#include "PPCPredicates.h"
#include "PPCTargetMachine.h"
#include "PPCPerfectShuffle.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/VectorExtras.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/CodeGen/CallingConvLower.h"
setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
+ // Shortening conversions involving ppcf128 get expanded (2 regs -> 1 reg)
+ setConvertAction(MVT::ppcf128, MVT::f64, Expand);
+ setConvertAction(MVT::ppcf128, MVT::f32, Expand);
+ // This is used in the ppcf128->int sequence. Note it has different semantics
+ // from FP_ROUND: that rounds to nearest, this rounds to zero.
+ setOperationAction(ISD::FP_ROUND_INREG, MVT::ppcf128, Custom);
+
// PowerPC has no intrinsics for these particular operations
setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
setOperationAction(ISD::MEMSET, MVT::Other, Expand);
setOperationAction(ISD::UREM, MVT::i32, Expand);
setOperationAction(ISD::SREM, MVT::i64, Expand);
setOperationAction(ISD::UREM, MVT::i64, Expand);
-
- // We don't support sin/cos/sqrt/fmod
+
+ // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM.
+ setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
+ setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
+ setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
+ setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
+ setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
+ setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
+ setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
+ setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
+
+ // We don't support sin/cos/sqrt/fmod/pow
setOperationAction(ISD::FSIN , MVT::f64, Expand);
setOperationAction(ISD::FCOS , MVT::f64, Expand);
setOperationAction(ISD::FREM , MVT::f64, Expand);
+ setOperationAction(ISD::FPOW , MVT::f64, Expand);
setOperationAction(ISD::FSIN , MVT::f32, Expand);
setOperationAction(ISD::FCOS , MVT::f32, Expand);
setOperationAction(ISD::FREM , MVT::f32, Expand);
+ setOperationAction(ISD::FPOW , MVT::f32, Expand);
// If we're enabling GP optimizations, use hardware square root
if (!TM.getSubtarget<PPCSubtarget>().hasFSQRT()) {
// RET must be custom lowered, to meet ABI requirements
setOperationAction(ISD::RET , MVT::Other, Custom);
- setOperationAction(ISD::ADJUST_TRAMP, MVT::i32, Expand);
- setOperationAction(ISD::ADJUST_TRAMP, MVT::i64, Expand);
-
// VASTART needs to be custom lowered to use the VarArgsFrameIndex
setOperationAction(ISD::VASTART , MVT::Other, Custom);
}
if (TM.getSubtarget<PPCSubtarget>().use64BitRegs()) {
- // 64 bit PowerPC implementations can support i64 types directly
+ // 64-bit PowerPC implementations can support i64 types directly
addRegisterClass(MVT::i64, PPC::G8RCRegisterClass);
// BUILD_PAIR can't be handled natively, and should be expanded to shl/or
setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
} else {
- // 32 bit PowerPC wants to expand i64 shifts itself.
+ // 32-bit PowerPC wants to expand i64 shifts itself.
setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
-
+ setOperationAction(ISD::UMUL_LOHI, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::SMUL_LOHI, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::UDIVREM, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::SDIVREM, (MVT::ValueType)VT, Expand);
setOperationAction(ISD::SCALAR_TO_VECTOR, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::FPOW, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::CTPOP, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::CTLZ, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::CTTZ, (MVT::ValueType)VT, Expand);
}
// We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle
setTargetDAGCombine(ISD::BR_CC);
setTargetDAGCombine(ISD::BSWAP);
+ // Darwin long double math library functions have $LDBL128 appended.
+ if (TM.getSubtarget<PPCSubtarget>().isDarwin()) {
+ setLibcallName(RTLIB::SQRT_PPCF128, "sqrtl$LDBL128");
+ setLibcallName(RTLIB::POW_PPCF128, "powl$LDBL128");
+ setLibcallName(RTLIB::REM_PPCF128, "fmodl$LDBL128");
+ }
+
computeRegisterProperties();
}
PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
};
- const unsigned Num_GPR_Regs = sizeof(GPR_32)/sizeof(GPR_32[0]);
+ const unsigned Num_GPR_Regs = array_lengthof(GPR_32);
const unsigned Num_FPR_Regs = isMachoABI ? 13 : 8;
- const unsigned Num_VR_Regs = sizeof( VR)/sizeof( VR[0]);
+ const unsigned Num_VR_Regs = array_lengthof( VR);
unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8,
PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
};
- const unsigned NumGPRs = sizeof(GPR_32)/sizeof(GPR_32[0]);
+ const unsigned NumGPRs = array_lengthof(GPR_32);
const unsigned NumFPRs = isMachoABI ? 13 : 8;
- const unsigned NumVRs = sizeof( VR)/sizeof( VR[0]);
+ const unsigned NumVRs = array_lengthof( VR);
const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
NumResults = 1;
NodeTys.push_back(MVT::i64);
break;
- case MVT::f32:
case MVT::f64:
+ if (Op.Val->getValueType(1) == MVT::f64) {
+ Chain = DAG.getCopyFromReg(Chain, PPC::F1, MVT::f64, InFlag).getValue(1);
+ ResultVals[0] = Chain.getValue(0);
+ Chain = DAG.getCopyFromReg(Chain, PPC::F2, MVT::f64,
+ Chain.getValue(2)).getValue(1);
+ ResultVals[1] = Chain.getValue(0);
+ NumResults = 2;
+ NodeTys.push_back(MVT::f64);
+ NodeTys.push_back(MVT::f64);
+ break;
+ }
+ // else fall through
+ case MVT::f32:
Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
InFlag).getValue(1);
ResultVals[0] = Chain.getValue(0);
DAG.getNode(ISD::FNEG, MVT::f64, LHS), TV, FV);
}
- SDOperand Cmp;
+ SDOperand Cmp;
switch (CC) {
default: break; // SETUO etc aren't handled by fsel.
case ISD::SETULT:
}
// Convert the FP value to an int value through memory.
- SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::i64, Tmp);
+ SDOperand FIPtr = DAG.CreateStackTemporary(MVT::f64);
+
+ // Emit a store to the stack slot.
+ SDOperand Chain = DAG.getStore(DAG.getEntryNode(), Tmp, FIPtr, NULL, 0);
+
+ // Result is a load from the stack slot. If loading 4 bytes, make sure to
+ // add in a bias.
if (Op.getValueType() == MVT::i32)
- Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
- return Bits;
+ FIPtr = DAG.getNode(ISD::ADD, FIPtr.getValueType(), FIPtr,
+ DAG.getConstant(4, FIPtr.getValueType()));
+ return DAG.getLoad(Op.getValueType(), Chain, FIPtr, NULL, 0);
+}
+
+static SDOperand LowerFP_ROUND_INREG(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getValueType() == MVT::ppcf128);
+ SDNode *Node = Op.Val;
+ assert(Node->getOperand(0).getValueType() == MVT::ppcf128);
+ assert(Node->getOperand(0).Val->getOpcode() == ISD::BUILD_PAIR);
+ SDOperand Lo = Node->getOperand(0).Val->getOperand(0);
+ SDOperand Hi = Node->getOperand(0).Val->getOperand(1);
+
+ // This sequence changes FPSCR to do round-to-zero, adds the two halves
+ // of the long double, and puts FPSCR back the way it was. We do not
+ // actually model FPSCR.
+ std::vector<MVT::ValueType> NodeTys;
+ SDOperand Ops[4], Result, MFFSreg, InFlag, FPreg;
+
+ NodeTys.push_back(MVT::f64); // Return register
+ NodeTys.push_back(MVT::Flag); // Returns a flag for later insns
+ Result = DAG.getNode(PPCISD::MFFS, NodeTys, &InFlag, 0);
+ MFFSreg = Result.getValue(0);
+ InFlag = Result.getValue(1);
+
+ NodeTys.clear();
+ NodeTys.push_back(MVT::Flag); // Returns a flag
+ Ops[0] = DAG.getConstant(31, MVT::i32);
+ Ops[1] = InFlag;
+ Result = DAG.getNode(PPCISD::MTFSB1, NodeTys, Ops, 2);
+ InFlag = Result.getValue(0);
+
+ NodeTys.clear();
+ NodeTys.push_back(MVT::Flag); // Returns a flag
+ Ops[0] = DAG.getConstant(30, MVT::i32);
+ Ops[1] = InFlag;
+ Result = DAG.getNode(PPCISD::MTFSB0, NodeTys, Ops, 2);
+ InFlag = Result.getValue(0);
+
+ NodeTys.clear();
+ NodeTys.push_back(MVT::f64); // result of add
+ NodeTys.push_back(MVT::Flag); // Returns a flag
+ Ops[0] = Lo;
+ Ops[1] = Hi;
+ Ops[2] = InFlag;
+ Result = DAG.getNode(PPCISD::FADDRTZ, NodeTys, Ops, 3);
+ FPreg = Result.getValue(0);
+ InFlag = Result.getValue(1);
+
+ NodeTys.clear();
+ NodeTys.push_back(MVT::f64);
+ Ops[0] = DAG.getConstant(1, MVT::i32);
+ Ops[1] = MFFSreg;
+ Ops[2] = FPreg;
+ Ops[3] = InFlag;
+ Result = DAG.getNode(PPCISD::MTFSF, NodeTys, Ops, 4);
+ FPreg = Result.getValue(0);
+
+ // We know the low half is about to be thrown away, so just use something
+ // convenient.
+ return DAG.getNode(ISD::BUILD_PAIR, Lo.getValueType(), FPreg, FPreg);
}
static SDOperand LowerSINT_TO_FP(SDOperand Op, SelectionDAG &DAG) {
-8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
};
- for (unsigned idx = 0; idx < sizeof(SplatCsts)/sizeof(SplatCsts[0]); ++idx){
+ for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) {
// Indirect through the SplatCsts array so that we favor 'vsplti -1' for
// cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
int i = SplatCsts[idx];
if (SextVal >= 0 && SextVal <= 31) {
SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
- LHS = DAG.getNode(ISD::SUB, Op.getValueType(), LHS, RHS);
+ LHS = DAG.getNode(ISD::SUB, LHS.getValueType(), LHS, RHS);
return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
}
// Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
if (SextVal >= -31 && SextVal <= 0) {
SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
- LHS = DAG.getNode(ISD::ADD, Op.getValueType(), LHS, RHS);
+ LHS = DAG.getNode(ISD::ADD, LHS.getValueType(), LHS, RHS);
return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
}
}
case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
+ case ISD::FP_ROUND_INREG: return LowerFP_ROUND_INREG(Op, DAG);
// Lower 64-bit shifts.
case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
projects / oota-llvm.git / blobdiff