From 614051a1c534aff052152b0162a414b3271e8fca Mon Sep 17 00:00:00 2001 From: Akira Hatanaka Date: Tue, 16 Aug 2011 03:51:51 +0000 Subject: [PATCH] Fix handling of double precision loads and stores when Mips1 is targeted. Mips1 does not support double precision loads or stores, therefore two single precision loads or stores must be used in place of these instructions. This patch treats double precision loads and stores as if they are legal instructions until MCInstLowering, instead of generating the single precision instructions during instruction selection or Prolog/Epilog code insertion. Without the changes made in this patch, llc produces code that has the same problem described in r137484 or bails out when MipsInstrInfo::storeRegToStackSlot or loadRegFromStackSlot is called before register allocation. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@137711 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/Mips/MipsAsmPrinter.cpp | 21 +++- lib/Target/Mips/MipsISelDAGToDAG.cpp | 139 --------------------------- lib/Target/Mips/MipsInstrFPU.td | 2 +- lib/Target/Mips/MipsInstrInfo.cpp | 30 +----- lib/Target/Mips/MipsMCInstLower.cpp | 44 +++++++++ lib/Target/Mips/MipsMCInstLower.h | 4 + test/CodeGen/Mips/select.ll | 2 +- 7 files changed, 74 insertions(+), 168 deletions(-) diff --git a/lib/Target/Mips/MipsAsmPrinter.cpp b/lib/Target/Mips/MipsAsmPrinter.cpp index 4013d6c1100..28358abc00f 100644 --- a/lib/Target/Mips/MipsAsmPrinter.cpp +++ b/lib/Target/Mips/MipsAsmPrinter.cpp @@ -18,6 +18,7 @@ #include "MipsInstrInfo.h" #include "MipsMachineFunction.h" #include "MipsMCInstLower.h" +#include "MipsMCSymbolRefExpr.h" #include "InstPrinter/MipsInstPrinter.h" #include "llvm/BasicBlock.h" #include "llvm/Instructions.h" @@ -36,6 +37,7 @@ #include "llvm/Target/TargetOptions.h" #include "llvm/Target/TargetRegistry.h" #include "llvm/ADT/SmallString.h" +#include "llvm/ADT/SmallVector.h" #include "llvm/ADT/StringExtras.h" #include "llvm/ADT/Twine.h" #include "llvm/Support/raw_ostream.h" @@ -53,9 +55,26 @@ void MipsAsmPrinter::EmitInstruction(const MachineInstr *MI) { } MipsMCInstLower MCInstLowering(Mang, *MF, *this); + unsigned Opc = MI->getOpcode(); + + // If target is Mips1, expand double precision load/store to two single + // precision loads/stores (and delay slot if MI is a load). + if (Subtarget->isMips1() && (Opc == Mips::LDC1 || Opc == Mips::SDC1)) { + SmallVector MCInsts; + const unsigned* SubReg = + TM.getRegisterInfo()->getSubRegisters(MI->getOperand(0).getReg()); + MCInstLowering.LowerMips1F64LoadStore(MI, Opc, MCInsts, + Subtarget->isLittle(), SubReg); + + for (SmallVector::iterator I = MCInsts.begin(); + I != MCInsts.end(); ++I) + OutStreamer.EmitInstruction(*I); + + return; + } + MCInst TmpInst0; MCInstLowering.Lower(MI, TmpInst0); - unsigned Opc = MI->getOpcode(); // Convert aligned loads/stores to their unaligned counterparts. // FIXME: expand other unaligned memory accesses too. diff --git a/lib/Target/Mips/MipsISelDAGToDAG.cpp b/lib/Target/Mips/MipsISelDAGToDAG.cpp index 47015056e14..63463bb5048 100644 --- a/lib/Target/Mips/MipsISelDAGToDAG.cpp +++ b/lib/Target/Mips/MipsISelDAGToDAG.cpp @@ -86,9 +86,6 @@ private: // Complex Pattern. bool SelectAddr(SDValue N, SDValue &Base, SDValue &Offset); - SDNode *SelectLoadFp64(SDNode *N); - SDNode *SelectStoreFp64(SDNode *N); - // getI32Imm - Return a target constant with the specified // value, of type i32. inline SDValue getI32Imm(unsigned Imm) { @@ -184,130 +181,6 @@ SelectAddr(SDValue Addr, SDValue &Base, SDValue &Offset) { return true; } -SDNode *MipsDAGToDAGISel::SelectLoadFp64(SDNode *N) { - MVT::SimpleValueType NVT = - N->getValueType(0).getSimpleVT().SimpleTy; - - if (!Subtarget.isMips1() || NVT != MVT::f64) - return NULL; - - LoadSDNode *LN = cast(N); - if (LN->getExtensionType() != ISD::NON_EXTLOAD || - LN->getAddressingMode() != ISD::UNINDEXED) - return NULL; - - SDValue Chain = N->getOperand(0); - SDValue N1 = N->getOperand(1); - SDValue Offset0, Offset1, Base; - - if (!SelectAddr(N1, Base, Offset0) || - N1.getValueType() != MVT::i32) - return NULL; - - MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); - MemRefs0[0] = cast(N)->getMemOperand(); - DebugLoc dl = N->getDebugLoc(); - - // The second load should start after for 4 bytes. - if (ConstantSDNode *C = dyn_cast(Offset0)) - Offset1 = CurDAG->getTargetConstant(C->getSExtValue()+4, MVT::i32); - else if (ConstantPoolSDNode *CP = dyn_cast(Offset0)) - Offset1 = CurDAG->getTargetConstantPool(CP->getConstVal(), - MVT::i32, - CP->getAlignment(), - CP->getOffset()+4, - CP->getTargetFlags()); - else - return NULL; - - // Choose the offsets depending on the endianess - if (TM.getTargetData()->isBigEndian()) - std::swap(Offset0, Offset1); - - // Instead of: - // ldc $f0, X($3) - // Generate: - // lwc $f0, X($3) - // lwc $f1, X+4($3) - SDNode *LD0 = CurDAG->getMachineNode(Mips::LWC1, dl, MVT::f32, - MVT::Other, Base, Offset0, Chain); - SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, - dl, NVT), 0); - SDValue I0 = CurDAG->getTargetInsertSubreg(Mips::sub_fpeven, dl, - MVT::f64, Undef, SDValue(LD0, 0)); - - SDNode *LD1 = CurDAG->getMachineNode(Mips::LWC1, dl, MVT::f32, - MVT::Other, Base, Offset1, SDValue(LD0, 1)); - SDValue I1 = CurDAG->getTargetInsertSubreg(Mips::sub_fpodd, dl, - MVT::f64, I0, SDValue(LD1, 0)); - - ReplaceUses(SDValue(N, 0), I1); - ReplaceUses(SDValue(N, 1), Chain); - cast(LD0)->setMemRefs(MemRefs0, MemRefs0 + 1); - cast(LD1)->setMemRefs(MemRefs0, MemRefs0 + 1); - return I1.getNode(); -} - -SDNode *MipsDAGToDAGISel::SelectStoreFp64(SDNode *N) { - - if (!Subtarget.isMips1() || - N->getOperand(1).getValueType() != MVT::f64) - return NULL; - - SDValue Chain = N->getOperand(0); - - StoreSDNode *SN = cast(N); - if (SN->isTruncatingStore() || SN->getAddressingMode() != ISD::UNINDEXED) - return NULL; - - SDValue N1 = N->getOperand(1); - SDValue N2 = N->getOperand(2); - SDValue Offset0, Offset1, Base; - - if (!SelectAddr(N2, Base, Offset0) || - N1.getValueType() != MVT::f64 || - N2.getValueType() != MVT::i32) - return NULL; - - MachineSDNode::mmo_iterator MemRefs0 = MF->allocateMemRefsArray(1); - MemRefs0[0] = cast(N)->getMemOperand(); - DebugLoc dl = N->getDebugLoc(); - - // Get the even and odd part from the f64 register - SDValue FPOdd = CurDAG->getTargetExtractSubreg(Mips::sub_fpodd, - dl, MVT::f32, N1); - SDValue FPEven = CurDAG->getTargetExtractSubreg(Mips::sub_fpeven, - dl, MVT::f32, N1); - - // The second store should start after for 4 bytes. - if (ConstantSDNode *C = dyn_cast(Offset0)) - Offset1 = CurDAG->getTargetConstant(C->getSExtValue()+4, MVT::i32); - else - return NULL; - - // Choose the offsets depending on the endianess - if (TM.getTargetData()->isBigEndian()) - std::swap(Offset0, Offset1); - - // Instead of: - // sdc $f0, X($3) - // Generate: - // swc $f0, X($3) - // swc $f1, X+4($3) - SDValue Ops0[] = { FPEven, Base, Offset0, Chain }; - Chain = SDValue(CurDAG->getMachineNode(Mips::SWC1, dl, - MVT::Other, Ops0, 4), 0); - cast(Chain.getNode())->setMemRefs(MemRefs0, MemRefs0 + 1); - - SDValue Ops1[] = { FPOdd, Base, Offset1, Chain }; - Chain = SDValue(CurDAG->getMachineNode(Mips::SWC1, dl, - MVT::Other, Ops1, 4), 0); - cast(Chain.getNode())->setMemRefs(MemRefs0, MemRefs0 + 1); - - ReplaceUses(SDValue(N, 0), Chain); - return Chain.getNode(); -} - /// Select instructions not customized! Used for /// expanded, promoted and normal instructions SDNode* MipsDAGToDAGISel::Select(SDNode *Node) { @@ -423,18 +296,6 @@ SDNode* MipsDAGToDAGISel::Select(SDNode *Node) { break; } - case ISD::LOAD: - if (SDNode *ResNode = SelectLoadFp64(Node)) - return ResNode; - // Other cases are autogenerated. - break; - - case ISD::STORE: - if (SDNode *ResNode = SelectStoreFp64(Node)) - return ResNode; - // Other cases are autogenerated. - break; - case MipsISD::ThreadPointer: { unsigned SrcReg = Mips::HWR29; unsigned DestReg = Mips::V1; diff --git a/lib/Target/Mips/MipsInstrFPU.td b/lib/Target/Mips/MipsInstrFPU.td index 021c1679977..24e95aec4c7 100644 --- a/lib/Target/Mips/MipsInstrFPU.td +++ b/lib/Target/Mips/MipsInstrFPU.td @@ -190,7 +190,7 @@ def FMOV_D32 : FFR<0x11, 0b000110, 0x1, (outs AFGR64:$fd), (ins AFGR64:$fs), "mov.d\t$fd, $fs", []>; /// Floating Point Memory Instructions -let Predicates = [IsNotSingleFloat, IsNotMipsI] in { +let Predicates = [IsNotSingleFloat] in { def LDC1 : FFI<0b110101, (outs AFGR64:$ft), (ins mem:$addr), "ldc1\t$ft, $addr", [(set AFGR64:$ft, (load addr:$addr))]>; diff --git a/lib/Target/Mips/MipsInstrInfo.cpp b/lib/Target/Mips/MipsInstrInfo.cpp index 0a7a7f2dfe4..d04cc3b066f 100644 --- a/lib/Target/Mips/MipsInstrInfo.cpp +++ b/lib/Target/Mips/MipsInstrInfo.cpp @@ -175,21 +175,9 @@ storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, BuildMI(MBB, I, DL, get(Mips::SWC1)).addReg(SrcReg, getKillRegState(isKill)) .addFrameIndex(FI).addImm(0); else if (RC == Mips::AFGR64RegisterClass) { - if (!TM.getSubtarget().isMips1()) { - BuildMI(MBB, I, DL, get(Mips::SDC1)) - .addReg(SrcReg, getKillRegState(isKill)) - .addFrameIndex(FI).addImm(0); - } else { - const TargetRegisterInfo *TRI = - MBB.getParent()->getTarget().getRegisterInfo(); - const unsigned *SubSet = TRI->getSubRegisters(SrcReg); - BuildMI(MBB, I, DL, get(Mips::SWC1)) - .addReg(SubSet[0], getKillRegState(isKill)) - .addFrameIndex(FI).addImm(0); - BuildMI(MBB, I, DL, get(Mips::SWC1)) - .addReg(SubSet[1], getKillRegState(isKill)) - .addFrameIndex(FI).addImm(4); - } + BuildMI(MBB, I, DL, get(Mips::SDC1)) + .addReg(SrcReg, getKillRegState(isKill)) + .addFrameIndex(FI).addImm(0); } else llvm_unreachable("Register class not handled!"); } @@ -208,17 +196,7 @@ loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, else if (RC == Mips::FGR32RegisterClass) BuildMI(MBB, I, DL, get(Mips::LWC1), DestReg).addFrameIndex(FI).addImm(0); else if (RC == Mips::AFGR64RegisterClass) { - if (!TM.getSubtarget().isMips1()) { - BuildMI(MBB, I, DL, get(Mips::LDC1), DestReg).addFrameIndex(FI).addImm(0); - } else { - const TargetRegisterInfo *TRI = - MBB.getParent()->getTarget().getRegisterInfo(); - const unsigned *SubSet = TRI->getSubRegisters(DestReg); - BuildMI(MBB, I, DL, get(Mips::LWC1), SubSet[0]) - .addFrameIndex(FI).addImm(0); - BuildMI(MBB, I, DL, get(Mips::LWC1), SubSet[1]) - .addFrameIndex(FI).addImm(4); - } + BuildMI(MBB, I, DL, get(Mips::LDC1), DestReg).addFrameIndex(FI).addImm(0); } else llvm_unreachable("Register class not handled!"); } diff --git a/lib/Target/Mips/MipsMCInstLower.cpp b/lib/Target/Mips/MipsMCInstLower.cpp index 16ec41ef51b..58e18443055 100644 --- a/lib/Target/Mips/MipsMCInstLower.cpp +++ b/lib/Target/Mips/MipsMCInstLower.cpp @@ -83,6 +83,50 @@ MCOperand MipsMCInstLower::LowerSymbolOperand(const MachineOperand &MO, Ctx)); } +// If target is Mips1, expand double precision load/store to two single +// precision loads/stores. +// +// ldc1 $f0, lo($CPI0_0)($5) gets expanded to the following two instructions: +// (little endian) +// lwc1 $f0, lo($CPI0_0)($5) and +// lwc1 $f1, lo($CPI0_0+4)($5) +// (big endian) +// lwc1 $f1, lo($CPI0_0)($5) and +// lwc1 $f0, lo($CPI0_0+4)($5) +void MipsMCInstLower::LowerMips1F64LoadStore(const MachineInstr *MI, + unsigned Opc, + SmallVector& MCInsts, + bool isLittle, + const unsigned *SubReg) const { + MCInst InstLo, InstHi, DelaySlot; + unsigned SingleOpc = (Opc == Mips::LDC1 ? Mips::LWC1 : Mips::SWC1); + unsigned RegLo = isLittle ? *SubReg : *(SubReg + 1); + unsigned RegHi = isLittle ? *(SubReg + 1) : *SubReg; + const MachineOperand &MO1 = MI->getOperand(1); + const MachineOperand &MO2 = MI->getOperand(2); + + InstLo.setOpcode(SingleOpc); + InstLo.addOperand(MCOperand::CreateReg(RegLo)); + InstLo.addOperand(LowerOperand(MO1)); + InstLo.addOperand(LowerOperand(MO2)); + MCInsts.push_back(InstLo); + + InstHi.setOpcode(SingleOpc); + InstHi.addOperand(MCOperand::CreateReg(RegHi)); + InstHi.addOperand(LowerOperand(MO1)); + if (MO2.isImm())// The offset of addr operand is an immediate: e.g. 0($sp) + InstHi.addOperand(MCOperand::CreateImm(MO2.getImm() + 4)); + else// Otherwise, the offset must be a symbol: e.g. lo($CPI0_0)($5) + InstHi.addOperand(LowerSymbolOperand(MO2, MO2.getType(), 4)); + MCInsts.push_back(InstHi); + + // Need to insert a NOP in LWC1's delay slot. + if (SingleOpc == Mips::LWC1) { + DelaySlot.setOpcode(Mips::NOP); + MCInsts.push_back(DelaySlot); + } +} + MCOperand MipsMCInstLower::LowerOperand(const MachineOperand& MO) const { MachineOperandType MOTy = MO.getType(); diff --git a/lib/Target/Mips/MipsMCInstLower.h b/lib/Target/Mips/MipsMCInstLower.h index f401536e01c..d59f3f107ff 100644 --- a/lib/Target/Mips/MipsMCInstLower.h +++ b/lib/Target/Mips/MipsMCInstLower.h @@ -9,6 +9,7 @@ #ifndef MIPSMCINSTLOWER_H #define MIPSMCINSTLOWER_H +#include "llvm/ADT/SmallVector.h" #include "llvm/CodeGen/MachineOperand.h" #include "llvm/Support/Compiler.h" @@ -34,6 +35,9 @@ public: MipsMCInstLower(Mangler *mang, const MachineFunction &MF, MipsAsmPrinter &asmprinter); void Lower(const MachineInstr *MI, MCInst &OutMI) const; + void LowerMips1F64LoadStore(const MachineInstr *MI, unsigned Opc, + SmallVector& MCInsts, + bool isLittle, const unsigned *SubReg) const; private: MCOperand LowerSymbolOperand(const MachineOperand &MO, MachineOperandType MOTy, unsigned Offset) const; diff --git a/test/CodeGen/Mips/select.ll b/test/CodeGen/Mips/select.ll index c83fa3ece02..3ea4c4b2e33 100644 --- a/test/CodeGen/Mips/select.ll +++ b/test/CodeGen/Mips/select.ll @@ -25,7 +25,7 @@ entry: define double @sel2_1(i32 %s, double %f0, double %f1) nounwind readnone { entry: ; CHECK-MIPS32R2: movn.d -; CHECK-MIPS1: beq +; CHECK-MIPS1: bne %tobool = icmp ne i32 %s, 0 %cond = select i1 %tobool, double %f0, double %f1 ret double %cond