diff options
Diffstat (limited to 'lib/Target/ARM/ARMISelDAGToDAG.cpp')
| -rw-r--r-- | lib/Target/ARM/ARMISelDAGToDAG.cpp | 144 |
1 files changed, 71 insertions, 73 deletions
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp index 013e00a..71207c8 100644 --- a/lib/Target/ARM/ARMISelDAGToDAG.cpp +++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp @@ -80,8 +80,7 @@ public: SDValue &Mode); bool SelectAddrMode5(SDNode *Op, SDValue N, SDValue &Base, SDValue &Offset); - bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Update, - SDValue &Opc, SDValue &Align); + bool SelectAddrMode6(SDNode *Op, SDValue N, SDValue &Addr, SDValue &Align); bool SelectAddrModePC(SDNode *Op, SDValue N, SDValue &Offset, SDValue &Label); @@ -502,12 +501,8 @@ bool ARMDAGToDAGISel::SelectAddrMode5(SDNode *Op, SDValue N, } bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Op, SDValue N, - SDValue &Addr, SDValue &Update, - SDValue &Opc, SDValue &Align) { + SDValue &Addr, SDValue &Align) { Addr = N; - // Default to no writeback. - Update = CurDAG->getRegister(0, MVT::i32); - Opc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(false), MVT::i32); // Default to no alignment. Align = CurDAG->getTargetConstant(0, MVT::i32); return true; @@ -1030,8 +1025,8 @@ SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs, assert(NumVecs >=2 && NumVecs <= 4 && "VLD NumVecs out-of-range"); DebugLoc dl = N->getDebugLoc(); - SDValue MemAddr, MemUpdate, MemOpc, Align; - if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, MemUpdate, MemOpc, Align)) + SDValue MemAddr, Align; + if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align)) return NULL; SDValue Chain = N->getOperand(0); @@ -1055,14 +1050,13 @@ SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs, } SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32); - SDValue PredReg = CurDAG->getRegister(0, MVT::i32); + SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); if (is64BitVector) { unsigned Opc = DOpcodes[OpcodeIndex]; - const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Align, - Pred, PredReg, Chain }; + const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain }; std::vector<EVT> ResTys(NumVecs, VT); ResTys.push_back(MVT::Other); - return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 7); + return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5); } EVT RegVT = GetNEONSubregVT(VT); @@ -1070,11 +1064,10 @@ SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs, // Quad registers are directly supported for VLD2, // loading 2 pairs of D regs. unsigned Opc = QOpcodes0[OpcodeIndex]; - const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Align, - Pred, PredReg, Chain }; + const SDValue Ops[] = { MemAddr, Align, Pred, Reg0, Chain }; std::vector<EVT> ResTys(4, VT); ResTys.push_back(MVT::Other); - SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 7); + SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 5); Chain = SDValue(VLd, 4); // Combine the even and odd subregs to produce the result. @@ -1086,25 +1079,21 @@ SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, unsigned NumVecs, // Otherwise, quad registers are loaded with two separate instructions, // where one loads the even registers and the other loads the odd registers. - // Enable writeback to the address register. - MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32); - std::vector<EVT> ResTys(NumVecs, RegVT); ResTys.push_back(MemAddr.getValueType()); ResTys.push_back(MVT::Other); // Load the even subregs. unsigned Opc = QOpcodes0[OpcodeIndex]; - const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Align, - Pred, PredReg, Chain }; - SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 7); + const SDValue OpsA[] = { MemAddr, Align, Reg0, Pred, Reg0, Chain }; + SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 6); Chain = SDValue(VLdA, NumVecs+1); // Load the odd subregs. Opc = QOpcodes1[OpcodeIndex]; - const SDValue OpsB[] = { SDValue(VLdA, NumVecs), MemUpdate, MemOpc, - Align, Pred, PredReg, Chain }; - SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 7); + const SDValue OpsB[] = { SDValue(VLdA, NumVecs), + Align, Reg0, Pred, Reg0, Chain }; + SDNode *VLdB = CurDAG->getMachineNode(Opc, dl, ResTys, OpsB, 6); Chain = SDValue(VLdB, NumVecs+1); // Combine the even and odd subregs to produce the result. @@ -1123,8 +1112,8 @@ SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs, assert(NumVecs >=2 && NumVecs <= 4 && "VST NumVecs out-of-range"); DebugLoc dl = N->getDebugLoc(); - SDValue MemAddr, MemUpdate, MemOpc, Align; - if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, MemUpdate, MemOpc, Align)) + SDValue MemAddr, Align; + if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align)) return NULL; SDValue Chain = N->getOperand(0); @@ -1148,12 +1137,10 @@ SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs, } SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32); - SDValue PredReg = CurDAG->getRegister(0, MVT::i32); + SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); - SmallVector<SDValue, 8> Ops; + SmallVector<SDValue, 10> Ops; Ops.push_back(MemAddr); - Ops.push_back(MemUpdate); - Ops.push_back(MemOpc); Ops.push_back(Align); if (is64BitVector) { @@ -1161,9 +1148,9 @@ SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs, for (unsigned Vec = 0; Vec < NumVecs; ++Vec) Ops.push_back(N->getOperand(Vec+3)); Ops.push_back(Pred); - Ops.push_back(PredReg); + Ops.push_back(Reg0); // predicate register Ops.push_back(Chain); - return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+7); + return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+5); } EVT RegVT = GetNEONSubregVT(VT); @@ -1178,40 +1165,37 @@ SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, unsigned NumVecs, N->getOperand(Vec+3))); } Ops.push_back(Pred); - Ops.push_back(PredReg); + Ops.push_back(Reg0); // predicate register Ops.push_back(Chain); - return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 11); + return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), 9); } // Otherwise, quad registers are stored with two separate instructions, // where one stores the even registers and the other stores the odd registers. - // Enable writeback to the address register. - MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32); + Ops.push_back(Reg0); // post-access address offset // Store the even subregs. for (unsigned Vec = 0; Vec < NumVecs; ++Vec) Ops.push_back(CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT, N->getOperand(Vec+3))); Ops.push_back(Pred); - Ops.push_back(PredReg); + Ops.push_back(Reg0); // predicate register Ops.push_back(Chain); unsigned Opc = QOpcodes0[OpcodeIndex]; SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(), - MVT::Other, Ops.data(), NumVecs+7); + MVT::Other, Ops.data(), NumVecs+6); Chain = SDValue(VStA, 1); // Store the odd subregs. Ops[0] = SDValue(VStA, 0); // MemAddr for (unsigned Vec = 0; Vec < NumVecs; ++Vec) - Ops[Vec+4] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT, + Ops[Vec+3] = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT, N->getOperand(Vec+3)); - Ops[NumVecs+4] = Pred; - Ops[NumVecs+5] = PredReg; - Ops[NumVecs+6] = Chain; + Ops[NumVecs+5] = Chain; Opc = QOpcodes1[OpcodeIndex]; SDNode *VStB = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(), - MVT::Other, Ops.data(), NumVecs+7); + MVT::Other, Ops.data(), NumVecs+6); Chain = SDValue(VStB, 1); ReplaceUses(SDValue(N, 0), Chain); return NULL; @@ -1224,8 +1208,8 @@ SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad, assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range"); DebugLoc dl = N->getDebugLoc(); - SDValue MemAddr, MemUpdate, MemOpc, Align; - if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, MemUpdate, MemOpc, Align)) + SDValue MemAddr, Align; + if (!SelectAddrMode6(N, N->getOperand(2), MemAddr, Align)) return NULL; SDValue Chain = N->getOperand(0); @@ -1259,12 +1243,10 @@ SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad, } SDValue Pred = CurDAG->getTargetConstant(14, MVT::i32); - SDValue PredReg = CurDAG->getRegister(0, MVT::i32); + SDValue Reg0 = CurDAG->getRegister(0, MVT::i32); - SmallVector<SDValue, 9> Ops; + SmallVector<SDValue, 10> Ops; Ops.push_back(MemAddr); - Ops.push_back(MemUpdate); - Ops.push_back(MemOpc); Ops.push_back(Align); unsigned Opc = 0; @@ -1287,16 +1269,16 @@ SDNode *ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad, } Ops.push_back(getI32Imm(Lane)); Ops.push_back(Pred); - Ops.push_back(PredReg); + Ops.push_back(Reg0); Ops.push_back(Chain); if (!IsLoad) - return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+8); + return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops.data(), NumVecs+6); std::vector<EVT> ResTys(NumVecs, RegVT); ResTys.push_back(MVT::Other); SDNode *VLdLn = - CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), NumVecs+8); + CurDAG->getMachineNode(Opc, dl, ResTys, Ops.data(), NumVecs+6); // For a 64-bit vector load to D registers, nothing more needs to be done. if (is64BitVector) return VLdLn; @@ -1859,37 +1841,45 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) { case Intrinsic::arm_neon_vld3: { unsigned DOpcodes[] = { ARM::VLD3d8, ARM::VLD3d16, ARM::VLD3d32, ARM::VLD3d64 }; - unsigned QOpcodes0[] = { ARM::VLD3q8a, ARM::VLD3q16a, ARM::VLD3q32a }; - unsigned QOpcodes1[] = { ARM::VLD3q8b, ARM::VLD3q16b, ARM::VLD3q32b }; + unsigned QOpcodes0[] = { ARM::VLD3q8_UPD, + ARM::VLD3q16_UPD, + ARM::VLD3q32_UPD }; + unsigned QOpcodes1[] = { ARM::VLD3q8odd_UPD, + ARM::VLD3q16odd_UPD, + ARM::VLD3q32odd_UPD }; return SelectVLD(N, 3, DOpcodes, QOpcodes0, QOpcodes1); } case Intrinsic::arm_neon_vld4: { unsigned DOpcodes[] = { ARM::VLD4d8, ARM::VLD4d16, ARM::VLD4d32, ARM::VLD4d64 }; - unsigned QOpcodes0[] = { ARM::VLD4q8a, ARM::VLD4q16a, ARM::VLD4q32a }; - unsigned QOpcodes1[] = { ARM::VLD4q8b, ARM::VLD4q16b, ARM::VLD4q32b }; + unsigned QOpcodes0[] = { ARM::VLD4q8_UPD, + ARM::VLD4q16_UPD, + ARM::VLD4q32_UPD }; + unsigned QOpcodes1[] = { ARM::VLD4q8odd_UPD, + ARM::VLD4q16odd_UPD, + ARM::VLD4q32odd_UPD }; return SelectVLD(N, 4, DOpcodes, QOpcodes0, QOpcodes1); } case Intrinsic::arm_neon_vld2lane: { unsigned DOpcodes[] = { ARM::VLD2LNd8, ARM::VLD2LNd16, ARM::VLD2LNd32 }; - unsigned QOpcodes0[] = { ARM::VLD2LNq16a, ARM::VLD2LNq32a }; - unsigned QOpcodes1[] = { ARM::VLD2LNq16b, ARM::VLD2LNq32b }; + unsigned QOpcodes0[] = { ARM::VLD2LNq16, ARM::VLD2LNq32 }; + unsigned QOpcodes1[] = { ARM::VLD2LNq16odd, ARM::VLD2LNq32odd }; return SelectVLDSTLane(N, true, 2, DOpcodes, QOpcodes0, QOpcodes1); } case Intrinsic::arm_neon_vld3lane: { unsigned DOpcodes[] = { ARM::VLD3LNd8, ARM::VLD3LNd16, ARM::VLD3LNd32 }; - unsigned QOpcodes0[] = { ARM::VLD3LNq16a, ARM::VLD3LNq32a }; - unsigned QOpcodes1[] = { ARM::VLD3LNq16b, ARM::VLD3LNq32b }; + unsigned QOpcodes0[] = { ARM::VLD3LNq16, ARM::VLD3LNq32 }; + unsigned QOpcodes1[] = { ARM::VLD3LNq16odd, ARM::VLD3LNq32odd }; return SelectVLDSTLane(N, true, 3, DOpcodes, QOpcodes0, QOpcodes1); } case Intrinsic::arm_neon_vld4lane: { unsigned DOpcodes[] = { ARM::VLD4LNd8, ARM::VLD4LNd16, ARM::VLD4LNd32 }; - unsigned QOpcodes0[] = { ARM::VLD4LNq16a, ARM::VLD4LNq32a }; - unsigned QOpcodes1[] = { ARM::VLD4LNq16b, ARM::VLD4LNq32b }; + unsigned QOpcodes0[] = { ARM::VLD4LNq16, ARM::VLD4LNq32 }; + unsigned QOpcodes1[] = { ARM::VLD4LNq16odd, ARM::VLD4LNq32odd }; return SelectVLDSTLane(N, true, 4, DOpcodes, QOpcodes0, QOpcodes1); } @@ -1903,37 +1893,45 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) { case Intrinsic::arm_neon_vst3: { unsigned DOpcodes[] = { ARM::VST3d8, ARM::VST3d16, ARM::VST3d32, ARM::VST3d64 }; - unsigned QOpcodes0[] = { ARM::VST3q8a, ARM::VST3q16a, ARM::VST3q32a }; - unsigned QOpcodes1[] = { ARM::VST3q8b, ARM::VST3q16b, ARM::VST3q32b }; + unsigned QOpcodes0[] = { ARM::VST3q8_UPD, + ARM::VST3q16_UPD, + ARM::VST3q32_UPD }; + unsigned QOpcodes1[] = { ARM::VST3q8odd_UPD, + ARM::VST3q16odd_UPD, + ARM::VST3q32odd_UPD }; return SelectVST(N, 3, DOpcodes, QOpcodes0, QOpcodes1); } case Intrinsic::arm_neon_vst4: { unsigned DOpcodes[] = { ARM::VST4d8, ARM::VST4d16, ARM::VST4d32, ARM::VST4d64 }; - unsigned QOpcodes0[] = { ARM::VST4q8a, ARM::VST4q16a, ARM::VST4q32a }; - unsigned QOpcodes1[] = { ARM::VST4q8b, ARM::VST4q16b, ARM::VST4q32b }; + unsigned QOpcodes0[] = { ARM::VST4q8_UPD, + ARM::VST4q16_UPD, + ARM::VST4q32_UPD }; + unsigned QOpcodes1[] = { ARM::VST4q8odd_UPD, + ARM::VST4q16odd_UPD, + ARM::VST4q32odd_UPD }; return SelectVST(N, 4, DOpcodes, QOpcodes0, QOpcodes1); } case Intrinsic::arm_neon_vst2lane: { unsigned DOpcodes[] = { ARM::VST2LNd8, ARM::VST2LNd16, ARM::VST2LNd32 }; - unsigned QOpcodes0[] = { ARM::VST2LNq16a, ARM::VST2LNq32a }; - unsigned QOpcodes1[] = { ARM::VST2LNq16b, ARM::VST2LNq32b }; + unsigned QOpcodes0[] = { ARM::VST2LNq16, ARM::VST2LNq32 }; + unsigned QOpcodes1[] = { ARM::VST2LNq16odd, ARM::VST2LNq32odd }; return SelectVLDSTLane(N, false, 2, DOpcodes, QOpcodes0, QOpcodes1); } case Intrinsic::arm_neon_vst3lane: { unsigned DOpcodes[] = { ARM::VST3LNd8, ARM::VST3LNd16, ARM::VST3LNd32 }; - unsigned QOpcodes0[] = { ARM::VST3LNq16a, ARM::VST3LNq32a }; - unsigned QOpcodes1[] = { ARM::VST3LNq16b, ARM::VST3LNq32b }; + unsigned QOpcodes0[] = { ARM::VST3LNq16, ARM::VST3LNq32 }; + unsigned QOpcodes1[] = { ARM::VST3LNq16odd, ARM::VST3LNq32odd }; return SelectVLDSTLane(N, false, 3, DOpcodes, QOpcodes0, QOpcodes1); } case Intrinsic::arm_neon_vst4lane: { unsigned DOpcodes[] = { ARM::VST4LNd8, ARM::VST4LNd16, ARM::VST4LNd32 }; - unsigned QOpcodes0[] = { ARM::VST4LNq16a, ARM::VST4LNq32a }; - unsigned QOpcodes1[] = { ARM::VST4LNq16b, ARM::VST4LNq32b }; + unsigned QOpcodes0[] = { ARM::VST4LNq16, ARM::VST4LNq32 }; + unsigned QOpcodes1[] = { ARM::VST4LNq16odd, ARM::VST4LNq32odd }; return SelectVLDSTLane(N, false, 4, DOpcodes, QOpcodes0, QOpcodes1); } } |
