diff options
Diffstat (limited to 'llvm/lib/Target/LoongArch')
5 files changed, 64 insertions, 39 deletions
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp index a31fa57..d96136c 100644 --- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp +++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp @@ -2385,23 +2385,6 @@ SDValue LoongArchTargetLowering::lowerBF16_TO_FP(SDValue Op, return Res; } -static bool isConstantOrUndef(const SDValue Op) { - if (Op->isUndef()) - return true; - if (isa<ConstantSDNode>(Op)) - return true; - if (isa<ConstantFPSDNode>(Op)) - return true; - return false; -} - -static bool isConstantOrUndefBUILD_VECTOR(const BuildVectorSDNode *Op) { - for (unsigned i = 0; i < Op->getNumOperands(); ++i) - if (isConstantOrUndef(Op->getOperand(i))) - return true; - return false; -} - // Lower BUILD_VECTOR as broadcast load (if possible). // For example: // %a = load i8, ptr %ptr @@ -2451,10 +2434,14 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const { BuildVectorSDNode *Node = cast<BuildVectorSDNode>(Op); EVT ResTy = Op->getValueType(0); + unsigned NumElts = ResTy.getVectorNumElements(); SDLoc DL(Op); APInt SplatValue, SplatUndef; unsigned SplatBitSize; bool HasAnyUndefs; + bool IsConstant = false; + bool UseSameConstant = true; + SDValue ConstantValue; bool Is128Vec = ResTy.is128BitVector(); bool Is256Vec = ResTy.is256BitVector(); @@ -2505,19 +2492,45 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op, if (DAG.isSplatValue(Op, /*AllowUndefs=*/false)) return Op; - if (!isConstantOrUndefBUILD_VECTOR(Node)) { + for (unsigned i = 0; i < NumElts; ++i) { + SDValue Opi = Node->getOperand(i); + if (isIntOrFPConstant(Opi)) { + IsConstant = true; + if (!ConstantValue.getNode()) + ConstantValue = Opi; + else if (ConstantValue != Opi) + UseSameConstant = false; + } + } + + // If the type of BUILD_VECTOR is v2f64, custom legalizing it has no benefits. + if (IsConstant && UseSameConstant && ResTy != MVT::v2f64) { + SDValue Result = DAG.getSplatBuildVector(ResTy, DL, ConstantValue); + for (unsigned i = 0; i < NumElts; ++i) { + SDValue Opi = Node->getOperand(i); + if (!isIntOrFPConstant(Opi)) + Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Result, Opi, + DAG.getConstant(i, DL, Subtarget.getGRLenVT())); + } + return Result; + } + + if (!IsConstant) { // Use INSERT_VECTOR_ELT operations rather than expand to stores. // The resulting code is the same length as the expansion, but it doesn't // use memory operations. - EVT ResTy = Node->getValueType(0); - assert(ResTy.isVector()); - unsigned NumElts = ResTy.getVectorNumElements(); + SDValue Op0 = Node->getOperand(0); SDValue Vector = DAG.getUNDEF(ResTy); - for (unsigned i = 0; i < NumElts; ++i) { - Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector, - Node->getOperand(i), + + if (!Op0.isUndef()) + Vector = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ResTy, Op0); + for (unsigned i = 1; i < NumElts; ++i) { + SDValue Opi = Node->getOperand(i); + if (Opi.isUndef()) + continue; + Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ResTy, Vector, Opi, DAG.getConstant(i, DL, Subtarget.getGRLenVT())); } return Vector; diff --git a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td index a0107e4..5096a8f 100644 --- a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td +++ b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td @@ -1651,18 +1651,20 @@ def : Pat<(vector_insert v8i32:$xd, GRLenVT:$rj, uimm3:$imm), (XVINSGR2VR_W v8i32:$xd, GRLenVT:$rj, uimm3:$imm)>; def : Pat<(vector_insert v4i64:$xd, GRLenVT:$rj, uimm2:$imm), (XVINSGR2VR_D v4i64:$xd, GRLenVT:$rj, uimm2:$imm)>; -def : Pat<(vector_insert v8f32:$vd, (loongarch_movgr2fr_w_la64 GPR:$rj), uimm3:$imm), - (XVINSGR2VR_W $vd, $rj, uimm3:$imm)>; -def : Pat<(vector_insert v4f64:$vd, (f64 (bitconvert i64:$rj)), uimm2:$imm), - (XVINSGR2VR_D $vd, $rj, uimm2:$imm)>; +def : Pat<(vector_insert v8f32:$xd, (loongarch_movgr2fr_w_la64 GPR:$rj), uimm3:$imm), + (XVINSGR2VR_W $xd, $rj, uimm3:$imm)>; +def : Pat<(vector_insert v4f64:$xd, (f64 (bitconvert i64:$rj)), uimm2:$imm), + (XVINSGR2VR_D $xd, $rj, uimm2:$imm)>; def : Pat<(vector_insert v8f32:$xd, (f32 (vector_extract v8f32:$xj, uimm3:$imm1)), uimm3:$imm2), (XVINSGR2VR_W $xd, (XVPICKVE2GR_W v8f32:$xj, uimm3:$imm1), uimm3:$imm2)>; def : Pat<(vector_insert v4f64:$xd, (f64 (vector_extract v4f64:$xj, uimm2:$imm1)), uimm2:$imm2), (XVINSGR2VR_D $xd, (XVPICKVE2GR_D v4f64:$xj, uimm2:$imm1), uimm2:$imm2)>; + +// XVINSVE0_{W/D} def : Pat<(vector_insert v8f32:$xd, FPR32:$fj, uimm3:$imm), - (XVINSGR2VR_W $xd, (COPY_TO_REGCLASS FPR32:$fj, GPR), uimm3:$imm)>; + (XVINSVE0_W $xd, (SUBREG_TO_REG (i64 0), FPR32:$fj, sub_32), uimm3:$imm)>; def : Pat<(vector_insert v4f64:$xd, FPR64:$fj, uimm2:$imm), - (XVINSGR2VR_D $xd, (COPY_TO_REGCLASS FPR64:$fj, GPR), uimm2:$imm)>; + (XVINSVE0_D $xd, (SUBREG_TO_REG (i64 0), FPR64:$fj, sub_64), uimm2:$imm)>; // scalar_to_vector def : Pat<(v8f32 (scalar_to_vector FPR32:$fj)), diff --git a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td index 962e7c2..3c9defb 100644 --- a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td +++ b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td @@ -1842,10 +1842,19 @@ def : Pat<(vector_insert v4f32:$vd, (loongarch_movgr2fr_w_la64 GPR:$rj), uimm2:$ (VINSGR2VR_W $vd, $rj, uimm2:$imm)>; def : Pat<(vector_insert v2f64:$vd, (f64 (bitconvert i64:$rj)), uimm1:$imm), (VINSGR2VR_D $vd, $rj, uimm1:$imm)>; -def : Pat<(vector_insert v4f32:$vd, FPR32:$fj, uimm2:$imm), - (VINSGR2VR_W $vd, (COPY_TO_REGCLASS FPR32:$fj, GPR), uimm2:$imm)>; -def : Pat<(vector_insert v2f64:$vd, FPR64:$fj, uimm1:$imm), - (VINSGR2VR_D $vd, (COPY_TO_REGCLASS FPR64:$fj, GPR), uimm1:$imm)>; + +// VEXTRINS_{W/D} +foreach imm = 0...3 in { + defvar Imm = !shl(imm, 4); + def : Pat<(vector_insert v4f32:$vd, FPR32:$fj, imm), + (VEXTRINS_W $vd, (SUBREG_TO_REG (i64 0), FPR32:$fj, sub_32), Imm)>; +} + +foreach imm = 0...1 in { + defvar Imm = !shl(imm, 4); + def : Pat<(vector_insert v2f64:$vd, FPR64:$fj, imm), + (VEXTRINS_D $vd, (SUBREG_TO_REG (i64 0), FPR64:$fj, sub_64), Imm)>; +} // scalar_to_vector def : Pat<(v4f32 (scalar_to_vector FPR32:$fj)), diff --git a/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp b/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp index 8fa72bc..d9ea88c 100644 --- a/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp +++ b/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp @@ -254,6 +254,7 @@ bool LoongArchAsmBackend::relaxAlign(MCFragment &F, unsigned &Size) { MCFixup Fixup = MCFixup::create(0, Expr, FirstLiteralRelocationKind + ELF::R_LARCH_ALIGN); F.setVarFixups({Fixup}); + F.setLinkerRelaxable(); F.getParent()->setLinkerRelaxable(); return true; } diff --git a/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchBaseInfo.cpp b/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchBaseInfo.cpp index 03ce004..7cefb3f 100644 --- a/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchBaseInfo.cpp +++ b/llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchBaseInfo.cpp @@ -52,6 +52,9 @@ static ABI getTripleABI(const Triple &TT) { bool Is64Bit = TT.isArch64Bit(); ABI TripleABI; switch (TT.getEnvironment()) { + case llvm::Triple::EnvironmentType::UnknownEnvironment: + TripleABI = ABI_Unknown; + break; case llvm::Triple::EnvironmentType::GNUSF: case llvm::Triple::EnvironmentType::MuslSF: TripleABI = Is64Bit ? ABI_LP64S : ABI_ILP32S; @@ -96,7 +99,7 @@ ABI computeTargetABI(const Triple &TT, const FeatureBitset &FeatureBits, // 1. If the '-target-abi' is valid, use it. if (IsABIValidForFeature(ArgProvidedABI)) { - if (TT.hasEnvironment() && ArgProvidedABI != TripleABI) + if (IsABIValidForFeature(TripleABI) && ArgProvidedABI != TripleABI) errs() << "warning: triple-implied ABI conflicts with provided target-abi '" << ABIName << "', using target-abi\n"; @@ -164,10 +167,7 @@ ABI computeTargetABI(const Triple &TT, const FeatureBitset &FeatureBits, return Is64Bit ? ABI_LP64F : ABI_ILP32F; return Is64Bit ? ABI_LP64S : ABI_ILP32S; }; - if (ABIName.empty()) - errs() << "warning: the triple-implied ABI is invalid, ignoring and using " - "feature-implied ABI\n"; - else + if (!ABIName.empty()) errs() << "warning: both target-abi and the triple-implied ABI are " "invalid, ignoring and using feature-implied ABI\n"; return checkABIStandardized(GetFeatureABI()); |