diff options
Diffstat (limited to 'llvm/lib/Target/X86')
| -rw-r--r-- | llvm/lib/Target/X86/X86InstrAVX512.td | 2 | ||||
| -rw-r--r-- | llvm/lib/Target/X86/X86InstrInfo.cpp | 34 |
2 files changed, 35 insertions, 1 deletions
diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td index b8f2999..2371ed4 100644 --- a/llvm/lib/Target/X86/X86InstrAVX512.td +++ b/llvm/lib/Target/X86/X86InstrAVX512.td @@ -3238,6 +3238,7 @@ multiclass avx512_load<bits<8> opc, string OpcodeStr, string Name, (_.VT _.RC:$src1), (_.VT _.RC:$src0))))], _.ExeDomain>, EVEX, EVEX_K, Sched<[Sched.RR]>; + let mayLoad = 1, canFoldAsLoad = 1 in def rmk : AVX512PI<opc, MRMSrcMem, (outs _.RC:$dst), (ins _.RC:$src0, _.KRCWM:$mask, _.MemOp:$src1), !strconcat(OpcodeStr, "\t{$src1, ${dst} {${mask}}|", @@ -3248,6 +3249,7 @@ multiclass avx512_load<bits<8> opc, string OpcodeStr, string Name, (_.VT _.RC:$src0))))], _.ExeDomain>, EVEX, EVEX_K, Sched<[Sched.RM]>; } + let mayLoad = 1, canFoldAsLoad = 1 in def rmkz : AVX512PI<opc, MRMSrcMem, (outs _.RC:$dst), (ins _.KRCWM:$mask, _.MemOp:$src), OpcodeStr #"\t{$src, ${dst} {${mask}} {z}|"# diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp index 03ac1d3..1d2cd39 100644 --- a/llvm/lib/Target/X86/X86InstrInfo.cpp +++ b/llvm/lib/Target/X86/X86InstrInfo.cpp @@ -8113,6 +8113,39 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl( MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, LiveIntervals *LIS) const { + // If LoadMI is a masked load, check MI having the same mask. + const MCInstrDesc &MCID = get(LoadMI.getOpcode()); + unsigned NumOps = MCID.getNumOperands(); + if (NumOps >= 3) { + Register MaskReg; + const MachineOperand &Op1 = LoadMI.getOperand(1); + const MachineOperand &Op2 = LoadMI.getOperand(2); + + auto IsVKWMClass = [](const TargetRegisterClass *RC) { + return RC == &X86::VK2WMRegClass || RC == &X86::VK4WMRegClass || + RC == &X86::VK8WMRegClass || RC == &X86::VK16WMRegClass || + RC == &X86::VK32WMRegClass || RC == &X86::VK64WMRegClass; + }; + + if (Op1.isReg() && IsVKWMClass(getRegClass(MCID, 1, &RI))) + MaskReg = Op1.getReg(); + else if (Op2.isReg() && IsVKWMClass(getRegClass(MCID, 2, &RI))) + MaskReg = Op2.getReg(); + + if (MaskReg) { + bool HasSameMask = false; + for (unsigned I = 1, E = MI.getDesc().getNumOperands(); I < E; ++I) { + const MachineOperand &Op = MI.getOperand(I); + if (Op.isReg() && Op.getReg() == MaskReg) { + HasSameMask = true; + break; + } + } + if (!HasSameMask) + return nullptr; + } + } + // TODO: Support the case where LoadMI loads a wide register, but MI // only uses a subreg. for (auto Op : Ops) { @@ -8121,7 +8154,6 @@ MachineInstr *X86InstrInfo::foldMemoryOperandImpl( } // If loading from a FrameIndex, fold directly from the FrameIndex. - unsigned NumOps = LoadMI.getDesc().getNumOperands(); int FrameIndex; if (isLoadFromStackSlot(LoadMI, FrameIndex)) { if (isNonFoldablePartialRegisterLoad(LoadMI, MI, MF)) |