diff options
Diffstat (limited to 'lib/Target/X86/X86ISelLowering.cpp')
-rw-r--r-- | lib/Target/X86/X86ISelLowering.cpp | 108 |
1 files changed, 78 insertions, 30 deletions
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp index 9b83aad..04299f3 100644 --- a/lib/Target/X86/X86ISelLowering.cpp +++ b/lib/Target/X86/X86ISelLowering.cpp @@ -2935,6 +2935,7 @@ static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT, case X86ISD::PSHUFHW: case X86ISD::PSHUFLW: case X86ISD::VPERMILP: + case X86ISD::VPERMI: return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8)); } } @@ -3976,6 +3977,27 @@ unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) { return Index / NumElemsPerChunk; } +/// getShuffleCLImmediate - Return the appropriate immediate to shuffle +/// the specified VECTOR_SHUFFLE mask with VPERMQ and VPERMPD instructions. +/// Handles 256-bit. +static unsigned getShuffleCLImmediate(ShuffleVectorSDNode *N) { + EVT VT = N->getValueType(0); + + unsigned NumElts = VT.getVectorNumElements(); + + assert((VT.is256BitVector() && NumElts == 4) && + "Unsupported vector type for VPERMQ/VPERMPD"); + + unsigned Mask = 0; + for (unsigned i = 0; i != NumElts; ++i) { + int Elt = N->getMaskElt(i); + if (Elt < 0) + continue; + Mask |= Elt << (i*2); + } + + return Mask; +} /// isZeroNode - Returns true if Elt is a constant zero or a floating point /// constant +0.0. bool X86::isZeroNode(SDValue Elt) { @@ -4408,6 +4430,7 @@ static bool getTargetShuffleMask(SDNode *N, EVT VT, case X86ISD::VPERM2X128: ImmN = N->getOperand(N->getNumOperands()-1); DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask); + if (Mask.empty()) return false; break; case X86ISD::MOVDDUP: case X86ISD::MOVLHPD: @@ -6628,6 +6651,23 @@ X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const { if (BlendOp.getNode()) return BlendOp; + if (V2IsUndef && HasAVX2 && (VT == MVT::v8i32 || VT == MVT::v8f32)) { + SmallVector<SDValue, 8> permclMask; + for (unsigned i = 0; i != 8; ++i) { + permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MVT::i32)); + } + SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, + &permclMask[0], 8); + // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32 + return DAG.getNode(X86ISD::VPERMV, dl, VT, + DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1); + } + + if (V2IsUndef && HasAVX2 && (VT == MVT::v4i64 || VT == MVT::v4f64)) + return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, + getShuffleCLImmediate(SVOp), DAG); + + //===--------------------------------------------------------------------===// // Since no target specific shuffle was selected for this generic one, // lower it into other known shuffles. FIXME: this isn't true yet, but @@ -9552,12 +9592,12 @@ X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const case Intrinsic::x86_avx2_vperm2i128: return DAG.getNode(X86ISD::VPERM2X128, dl, Op.getValueType(), Op.getOperand(1), Op.getOperand(2), Op.getOperand(3)); - case Intrinsic::x86_avx_vpermil_ps: - case Intrinsic::x86_avx_vpermil_pd: - case Intrinsic::x86_avx_vpermil_ps_256: - case Intrinsic::x86_avx_vpermil_pd_256: - return DAG.getNode(X86ISD::VPERMILP, dl, Op.getValueType(), - Op.getOperand(1), Op.getOperand(2)); + case Intrinsic::x86_avx2_permd: + case Intrinsic::x86_avx2_permps: + // Operands intentionally swapped. Mask is last operand to intrinsic, + // but second operand for node/intruction. + return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(), + Op.getOperand(2), Op.getOperand(1)); // ptest and testp intrinsics. The intrinsic these come from are designed to // return an integer value, not just an instruction so lower it to the ptest @@ -11141,6 +11181,8 @@ const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const { case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST"; case X86ISD::VPERMILP: return "X86ISD::VPERMILP"; case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128"; + case X86ISD::VPERMV: return "X86ISD::VPERMV"; + case X86ISD::VPERMI: return "X86ISD::VPERMI"; case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ"; case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS"; case X86ISD::VAARG_64: return "X86ISD::VAARG_64"; @@ -11298,14 +11340,15 @@ X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr, unsigned notOpc, unsigned EAXreg, const TargetRegisterClass *RC, - bool invSrc) const { + bool Invert) const { // For the atomic bitwise operator, we generate // thisMBB: // newMBB: // ld t1 = [bitinstr.addr] // op t2 = t1, [bitinstr.val] + // not t3 = t2 (if Invert) // mov EAX = t1 - // lcs dest = [bitinstr.addr], t2 [EAX is implicit] + // lcs dest = [bitinstr.addr], t3 [EAX is implicit] // bz newMBB // fallthrough -->nextMBB const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); @@ -11353,13 +11396,6 @@ X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr, for (int i=0; i <= lastAddrIndx; ++i) (*MIB).addOperand(*argOpers[i]); - unsigned tt = F->getRegInfo().createVirtualRegister(RC); - if (invSrc) { - MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1); - } - else - tt = t1; - unsigned t2 = F->getRegInfo().createVirtualRegister(RC); assert((argOpers[valArgIndx]->isReg() || argOpers[valArgIndx]->isImm()) && @@ -11368,16 +11404,23 @@ X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr, MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2); else MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2); - MIB.addReg(tt); + MIB.addReg(t1); (*MIB).addOperand(*argOpers[valArgIndx]); + unsigned t3 = F->getRegInfo().createVirtualRegister(RC); + if (Invert) { + MIB = BuildMI(newMBB, dl, TII->get(notOpc), t3).addReg(t2); + } + else + t3 = t2; + MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg); MIB.addReg(t1); MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc)); for (int i=0; i <= lastAddrIndx; ++i) (*MIB).addOperand(*argOpers[i]); - MIB.addReg(t2); + MIB.addReg(t3); assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand"); (*MIB).setMemRefs(bInstr->memoperands_begin(), bInstr->memoperands_end()); @@ -11400,7 +11443,7 @@ X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr, unsigned regOpcH, unsigned immOpcL, unsigned immOpcH, - bool invSrc) const { + bool Invert) const { // For the atomic bitwise operator, we generate // thisMBB (instructions are in pairs, except cmpxchg8b) // ld t1,t2 = [bitinstr.addr] @@ -11408,6 +11451,7 @@ X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr, // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4) // op t5, t6 <- out1, out2, [bitinstr.val] // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val]) + // neg t7, t8 < t5, t6 (if Invert) // mov ECX, EBX <- t5, t6 // mov EAX, EDX <- t1, t2 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit] @@ -11491,16 +11535,9 @@ X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr, .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB); // The subsequent operations should be using the destination registers of - //the PHI instructions. - if (invSrc) { - t1 = F->getRegInfo().createVirtualRegister(RC); - t2 = F->getRegInfo().createVirtualRegister(RC); - MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg()); - MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg()); - } else { - t1 = dest1Oper.getReg(); - t2 = dest2Oper.getReg(); - } + // the PHI instructions. + t1 = dest1Oper.getReg(); + t2 = dest2Oper.getReg(); int valArgIndx = lastAddrIndx + 1; assert((argOpers[valArgIndx]->isReg() || @@ -11527,15 +11564,26 @@ X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr, MIB.addReg(t2); (*MIB).addOperand(*argOpers[valArgIndx + 1]); + unsigned t7, t8; + if (Invert) { + t7 = F->getRegInfo().createVirtualRegister(RC); + t8 = F->getRegInfo().createVirtualRegister(RC); + MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t7).addReg(t5); + MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t8).addReg(t6); + } else { + t7 = t5; + t8 = t6; + } + MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX); MIB.addReg(t1); MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX); MIB.addReg(t2); MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX); - MIB.addReg(t5); + MIB.addReg(t7); MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX); - MIB.addReg(t6); + MIB.addReg(t8); MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B)); for (int i=0; i <= lastAddrIndx; ++i) |