diff options
Diffstat (limited to 'llvm/lib/Target/X86/X86InstrInfo.cpp')
| -rw-r--r-- | llvm/lib/Target/X86/X86InstrInfo.cpp | 30 |
1 files changed, 25 insertions, 5 deletions
diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp index 2e868a60ce4..acb9128db79 100644 --- a/llvm/lib/Target/X86/X86InstrInfo.cpp +++ b/llvm/lib/Target/X86/X86InstrInfo.cpp @@ -5962,6 +5962,19 @@ static const uint16_t ReplaceableInstrsAVX2[][3] = { { X86::VUNPCKHPSYrr, X86::VUNPCKHPSYrr, X86::VPUNPCKHDQYrr }, }; +static const uint16_t ReplaceableInstrsFP[][3] = { + //PackedSingle PackedDouble + { X86::MOVLPSrm, X86::MOVLPDrm, X86::INSTRUCTION_LIST_END }, + { X86::MOVHPSrm, X86::MOVHPDrm, X86::INSTRUCTION_LIST_END }, + { X86::MOVHPSmr, X86::MOVHPDmr, X86::INSTRUCTION_LIST_END }, + { X86::VMOVLPSrm, X86::VMOVLPDrm, X86::INSTRUCTION_LIST_END }, + { X86::VMOVHPSrm, X86::VMOVHPDrm, X86::INSTRUCTION_LIST_END }, + { X86::VMOVHPSmr, X86::VMOVHPDmr, X86::INSTRUCTION_LIST_END }, + { X86::VMOVLPSZ128rm, X86::VMOVLPDZ128rm, X86::INSTRUCTION_LIST_END }, + { X86::VMOVHPSZ128rm, X86::VMOVHPDZ128rm, X86::INSTRUCTION_LIST_END }, + { X86::VMOVHPSZ128mr, X86::VMOVHPDZ128mr, X86::INSTRUCTION_LIST_END }, +}; + static const uint16_t ReplaceableInstrsAVX2InsertExtract[][3] = { //PackedSingle PackedDouble PackedInt { X86::VEXTRACTF128mr, X86::VEXTRACTF128mr, X86::VEXTRACTI128mr }, @@ -6202,7 +6215,7 @@ static const uint16_t ReplaceableInstrsAVX512DQMasked[][4] = { }; // NOTE: These should only be used by the custom domain methods. -static const uint16_t ReplaceableCustomInstrs[][3] = { +static const uint16_t ReplaceableBlendInstrs[][3] = { //PackedSingle PackedDouble PackedInt { X86::BLENDPSrmi, X86::BLENDPDrmi, X86::PBLENDWrmi }, { X86::BLENDPSrri, X86::BLENDPDrri, X86::PBLENDWrri }, @@ -6211,7 +6224,7 @@ static const uint16_t ReplaceableCustomInstrs[][3] = { { X86::VBLENDPSYrmi, X86::VBLENDPDYrmi, X86::VPBLENDWYrmi }, { X86::VBLENDPSYrri, X86::VBLENDPDYrri, X86::VPBLENDWYrri }, }; -static const uint16_t ReplaceableCustomAVX2Instrs[][3] = { +static const uint16_t ReplaceableBlendAVX2Instrs[][3] = { //PackedSingle PackedDouble PackedInt { X86::VBLENDPSrmi, X86::VBLENDPDrmi, X86::VPBLENDDrmi }, { X86::VBLENDPSrri, X86::VBLENDPDrri, X86::VPBLENDDrri }, @@ -6405,9 +6418,9 @@ bool X86InstrInfo::setExecutionDomainCustom(MachineInstr &MI, Imm = (ImmWidth == 16 ? ((Imm << 8) | Imm) : Imm); unsigned NewImm = Imm; - const uint16_t *table = lookup(Opcode, dom, ReplaceableCustomInstrs); + const uint16_t *table = lookup(Opcode, dom, ReplaceableBlendInstrs); if (!table) - table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs); + table = lookup(Opcode, dom, ReplaceableBlendAVX2Instrs); if (Domain == 1) { // PackedSingle AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm); @@ -6417,7 +6430,7 @@ bool X86InstrInfo::setExecutionDomainCustom(MachineInstr &MI, if (Subtarget.hasAVX2()) { // If we are already VPBLENDW use that, else use VPBLENDD. if ((ImmWidth / (Is256 ? 2 : 1)) != 8) { - table = lookup(Opcode, dom, ReplaceableCustomAVX2Instrs); + table = lookup(Opcode, dom, ReplaceableBlendAVX2Instrs); AdjustBlendMask(Imm, ImmWidth, Is256 ? 8 : 4, &NewImm); } } else { @@ -6525,6 +6538,8 @@ X86InstrInfo::getExecutionDomain(const MachineInstr &MI) const { validDomains = 0xe; } else if (lookup(opcode, domain, ReplaceableInstrsAVX2)) { validDomains = Subtarget.hasAVX2() ? 0xe : 0x6; + } else if (lookup(opcode, domain, ReplaceableInstrsFP)) { + validDomains = 0x6; } else if (lookup(opcode, domain, ReplaceableInstrsAVX2InsertExtract)) { // Insert/extract instructions should only effect domain if AVX2 // is enabled. @@ -6564,6 +6579,11 @@ void X86InstrInfo::setExecutionDomain(MachineInstr &MI, unsigned Domain) const { "256-bit vector operations only available in AVX2"); table = lookup(MI.getOpcode(), dom, ReplaceableInstrsAVX2); } + if (!table) { // try the FP table + table = lookup(MI.getOpcode(), dom, ReplaceableInstrsFP); + assert((!table || Domain < 3) && + "Can only select PackedSingle or PackedDouble"); + } if (!table) { // try the other table assert(Subtarget.hasAVX2() && "256-bit insert/extract only available in AVX2"); |
