3, r3, 0xFFFF
- r3 = DAG.getNode(ISD::AND, DL, OVT, r3,
- DAG.getConstant(0xFFFF, OVT));
-
- // cmov_logical r3, r11, r3, 0
- r3 = DAG.getNode(AMDILISD::CMOVLOG, DL, OVT, r11, r3,
- DAG.getConstant(0, OVT));
- // umul r3, r3, r11
- r3 = DAG.getNode(AMDILISD::UMUL, DL, OVT, r3, r11);
-
- // sub r3, r10, r3
- r3 = DAG.getNode(ISD::SUB, DL, OVT, r10, r3);
-
- // and DST, r3, 0xFFFF
- SDValue DST = DAG.getNode(ISD::AND, DL, OVT, r3,
- DAG.getConstant(0xFFFF, OVT));
- return DST;
-}
-
-SDValue
-AMDILTargetLowering::LowerUREM32(SDValue Op, SelectionDAG &DAG) const
-{
- DebugLoc DL = Op.getDebugLoc();
- EVT OVT = Op.getValueType();
- SDValue LHS = Op.getOperand(0);
- SDValue RHS = Op.getOperand(1);
- // The LowerUREM32 function generates equivalent to the following IL.
- // udiv r20, LHS, RHS
- // umul r20, r20, RHS
- // sub DST, LHS, r20
-
- // udiv r20, LHS, RHS
- SDValue r20 = DAG.getNode(ISD::UDIV, DL, OVT, LHS, RHS);
-
- // umul r20, r20, RHS
- r20 = DAG.getNode(AMDILISD::UMUL, DL, OVT, r20, RHS);
-
- // sub DST, LHS, r20
- SDValue DST = DAG.getNode(ISD::SUB, DL, OVT, LHS, r20);
- return DST;
-}
-
-SDValue
-AMDILTargetLowering::LowerUREM64(SDValue Op, SelectionDAG &DAG) const
-{
- return SDValue(Op.getNode(), 0);
-}
-
-
-SDValue
-AMDILTargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const
-{
- DebugLoc DL = Op.getDebugLoc();
- EVT OVT = Op.getValueType();
- MVT INTTY = MVT::i32;
- if (OVT == MVT::v2f32) {
- INTTY = MVT::v2i32;
- } else if (OVT == MVT::v4f32) {
- INTTY = MVT::v4i32;
- }
- SDValue LHS = Op.getOperand(0);
- SDValue RHS = Op.getOperand(1);
- SDValue DST;
- const AMDILSubtarget &STM = getTargetMachine().getSubtarget<AMDILSubtarget>();
- if (STM.device()->getGeneration() == AMDILDeviceInfo::HD4XXX) {
- // TODO: This doesn't work for vector types yet
- // The LowerFDIV32 function generates equivalent to the following
- // IL:
- // mov r20, as_int(LHS)
- // mov r21, as_int(RHS)
- // and r30, r20, 0x7f800000
- // and r31, r20, 0x807FFFFF
- // and r32, r21, 0x7f800000
- // and r33, r21, 0x807FFFFF
- // ieq r40, r30, 0x7F800000
- // ieq r41, r31, 0x7F800000
- // ieq r42, r32, 0
- // ieq r43, r33, 0
- // and r50, r20, 0x80000000
- // and r51, r21, 0x80000000
- // ior r32, r32, 0x3f800000
- // ior r33, r33, 0x3f800000
- // cmov_logical r32, r42, r50, r32
- // cmov_logical r33, r43, r51, r33
- // cmov_logical r32, r40, r20, r32
- // cmov_logical r33, r41, r21, r33
- // ior r50, r40, r41
- // ior r51, r42, r43
- // ior r50, r50, r51
- // inegate r52, r31
- // iadd r30, r30, r52
- // cmov_logical r30, r50, 0, r30
- // div_zeroop(infinity) r21, 1.0, r33
- // mul_ieee r20, r32, r21
- // and r22, r20, 0x7FFFFFFF
- // and r23, r20, 0x80000000
- // ishr r60, r22, 0x00000017
- // ishr r61, r30, 0x00000017
- // iadd r20, r20, r30
- // iadd r21, r22, r30
- // iadd r60, r60, r61
- // ige r42, 0, R60
- // ior r41, r23, 0x7F800000
- // ige r40, r60, 0x000000FF
- // cmov_logical r40, r50, 0, r40
- // cmov_logical r20, r42, r23, r20
- // cmov_logical DST, r40, r41, r20
- // as_float(DST)
-
- // mov r20, as_int(LHS)
- SDValue R20 = DAG.getNode(ISDBITCAST, DL, INTTY, LHS);
-
- // mov r21, as_int(RHS)
- SDValue R21 = DAG.getNode(ISDBITCAST, DL, INTTY, RHS);
-
- // and r30, r20, 0x7f800000
- SDValue R30 = DAG.getNode(ISD::AND, DL, INTTY, R20,
- DAG.getConstant(0x7F800000, INTTY));
-
- // and r31, r21, 0x7f800000
- SDValue R31 = DAG.getNode(ISD::AND, DL, INTTY, R21,
- DAG.getConstant(0x7f800000, INTTY));
-
- // and r32, r20, 0x807FFFFF
- SDValue R32 = DAG.getNode(ISD::AND, DL, INTTY, R20,
- DAG.getConstant(0x807FFFFF, INTTY));
-
- // and r33, r21, 0x807FFFFF
- SDValue R33 = DAG.getNode(ISD::AND, DL, INTTY, R21,
- DAG.getConstant(0x807FFFFF, INTTY));
-
- // ieq r40, r30, 0x7F800000
- SDValue R40 = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::i32), MVT::i32),
- R30, DAG.getConstant(0x7F800000, INTTY));
-
- // ieq r41, r31, 0x7F800000
- SDValue R41 = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::i32), MVT::i32),
- R31, DAG.getConstant(0x7F800000, INTTY));
-
- // ieq r42, r30, 0
- SDValue R42 = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::i32), MVT::i32),
- R30, DAG.getConstant(0, INTTY));
-
- // ieq r43, r31, 0
- SDValue R43 = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETEQ, MVT::i32), MVT::i32),
- R31, DAG.getConstant(0, INTTY));
-
- // and r50, r20, 0x80000000
- SDValue R50 = DAG.getNode(ISD::AND, DL, INTTY, R20,
- DAG.getConstant(0x80000000, INTTY));
-
- // and r51, r21, 0x80000000
- SDValue R51 = DAG.getNode(ISD::AND, DL, INTTY, R21,
- DAG.getConstant(0x80000000, INTTY));
-
- // ior r32, r32, 0x3f800000
- R32 = DAG.getNode(ISD::OR, DL, INTTY, R32,
- DAG.getConstant(0x3F800000, INTTY));
-
- // ior r33, r33, 0x3f800000
- R33 = DAG.getNode(ISD::OR, DL, INTTY, R33,
- DAG.getConstant(0x3F800000, INTTY));
-
- // cmov_logical r32, r42, r50, r32
- R32 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, R42, R50, R32);
-
- // cmov_logical r33, r43, r51, r33
- R33 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, R43, R51, R33);
-
- // cmov_logical r32, r40, r20, r32
- R32 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, R40, R20, R32);
-
- // cmov_logical r33, r41, r21, r33
- R33 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, R41, R21, R33);
-
- // ior r50, r40, r41
- R50 = DAG.getNode(ISD::OR, DL, INTTY, R40, R41);
-
- // ior r51, r42, r43
- R51 = DAG.getNode(ISD::OR, DL, INTTY, R42, R43);
-
- // ior r50, r50, r51
- R50 = DAG.getNode(ISD::OR, DL, INTTY, R50, R51);
-
- // inegate r52, r31
- SDValue R52 = DAG.getNode(AMDILISD::INEGATE, DL, INTTY, R31);
-
- // iadd r30, r30, r52
- R30 = DAG.getNode(ISD::ADD, DL, INTTY, R30, R52);
-
- // cmov_logical r30, r50, 0, r30
- R30 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, R50,
- DAG.getConstant(0, INTTY), R30);
-
- // div_zeroop(infinity) r21, 1.0, as_float(r33)
- R33 = DAG.getNode(ISDBITCAST, DL, OVT, R33);
- R21 = DAG.getNode(AMDILISD::DIV_INF, DL, OVT,
- DAG.getConstantFP(1.0f, OVT), R33);
-
- // mul_ieee as_int(r20), as_float(r32), r21
- R32 = DAG.getNode(ISDBITCAST, DL, OVT, R32);
- R20 = DAG.getNode(ISD::FMUL, DL, OVT, R32, R21);
- R20 = DAG.getNode(ISDBITCAST, DL, INTTY, R20);
-
- // div_zeroop(infinity) r21, 1.0, as_float(r33)
- R33 = DAG.getNode(ISDBITCAST, DL, OVT, R33);
- R21 = DAG.getNode(AMDILISD::DIV_INF, DL, OVT,
- DAG.getConstantFP(1.0f, OVT), R33);
-
- // mul_ieee as_int(r20), as_float(r32), r21
- R32 = DAG.getNode(ISDBITCAST, DL, OVT, R32);
- R20 = DAG.getNode(ISD::FMUL, DL, OVT, R32, R21);
- R20 = DAG.getNode(ISDBITCAST, DL, INTTY, R20);
-
- // and r22, r20, 0x7FFFFFFF
- SDValue R22 = DAG.getNode(ISD::AND, DL, INTTY, R20,
- DAG.getConstant(0x7FFFFFFF, INTTY));
-
- // and r23, r20, 0x80000000
- SDValue R23 = DAG.getNode(ISD::AND, DL, INTTY, R20,
- DAG.getConstant(0x80000000, INTTY));
-
- // ishr r60, r22, 0x00000017
- SDValue R60 = DAG.getNode(ISD::SRA, DL, INTTY, R22,
- DAG.getConstant(0x00000017, INTTY));
-
- // ishr r61, r30, 0x00000017
- SDValue R61 = DAG.getNode(ISD::SRA, DL, INTTY, R30,
- DAG.getConstant(0x00000017, INTTY));
-
- // iadd r20, r20, r30
- R20 = DAG.getNode(ISD::ADD, DL, INTTY, R20, R30);
-
- // iadd r21, r22, r30
- R21 = DAG.getNode(ISD::ADD, DL, INTTY, R22, R30);
-
- // iadd r60, r60, r61
- R60 = DAG.getNode(ISD::ADD, DL, INTTY, R60, R61);
-
- // ige r42, 0, R60
- R42 = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETGE, MVT::i32), MVT::i32),
- DAG.getConstant(0, INTTY),
- R60);
-
- // ior r41, r23, 0x7F800000
- R41 = DAG.getNode(ISD::OR, DL, INTTY, R23,
- DAG.getConstant(0x7F800000, INTTY));
-
- // ige r40, r60, 0x000000FF
- R40 = DAG.getNode(AMDILISD::CMP, DL, INTTY,
- DAG.getConstant(CondCCodeToCC(ISD::SETGE, MVT::i32), MVT::i32),
- R60,
- DAG.getConstant(0x0000000FF, INTTY));
-
- // cmov_logical r40, r50, 0, r40
- R40 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, R50,
- DAG.getConstant(0, INTTY),
- R40);
-
- // cmov_logical r20, r42, r23, r20
- R20 = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, R42, R23, R20);
-
- // cmov_logical DST, r40, r41, r20
- DST = DAG.getNode(AMDILISD::CMOVLOG, DL, INTTY, R40, R41, R20);
-
- // as_float(DST)
- DST = DAG.getNode(ISDBITCAST, DL, OVT, DST);
- } else {
- // The following sequence of DAG nodes produce the following IL:
- // fabs r1, RHS
- // lt r2, 0x1.0p+96f, r1
- // cmov_logical r3, r2, 0x1.0p-23f, 1.0f
- // mul_ieee r1, RHS, r3
- // div_zeroop(infinity) r0, LHS, r1
- // mul_ieee DST, r0, r3
-
- // fabs r1, RHS
- SDValue r1 = DAG.getNode(ISD::FABS, DL, OVT, RHS);
- // lt r2, 0x1.0p+96f, r1
- SDValue r2 = DAG.getNode(AMDILISD::CMP, DL, OVT,
- DAG.getConstant(CondCCodeToCC(ISD::SETLT, MVT::f32), MVT::i32),
- DAG.getConstant(0x6f800000, INTTY), r1);
- // cmov_logical r3, r2, 0x1.0p-23f, 1.0f
- SDValue r3 = DAG.getNode(AMDILISD::CMOVLOG, DL, OVT, r2,
- DAG.getConstant(0x2f800000, INTTY),
- DAG.getConstant(0x3f800000, INTTY));
- // mul_ieee r1, RHS, r3
- r1 = DAG.getNode(ISD::FMUL, DL, OVT, RHS, r3);
- // div_zeroop(infinity) r0, LHS, r1
- SDValue r0 = DAG.getNode(AMDILISD::DIV_INF, DL, OVT, LHS, r1);
- // mul_ieee DST, r0, r3
- DST = DAG.getNode(ISD::FMUL, DL, OVT, r0, r3);
- }
- return DST;
-}
-
-SDValue
-AMDILTargetLowering::LowerFDIV64(SDValue Op, SelectionDAG &DAG) const
-{
- return SDValue(Op.getNode(), 0);
-}
projects / mesa.git / commitdiff