X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FSelectionDAG%2FInstrEmitter.cpp;h=5e631c2d8e963e9f7b30e3ca76add8c7e4740324;hb=6c7ccaa3fd1d6e96d0bf922554b09d2b17c3b0e3;hp=610536fec667f6ba93961f251ac81019388a6d8b;hpb=3d7d07ef038696cefcaf3ce5335072964199a78d;p=oota-llvm.git diff --git a/lib/CodeGen/SelectionDAG/InstrEmitter.cpp b/lib/CodeGen/SelectionDAG/InstrEmitter.cpp index 610536fec66..5e631c2d8e9 100644 --- a/lib/CodeGen/SelectionDAG/InstrEmitter.cpp +++ b/lib/CodeGen/SelectionDAG/InstrEmitter.cpp @@ -30,28 +30,51 @@ #include "llvm/Support/MathExtras.h" using namespace llvm; +/// MinRCSize - Smallest register class we allow when constraining virtual +/// registers. If satisfying all register class constraints would require +/// using a smaller register class, emit a COPY to a new virtual register +/// instead. +const unsigned MinRCSize = 4; + /// CountResults - The results of target nodes have register or immediate -/// operands first, then an optional chain, and optional flag operands (which do +/// operands first, then an optional chain, and optional glue operands (which do /// not go into the resulting MachineInstr). unsigned InstrEmitter::CountResults(SDNode *Node) { unsigned N = Node->getNumValues(); - while (N && Node->getValueType(N - 1) == MVT::Flag) + while (N && Node->getValueType(N - 1) == MVT::Glue) --N; if (N && Node->getValueType(N - 1) == MVT::Other) --N; // Skip over chain result. return N; } -/// CountOperands - The inputs to target nodes have any actual inputs first, -/// followed by an optional chain operand, then an optional flag operand. +/// countOperands - The inputs to target nodes have any actual inputs first, +/// followed by an optional chain operand, then an optional glue operand. /// Compute the number of actual operands that will go into the resulting /// MachineInstr. -unsigned InstrEmitter::CountOperands(SDNode *Node) { +/// +/// Also count physreg RegisterSDNode and RegisterMaskSDNode operands preceding +/// the chain and glue. These operands may be implicit on the machine instr. +static unsigned countOperands(SDNode *Node, unsigned NumExpUses, + unsigned &NumImpUses) { unsigned N = Node->getNumOperands(); - while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag) + while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue) --N; if (N && Node->getOperand(N - 1).getValueType() == MVT::Other) --N; // Ignore chain if it exists. + + // Count RegisterSDNode and RegisterMaskSDNode operands for NumImpUses. + NumImpUses = N - NumExpUses; + for (unsigned I = N; I > NumExpUses; --I) { + if (isa(Node->getOperand(I - 1))) + continue; + if (RegisterSDNode *RN = dyn_cast(Node->getOperand(I - 1))) + if (TargetRegisterInfo::isPhysicalRegister(RN->getReg())) + continue; + NumImpUses = N - I; + break; + } + return N; } @@ -67,7 +90,7 @@ EmitCopyFromReg(SDNode *Node, unsigned ResNo, bool IsClone, bool IsCloned, if (IsClone) VRBaseMap.erase(Op); bool isNew = VRBaseMap.insert(std::make_pair(Op, SrcReg)).second; - isNew = isNew; // Silence compiler warning. + (void)isNew; // Silence compiler warning. assert(isNew && "Node emitted out of order - early"); return; } @@ -76,12 +99,18 @@ EmitCopyFromReg(SDNode *Node, unsigned ResNo, bool IsClone, bool IsCloned, // the CopyToReg'd destination register instead of creating a new vreg. bool MatchReg = true; const TargetRegisterClass *UseRC = NULL; + EVT VT = Node->getValueType(ResNo); + + // Stick to the preferred register classes for legal types. + if (TLI->isTypeLegal(VT)) + UseRC = TLI->getRegClassFor(VT); + if (!IsClone && !IsCloned) for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end(); UI != E; ++UI) { SDNode *User = *UI; bool Match = true; - if (User->getOpcode() == ISD::CopyToReg && + if (User->getOpcode() == ISD::CopyToReg && User->getOperand(2).getNode() == Node && User->getOperand(2).getResNo() == ResNo) { unsigned DestReg = cast(User->getOperand(1))->getReg(); @@ -96,18 +125,21 @@ EmitCopyFromReg(SDNode *Node, unsigned ResNo, bool IsClone, bool IsCloned, if (Op.getNode() != Node || Op.getResNo() != ResNo) continue; EVT VT = Node->getValueType(Op.getResNo()); - if (VT == MVT::Other || VT == MVT::Flag) + if (VT == MVT::Other || VT == MVT::Glue) continue; Match = false; if (User->isMachineOpcode()) { - const TargetInstrDesc &II = TII->get(User->getMachineOpcode()); + const MCInstrDesc &II = TII->get(User->getMachineOpcode()); const TargetRegisterClass *RC = 0; - if (i+II.getNumDefs() < II.getNumOperands()) - RC = II.OpInfo[i+II.getNumDefs()].getRegClass(TRI); + if (i+II.getNumDefs() < II.getNumOperands()) { + RC = TRI->getAllocatableClass( + TII->getRegClass(II, i+II.getNumDefs(), TRI, *MF)); + } if (!UseRC) UseRC = RC; else if (RC) { - const TargetRegisterClass *ComRC = getCommonSubClass(UseRC, RC); + const TargetRegisterClass *ComRC = + TRI->getCommonSubClass(UseRC, RC); // If multiple uses expect disjoint register classes, we emit // copies in AddRegisterOperand. if (ComRC) @@ -121,10 +153,9 @@ EmitCopyFromReg(SDNode *Node, unsigned ResNo, bool IsClone, bool IsCloned, break; } - EVT VT = Node->getValueType(ResNo); const TargetRegisterClass *SrcRC = 0, *DstRC = 0; - SrcRC = TRI->getPhysicalRegisterRegClass(SrcReg, VT); - + SrcRC = TRI->getMinimalPhysRegClass(SrcReg, VT); + // Figure out the register class to create for the destreg. if (VRBase) { DstRC = MRI->getRegClass(VRBase); @@ -134,7 +165,7 @@ EmitCopyFromReg(SDNode *Node, unsigned ResNo, bool IsClone, bool IsCloned, } else { DstRC = TLI->getRegClassFor(VT); } - + // If all uses are reading from the src physical register and copying the // register is either impossible or very expensive, then don't create a copy. if (MatchReg && SrcRC->getCopyCost() < 0) { @@ -142,18 +173,15 @@ EmitCopyFromReg(SDNode *Node, unsigned ResNo, bool IsClone, bool IsCloned, } else { // Create the reg, emit the copy. VRBase = MRI->createVirtualRegister(DstRC); - bool Emitted = TII->copyRegToReg(*MBB, InsertPos, VRBase, SrcReg, - DstRC, SrcRC); - - assert(Emitted && "Unable to issue a copy instruction!\n"); - (void) Emitted; + BuildMI(*MBB, InsertPos, Node->getDebugLoc(), TII->get(TargetOpcode::COPY), + VRBase).addReg(SrcReg); } SDValue Op(Node, ResNo); if (IsClone) VRBaseMap.erase(Op); bool isNew = VRBaseMap.insert(std::make_pair(Op, VRBase)).second; - isNew = isNew; // Silence compiler warning. + (void)isNew; // Silence compiler warning. assert(isNew && "Node emitted out of order - early"); } @@ -165,7 +193,7 @@ unsigned InstrEmitter::getDstOfOnlyCopyToRegUse(SDNode *Node, return 0; SDNode *User = *Node->use_begin(); - if (User->getOpcode() == ISD::CopyToReg && + if (User->getOpcode() == ISD::CopyToReg && User->getOperand(2).getNode() == Node && User->getOperand(2).getResNo() == ResNo) { unsigned Reg = cast(User->getOperand(1))->getReg(); @@ -176,7 +204,7 @@ unsigned InstrEmitter::getDstOfOnlyCopyToRegUse(SDNode *Node, } void InstrEmitter::CreateVirtualRegisters(SDNode *Node, MachineInstr *MI, - const TargetInstrDesc &II, + const MCInstrDesc &II, bool IsClone, bool IsCloned, DenseMap &VRBaseMap) { assert(Node->getMachineOpcode() != TargetOpcode::IMPLICIT_DEF && @@ -187,7 +215,8 @@ void InstrEmitter::CreateVirtualRegisters(SDNode *Node, MachineInstr *MI, // is a vreg in the same register class, use the CopyToReg'd destination // register instead of creating a new vreg. unsigned VRBase = 0; - const TargetRegisterClass *RC = II.OpInfo[i].getRegClass(TRI); + const TargetRegisterClass *RC = + TRI->getAllocatableClass(TII->getRegClass(II, i, TRI, *MF)); if (II.OpInfo[i].isOptionalDef()) { // Optional def must be a physical register. unsigned NumResults = CountResults(Node); @@ -200,7 +229,7 @@ void InstrEmitter::CreateVirtualRegisters(SDNode *Node, MachineInstr *MI, for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end(); UI != E; ++UI) { SDNode *User = *UI; - if (User->getOpcode() == ISD::CopyToReg && + if (User->getOpcode() == ISD::CopyToReg && User->getOperand(2).getNode() == Node && User->getOperand(2).getResNo() == i) { unsigned Reg = cast(User->getOperand(1))->getReg(); @@ -227,7 +256,7 @@ void InstrEmitter::CreateVirtualRegisters(SDNode *Node, MachineInstr *MI, if (IsClone) VRBaseMap.erase(Op); bool isNew = VRBaseMap.insert(std::make_pair(Op, VRBase)).second; - isNew = isNew; // Silence compiler warning. + (void)isNew; // Silence compiler warning. assert(isNew && "Node emitted out of order - early"); } } @@ -240,13 +269,13 @@ unsigned InstrEmitter::getVR(SDValue Op, Op.getMachineOpcode() == TargetOpcode::IMPLICIT_DEF) { // Add an IMPLICIT_DEF instruction before every use. unsigned VReg = getDstOfOnlyCopyToRegUse(Op.getNode(), Op.getResNo()); - // IMPLICIT_DEF can produce any type of result so its TargetInstrDesc + // IMPLICIT_DEF can produce any type of result so its MCInstrDesc // does not include operand register class info. if (!VReg) { const TargetRegisterClass *RC = TLI->getRegClassFor(Op.getValueType()); VReg = MRI->createVirtualRegister(RC); } - BuildMI(MBB, Op.getDebugLoc(), + BuildMI(*MBB, InsertPos, Op.getDebugLoc(), TII->get(TargetOpcode::IMPLICIT_DEF), VReg); return VReg; } @@ -263,41 +292,59 @@ unsigned InstrEmitter::getVR(SDValue Op, void InstrEmitter::AddRegisterOperand(MachineInstr *MI, SDValue Op, unsigned IIOpNum, - const TargetInstrDesc *II, + const MCInstrDesc *II, DenseMap &VRBaseMap, - bool IsDebug) { + bool IsDebug, bool IsClone, bool IsCloned) { assert(Op.getValueType() != MVT::Other && - Op.getValueType() != MVT::Flag && - "Chain and flag operands should occur at end of operand list!"); + Op.getValueType() != MVT::Glue && + "Chain and glue operands should occur at end of operand list!"); // Get/emit the operand. unsigned VReg = getVR(Op, VRBaseMap); assert(TargetRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?"); - const TargetInstrDesc &TID = MI->getDesc(); - bool isOptDef = IIOpNum < TID.getNumOperands() && - TID.OpInfo[IIOpNum].isOptionalDef(); + const MCInstrDesc &MCID = MI->getDesc(); + bool isOptDef = IIOpNum < MCID.getNumOperands() && + MCID.OpInfo[IIOpNum].isOptionalDef(); // If the instruction requires a register in a different class, create - // a new virtual register and copy the value into it. + // a new virtual register and copy the value into it, but first attempt to + // shrink VReg's register class within reason. For example, if VReg == GR32 + // and II requires a GR32_NOSP, just constrain VReg to GR32_NOSP. if (II) { - const TargetRegisterClass *SrcRC = MRI->getRegClass(VReg); const TargetRegisterClass *DstRC = 0; if (IIOpNum < II->getNumOperands()) - DstRC = II->OpInfo[IIOpNum].getRegClass(TRI); - assert((DstRC || (TID.isVariadic() && IIOpNum >= TID.getNumOperands())) && - "Don't have operand info for this instruction!"); - if (DstRC && SrcRC != DstRC && !SrcRC->hasSuperClass(DstRC)) { + DstRC = TRI->getAllocatableClass(TII->getRegClass(*II,IIOpNum,TRI,*MF)); + if (DstRC && !MRI->constrainRegClass(VReg, DstRC, MinRCSize)) { unsigned NewVReg = MRI->createVirtualRegister(DstRC); - bool Emitted = TII->copyRegToReg(*MBB, InsertPos, NewVReg, VReg, - DstRC, SrcRC); - assert(Emitted && "Unable to issue a copy instruction!\n"); - (void) Emitted; + BuildMI(*MBB, InsertPos, Op.getNode()->getDebugLoc(), + TII->get(TargetOpcode::COPY), NewVReg).addReg(VReg); VReg = NewVReg; } } + // If this value has only one use, that use is a kill. This is a + // conservative approximation. InstrEmitter does trivial coalescing + // with CopyFromReg nodes, so don't emit kill flags for them. + // Avoid kill flags on Schedule cloned nodes, since there will be + // multiple uses. + // Tied operands are never killed, so we need to check that. And that + // means we need to determine the index of the operand. + bool isKill = Op.hasOneUse() && + Op.getNode()->getOpcode() != ISD::CopyFromReg && + !IsDebug && + !(IsClone || IsCloned); + if (isKill) { + unsigned Idx = MI->getNumOperands(); + while (Idx > 0 && + MI->getOperand(Idx-1).isReg() && MI->getOperand(Idx-1).isImplicit()) + --Idx; + bool isTied = MI->getDesc().getOperandConstraint(Idx, MCOI::TIED_TO) != -1; + if (isTied) + isKill = false; + } + MI->addOperand(MachineOperand::CreateReg(VReg, isOptDef, - false/*isImp*/, false/*isKill*/, + false/*isImp*/, isKill, false/*isDead*/, false/*isUndef*/, false/*isEarlyClobber*/, 0/*SubReg*/, IsDebug)); @@ -305,22 +352,28 @@ InstrEmitter::AddRegisterOperand(MachineInstr *MI, SDValue Op, /// AddOperand - Add the specified operand to the specified machine instr. II /// specifies the instruction information for the node, and IIOpNum is the -/// operand number (in the II) that we are adding. IIOpNum and II are used for -/// assertions only. +/// operand number (in the II) that we are adding. void InstrEmitter::AddOperand(MachineInstr *MI, SDValue Op, unsigned IIOpNum, - const TargetInstrDesc *II, + const MCInstrDesc *II, DenseMap &VRBaseMap, - bool IsDebug) { + bool IsDebug, bool IsClone, bool IsCloned) { if (Op.isMachineOpcode()) { - AddRegisterOperand(MI, Op, IIOpNum, II, VRBaseMap, IsDebug); + AddRegisterOperand(MI, Op, IIOpNum, II, VRBaseMap, + IsDebug, IsClone, IsCloned); } else if (ConstantSDNode *C = dyn_cast(Op)) { MI->addOperand(MachineOperand::CreateImm(C->getSExtValue())); } else if (ConstantFPSDNode *F = dyn_cast(Op)) { const ConstantFP *CFP = F->getConstantFPValue(); MI->addOperand(MachineOperand::CreateFPImm(CFP)); } else if (RegisterSDNode *R = dyn_cast(Op)) { - MI->addOperand(MachineOperand::CreateReg(R->getReg(), false)); + // Turn additional physreg operands into implicit uses on non-variadic + // instructions. This is used by call and return instructions passing + // arguments in registers. + bool Imp = II && (IIOpNum >= II->getNumOperands() && !II->isVariadic()); + MI->addOperand(MachineOperand::CreateReg(R->getReg(), false, Imp)); + } else if (RegisterMaskSDNode *RM = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateRegMask(RM->getRegMask())); } else if (GlobalAddressSDNode *TGA = dyn_cast(Op)) { MI->addOperand(MachineOperand::CreateGA(TGA->getGlobal(), TGA->getOffset(), TGA->getTargetFlags())); @@ -334,7 +387,7 @@ void InstrEmitter::AddOperand(MachineInstr *MI, SDValue Op, } else if (ConstantPoolSDNode *CP = dyn_cast(Op)) { int Offset = CP->getOffset(); unsigned Align = CP->getAlignment(); - const Type *Type = CP->getType(); + Type *Type = CP->getType(); // MachineConstantPool wants an explicit alignment. if (Align == 0) { Align = TM->getTargetData()->getPrefTypeAlignment(Type); @@ -343,7 +396,7 @@ void InstrEmitter::AddOperand(MachineInstr *MI, SDValue Op, Align = TM->getTargetData()->getTypeAllocSize(Type); } } - + unsigned Idx; MachineConstantPool *MCP = MF->getConstantPool(); if (CP->isMachineConstantPoolEntry()) @@ -357,43 +410,59 @@ void InstrEmitter::AddOperand(MachineInstr *MI, SDValue Op, ES->getTargetFlags())); } else if (BlockAddressSDNode *BA = dyn_cast(Op)) { MI->addOperand(MachineOperand::CreateBA(BA->getBlockAddress(), + BA->getOffset(), BA->getTargetFlags())); + } else if (TargetIndexSDNode *TI = dyn_cast(Op)) { + MI->addOperand(MachineOperand::CreateTargetIndex(TI->getIndex(), + TI->getOffset(), + TI->getTargetFlags())); } else { assert(Op.getValueType() != MVT::Other && - Op.getValueType() != MVT::Flag && - "Chain and flag operands should occur at end of operand list!"); - AddRegisterOperand(MI, Op, IIOpNum, II, VRBaseMap, IsDebug); + Op.getValueType() != MVT::Glue && + "Chain and glue operands should occur at end of operand list!"); + AddRegisterOperand(MI, Op, IIOpNum, II, VRBaseMap, + IsDebug, IsClone, IsCloned); } } -/// getSuperRegisterRegClass - Returns the register class of a superreg A whose -/// "SubIdx"'th sub-register class is the specified register class and whose -/// type matches the specified type. -static const TargetRegisterClass* -getSuperRegisterRegClass(const TargetRegisterClass *TRC, - unsigned SubIdx, EVT VT) { - // Pick the register class of the superegister for this type - for (TargetRegisterInfo::regclass_iterator I = TRC->superregclasses_begin(), - E = TRC->superregclasses_end(); I != E; ++I) - if ((*I)->hasType(VT) && (*I)->getSubRegisterRegClass(SubIdx) == TRC) - return *I; - assert(false && "Couldn't find the register class"); - return 0; +unsigned InstrEmitter::ConstrainForSubReg(unsigned VReg, unsigned SubIdx, + EVT VT, DebugLoc DL) { + const TargetRegisterClass *VRC = MRI->getRegClass(VReg); + const TargetRegisterClass *RC = TRI->getSubClassWithSubReg(VRC, SubIdx); + + // RC is a sub-class of VRC that supports SubIdx. Try to constrain VReg + // within reason. + if (RC && RC != VRC) + RC = MRI->constrainRegClass(VReg, RC, MinRCSize); + + // VReg has been adjusted. It can be used with SubIdx operands now. + if (RC) + return VReg; + + // VReg couldn't be reasonably constrained. Emit a COPY to a new virtual + // register instead. + RC = TRI->getSubClassWithSubReg(TLI->getRegClassFor(VT), SubIdx); + assert(RC && "No legal register class for VT supports that SubIdx"); + unsigned NewReg = MRI->createVirtualRegister(RC); + BuildMI(*MBB, InsertPos, DL, TII->get(TargetOpcode::COPY), NewReg) + .addReg(VReg); + return NewReg; } /// EmitSubregNode - Generate machine code for subreg nodes. /// -void InstrEmitter::EmitSubregNode(SDNode *Node, - DenseMap &VRBaseMap){ +void InstrEmitter::EmitSubregNode(SDNode *Node, + DenseMap &VRBaseMap, + bool IsClone, bool IsCloned) { unsigned VRBase = 0; unsigned Opc = Node->getMachineOpcode(); - + // If the node is only used by a CopyToReg and the dest reg is a vreg, use // the CopyToReg'd destination register instead of creating a new vreg. for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end(); UI != E; ++UI) { SDNode *User = *UI; - if (User->getOpcode() == ISD::CopyToReg && + if (User->getOpcode() == ISD::CopyToReg && User->getOperand(2).getNode() == Node) { unsigned DestReg = cast(User->getOperand(1))->getReg(); if (TargetRegisterInfo::isVirtualRegister(DestReg)) { @@ -402,78 +471,97 @@ void InstrEmitter::EmitSubregNode(SDNode *Node, } } } - + if (Opc == TargetOpcode::EXTRACT_SUBREG) { + // EXTRACT_SUBREG is lowered as %dst = COPY %src:sub. There are no + // constraints on the %dst register, COPY can target all legal register + // classes. unsigned SubIdx = cast(Node->getOperand(1))->getZExtValue(); + const TargetRegisterClass *TRC = TLI->getRegClassFor(Node->getValueType(0)); - // Create the extract_subreg machine instruction. - MachineInstr *MI = BuildMI(*MF, Node->getDebugLoc(), - TII->get(TargetOpcode::EXTRACT_SUBREG)); - - // Figure out the register class to create for the destreg. unsigned VReg = getVR(Node->getOperand(0), VRBaseMap); - const TargetRegisterClass *TRC = MRI->getRegClass(VReg); - const TargetRegisterClass *SRC = TRC->getSubRegisterRegClass(SubIdx); - assert(SRC && "Invalid subregister index in EXTRACT_SUBREG"); - - // Figure out the register class to create for the destreg. - // Note that if we're going to directly use an existing register, - // it must be precisely the required class, and not a subclass - // thereof. - if (VRBase == 0 || SRC != MRI->getRegClass(VRBase)) { - // Create the reg - assert(SRC && "Couldn't find source register class"); - VRBase = MRI->createVirtualRegister(SRC); + MachineInstr *DefMI = MRI->getVRegDef(VReg); + unsigned SrcReg, DstReg, DefSubIdx; + if (DefMI && + TII->isCoalescableExtInstr(*DefMI, SrcReg, DstReg, DefSubIdx) && + SubIdx == DefSubIdx && + TRC == MRI->getRegClass(SrcReg)) { + // Optimize these: + // r1025 = s/zext r1024, 4 + // r1026 = extract_subreg r1025, 4 + // to a copy + // r1026 = copy r1024 + VRBase = MRI->createVirtualRegister(TRC); + BuildMI(*MBB, InsertPos, Node->getDebugLoc(), + TII->get(TargetOpcode::COPY), VRBase).addReg(SrcReg); + MRI->clearKillFlags(SrcReg); + } else { + // VReg may not support a SubIdx sub-register, and we may need to + // constrain its register class or issue a COPY to a compatible register + // class. + VReg = ConstrainForSubReg(VReg, SubIdx, + Node->getOperand(0).getValueType(), + Node->getDebugLoc()); + + // Create the destreg if it is missing. + if (VRBase == 0) + VRBase = MRI->createVirtualRegister(TRC); + + // Create the extract_subreg machine instruction. + BuildMI(*MBB, InsertPos, Node->getDebugLoc(), + TII->get(TargetOpcode::COPY), VRBase).addReg(VReg, 0, SubIdx); } - - // Add def, source, and subreg index - MI->addOperand(MachineOperand::CreateReg(VRBase, true)); - AddOperand(MI, Node->getOperand(0), 0, 0, VRBaseMap); - MI->addOperand(MachineOperand::CreateImm(SubIdx)); - MBB->insert(InsertPos, MI); } else if (Opc == TargetOpcode::INSERT_SUBREG || Opc == TargetOpcode::SUBREG_TO_REG) { SDValue N0 = Node->getOperand(0); SDValue N1 = Node->getOperand(1); SDValue N2 = Node->getOperand(2); - unsigned SubReg = getVR(N1, VRBaseMap); unsigned SubIdx = cast(N2)->getZExtValue(); - const TargetRegisterClass *TRC = MRI->getRegClass(SubReg); - const TargetRegisterClass *SRC = - getSuperRegisterRegClass(TRC, SubIdx, - Node->getValueType(0)); - - // Figure out the register class to create for the destreg. - // Note that if we're going to directly use an existing register, - // it must be precisely the required class, and not a subclass - // thereof. - if (VRBase == 0 || SRC != MRI->getRegClass(VRBase)) { - // Create the reg - assert(SRC && "Couldn't find source register class"); + + // Figure out the register class to create for the destreg. It should be + // the largest legal register class supporting SubIdx sub-registers. + // RegisterCoalescer will constrain it further if it decides to eliminate + // the INSERT_SUBREG instruction. + // + // %dst = INSERT_SUBREG %src, %sub, SubIdx + // + // is lowered by TwoAddressInstructionPass to: + // + // %dst = COPY %src + // %dst:SubIdx = COPY %sub + // + // There is no constraint on the %src register class. + // + const TargetRegisterClass *SRC = TLI->getRegClassFor(Node->getValueType(0)); + SRC = TRI->getSubClassWithSubReg(SRC, SubIdx); + assert(SRC && "No register class supports VT and SubIdx for INSERT_SUBREG"); + + if (VRBase == 0 || !SRC->hasSubClassEq(MRI->getRegClass(VRBase))) VRBase = MRI->createVirtualRegister(SRC); - } // Create the insert_subreg or subreg_to_reg machine instruction. MachineInstr *MI = BuildMI(*MF, Node->getDebugLoc(), TII->get(Opc)); MI->addOperand(MachineOperand::CreateReg(VRBase, true)); - + // If creating a subreg_to_reg, then the first input operand // is an implicit value immediate, otherwise it's a register if (Opc == TargetOpcode::SUBREG_TO_REG) { const ConstantSDNode *SD = cast(N0); MI->addOperand(MachineOperand::CreateImm(SD->getZExtValue())); } else - AddOperand(MI, N0, 0, 0, VRBaseMap); + AddOperand(MI, N0, 0, 0, VRBaseMap, /*IsDebug=*/false, + IsClone, IsCloned); // Add the subregster being inserted - AddOperand(MI, N1, 0, 0, VRBaseMap); + AddOperand(MI, N1, 0, 0, VRBaseMap, /*IsDebug=*/false, + IsClone, IsCloned); MI->addOperand(MachineOperand::CreateImm(SubIdx)); MBB->insert(InsertPos, MI); } else llvm_unreachable("Node is not insert_subreg, extract_subreg, or subreg_to_reg"); - + SDValue Op(Node, 0); bool isNew = VRBaseMap.insert(std::make_pair(Op, VRBase)).second; - isNew = isNew; // Silence compiler warning. + (void)isNew; // Silence compiler warning. assert(isNew && "Node emitted out of order - early"); } @@ -485,60 +573,110 @@ void InstrEmitter::EmitCopyToRegClassNode(SDNode *Node, DenseMap &VRBaseMap) { unsigned VReg = getVR(Node->getOperand(0), VRBaseMap); - const TargetRegisterClass *SrcRC = MRI->getRegClass(VReg); - - unsigned DstRCIdx = cast(Node->getOperand(1))->getZExtValue(); - const TargetRegisterClass *DstRC = TRI->getRegClass(DstRCIdx); // Create the new VReg in the destination class and emit a copy. + unsigned DstRCIdx = cast(Node->getOperand(1))->getZExtValue(); + const TargetRegisterClass *DstRC = + TRI->getAllocatableClass(TRI->getRegClass(DstRCIdx)); unsigned NewVReg = MRI->createVirtualRegister(DstRC); - bool Emitted = TII->copyRegToReg(*MBB, InsertPos, NewVReg, VReg, - DstRC, SrcRC); - assert(Emitted && - "Unable to issue a copy instruction for a COPY_TO_REGCLASS node!\n"); - (void) Emitted; + BuildMI(*MBB, InsertPos, Node->getDebugLoc(), TII->get(TargetOpcode::COPY), + NewVReg).addReg(VReg); + + SDValue Op(Node, 0); + bool isNew = VRBaseMap.insert(std::make_pair(Op, NewVReg)).second; + (void)isNew; // Silence compiler warning. + assert(isNew && "Node emitted out of order - early"); +} + +/// EmitRegSequence - Generate machine code for REG_SEQUENCE nodes. +/// +void InstrEmitter::EmitRegSequence(SDNode *Node, + DenseMap &VRBaseMap, + bool IsClone, bool IsCloned) { + unsigned DstRCIdx = cast(Node->getOperand(0))->getZExtValue(); + const TargetRegisterClass *RC = TRI->getRegClass(DstRCIdx); + unsigned NewVReg = MRI->createVirtualRegister(TRI->getAllocatableClass(RC)); + MachineInstr *MI = BuildMI(*MF, Node->getDebugLoc(), + TII->get(TargetOpcode::REG_SEQUENCE), NewVReg); + unsigned NumOps = Node->getNumOperands(); + assert((NumOps & 1) == 1 && + "REG_SEQUENCE must have an odd number of operands!"); + const MCInstrDesc &II = TII->get(TargetOpcode::REG_SEQUENCE); + for (unsigned i = 1; i != NumOps; ++i) { + SDValue Op = Node->getOperand(i); + if ((i & 1) == 0) { + RegisterSDNode *R = dyn_cast(Node->getOperand(i-1)); + // Skip physical registers as they don't have a vreg to get and we'll + // insert copies for them in TwoAddressInstructionPass anyway. + if (!R || !TargetRegisterInfo::isPhysicalRegister(R->getReg())) { + unsigned SubIdx = cast(Op)->getZExtValue(); + unsigned SubReg = getVR(Node->getOperand(i-1), VRBaseMap); + const TargetRegisterClass *TRC = MRI->getRegClass(SubReg); + const TargetRegisterClass *SRC = + TRI->getMatchingSuperRegClass(RC, TRC, SubIdx); + if (SRC && SRC != RC) { + MRI->setRegClass(NewVReg, SRC); + RC = SRC; + } + } + } + AddOperand(MI, Op, i+1, &II, VRBaseMap, /*IsDebug=*/false, + IsClone, IsCloned); + } + MBB->insert(InsertPos, MI); SDValue Op(Node, 0); bool isNew = VRBaseMap.insert(std::make_pair(Op, NewVReg)).second; - isNew = isNew; // Silence compiler warning. + (void)isNew; // Silence compiler warning. assert(isNew && "Node emitted out of order - early"); } /// EmitDbgValue - Generate machine instruction for a dbg_value node. /// -MachineInstr *InstrEmitter::EmitDbgValue(SDDbgValue *SD, - MachineBasicBlock *InsertBB, - DenseMap &VRBaseMap, - DenseMap *EM) { +MachineInstr * +InstrEmitter::EmitDbgValue(SDDbgValue *SD, + DenseMap &VRBaseMap) { uint64_t Offset = SD->getOffset(); MDNode* MDPtr = SD->getMDPtr(); DebugLoc DL = SD->getDebugLoc(); - const TargetInstrDesc &II = TII->get(TargetOpcode::DBG_VALUE); + if (SD->getKind() == SDDbgValue::FRAMEIX) { + // Stack address; this needs to be lowered in target-dependent fashion. + // EmitTargetCodeForFrameDebugValue is responsible for allocation. + unsigned FrameIx = SD->getFrameIx(); + return TII->emitFrameIndexDebugValue(*MF, FrameIx, Offset, MDPtr, DL); + } + // Otherwise, we're going to create an instruction here. + const MCInstrDesc &II = TII->get(TargetOpcode::DBG_VALUE); MachineInstrBuilder MIB = BuildMI(*MF, DL, II); if (SD->getKind() == SDDbgValue::SDNODE) { - AddOperand(&*MIB, SDValue(SD->getSDNode(), SD->getResNo()), - (*MIB).getNumOperands(), &II, VRBaseMap, true /*IsDebug*/); + SDNode *Node = SD->getSDNode(); + SDValue Op = SDValue(Node, SD->getResNo()); + // It's possible we replaced this SDNode with other(s) and therefore + // didn't generate code for it. It's better to catch these cases where + // they happen and transfer the debug info, but trying to guarantee that + // in all cases would be very fragile; this is a safeguard for any + // that were missed. + DenseMap::iterator I = VRBaseMap.find(Op); + if (I==VRBaseMap.end()) + MIB.addReg(0U); // undef + else + AddOperand(&*MIB, Op, (*MIB).getNumOperands(), &II, VRBaseMap, + /*IsDebug=*/true, /*IsClone=*/false, /*IsCloned=*/false); } else if (SD->getKind() == SDDbgValue::CONST) { - Value *V = SD->getConst(); - if (ConstantInt *CI = dyn_cast(V)) { - MIB.addImm(CI->getSExtValue()); - } else if (ConstantFP *CF = dyn_cast(V)) { + const Value *V = SD->getConst(); + if (const ConstantInt *CI = dyn_cast(V)) { + if (CI->getBitWidth() > 64) + MIB.addCImm(CI); + else + MIB.addImm(CI->getSExtValue()); + } else if (const ConstantFP *CF = dyn_cast(V)) { MIB.addFPImm(CF); } else { // Could be an Undef. In any case insert an Undef so we can see what we // dropped. MIB.addReg(0U); } - } else if (SD->getKind() == SDDbgValue::FRAMEIX) { - unsigned FrameIx = SD->getFrameIx(); - // Stack address; this needs to be lowered in target-dependent fashion. - // FIXME test that the target supports this somehow; if not emit Undef. - // Create a pseudo for EmitInstrWithCustomInserter's consumption. - MIB.addImm(FrameIx).addImm(Offset).addMetadata(MDPtr); - abort(); - TLI->EmitInstrWithCustomInserter(&*MIB, InsertBB, EM); - return 0; } else { // Insert an Undef so we can see what we dropped. MIB.addReg(0U); @@ -553,15 +691,14 @@ MachineInstr *InstrEmitter::EmitDbgValue(SDDbgValue *SD, /// void InstrEmitter:: EmitMachineNode(SDNode *Node, bool IsClone, bool IsCloned, - DenseMap &VRBaseMap, - DenseMap *EM) { + DenseMap &VRBaseMap) { unsigned Opc = Node->getMachineOpcode(); - + // Handle subreg insert/extract specially - if (Opc == TargetOpcode::EXTRACT_SUBREG || + if (Opc == TargetOpcode::EXTRACT_SUBREG || Opc == TargetOpcode::INSERT_SUBREG || Opc == TargetOpcode::SUBREG_TO_REG) { - EmitSubregNode(Node, VRBaseMap); + EmitSubregNode(Node, VRBaseMap, IsClone, IsCloned); return; } @@ -571,30 +708,42 @@ EmitMachineNode(SDNode *Node, bool IsClone, bool IsCloned, return; } + // Handle REG_SEQUENCE specially. + if (Opc == TargetOpcode::REG_SEQUENCE) { + EmitRegSequence(Node, VRBaseMap, IsClone, IsCloned); + return; + } + if (Opc == TargetOpcode::IMPLICIT_DEF) // We want a unique VR for each IMPLICIT_DEF use. return; - - const TargetInstrDesc &II = TII->get(Opc); + + const MCInstrDesc &II = TII->get(Opc); unsigned NumResults = CountResults(Node); - unsigned NodeOperands = CountOperands(Node); - bool HasPhysRegOuts = (NumResults > II.getNumDefs()) && - II.getImplicitDefs() != 0; + unsigned NumImpUses = 0; + unsigned NodeOperands = + countOperands(Node, II.getNumOperands() - II.getNumDefs(), NumImpUses); + bool HasPhysRegOuts = NumResults > II.getNumDefs() && II.getImplicitDefs()!=0; #ifndef NDEBUG unsigned NumMIOperands = NodeOperands + NumResults; - assert((II.getNumOperands() == NumMIOperands || - HasPhysRegOuts || II.isVariadic()) && - "#operands for dag node doesn't match .td file!"); + if (II.isVariadic()) + assert(NumMIOperands >= II.getNumOperands() && + "Too few operands for a variadic node!"); + else + assert(NumMIOperands >= II.getNumOperands() && + NumMIOperands <= II.getNumOperands() + II.getNumImplicitDefs() + + NumImpUses && + "#operands for dag node doesn't match .td file!"); #endif // Create the new machine instruction. MachineInstr *MI = BuildMI(*MF, Node->getDebugLoc(), II); - + // Add result register values for things that are defined by this // instruction. if (NumResults) CreateVirtualRegisters(Node, MI, II, IsClone, IsCloned, VRBaseMap); - + // Emit all of the actual operands of this instruction, adding them to the // instruction as appropriate. bool HasOptPRefs = II.getNumDefs() > NumResults; @@ -603,36 +752,80 @@ EmitMachineNode(SDNode *Node, bool IsClone, bool IsCloned, unsigned NumSkip = HasOptPRefs ? II.getNumDefs() - NumResults : 0; for (unsigned i = NumSkip; i != NodeOperands; ++i) AddOperand(MI, Node->getOperand(i), i-NumSkip+II.getNumDefs(), &II, - VRBaseMap); + VRBaseMap, /*IsDebug=*/false, IsClone, IsCloned); // Transfer all of the memory reference descriptions of this instruction. MI->setMemRefs(cast(Node)->memoperands_begin(), cast(Node)->memoperands_end()); - if (II.usesCustomInsertionHook()) { - // Insert this instruction into the basic block using a target - // specific inserter which may returns a new basic block. - MBB = TLI->EmitInstrWithCustomInserter(MI, MBB, EM); - InsertPos = MBB->end(); - } else { - MBB->insert(InsertPos, MI); - } - - // Additional results must be an physical register def. + // Insert the instruction into position in the block. This needs to + // happen before any custom inserter hook is called so that the + // hook knows where in the block to insert the replacement code. + MBB->insert(InsertPos, MI); + + // The MachineInstr may also define physregs instead of virtregs. These + // physreg values can reach other instructions in different ways: + // + // 1. When there is a use of a Node value beyond the explicitly defined + // virtual registers, we emit a CopyFromReg for one of the implicitly + // defined physregs. This only happens when HasPhysRegOuts is true. + // + // 2. A CopyFromReg reading a physreg may be glued to this instruction. + // + // 3. A glued instruction may implicitly use a physreg. + // + // 4. A glued instruction may use a RegisterSDNode operand. + // + // Collect all the used physreg defs, and make sure that any unused physreg + // defs are marked as dead. + SmallVector UsedRegs; + + // Additional results must be physical register defs. if (HasPhysRegOuts) { for (unsigned i = II.getNumDefs(); i < NumResults; ++i) { unsigned Reg = II.getImplicitDefs()[i - II.getNumDefs()]; - if (Node->hasAnyUseOfValue(i)) - EmitCopyFromReg(Node, i, IsClone, IsCloned, Reg, VRBaseMap); - // If there are no uses, mark the register as dead now, so that - // MachineLICM/Sink can see that it's dead. Don't do this if the - // node has a Flag value, for the benefit of targets still using - // Flag for values in physregs. - else if (Node->getValueType(Node->getNumValues()-1) != MVT::Flag) - MI->addRegisterDead(Reg, TRI); + if (!Node->hasAnyUseOfValue(i)) + continue; + // This implicitly defined physreg has a use. + UsedRegs.push_back(Reg); + EmitCopyFromReg(Node, i, IsClone, IsCloned, Reg, VRBaseMap); + } + } + + // Scan the glue chain for any used physregs. + if (Node->getValueType(Node->getNumValues()-1) == MVT::Glue) { + for (SDNode *F = Node->getGluedUser(); F; F = F->getGluedUser()) { + if (F->getOpcode() == ISD::CopyFromReg) { + UsedRegs.push_back(cast(F->getOperand(1))->getReg()); + continue; + } else if (F->getOpcode() == ISD::CopyToReg) { + // Skip CopyToReg nodes that are internal to the glue chain. + continue; + } + // Collect declared implicit uses. + const MCInstrDesc &MCID = TII->get(F->getMachineOpcode()); + UsedRegs.append(MCID.getImplicitUses(), + MCID.getImplicitUses() + MCID.getNumImplicitUses()); + // In addition to declared implicit uses, we must also check for + // direct RegisterSDNode operands. + for (unsigned i = 0, e = F->getNumOperands(); i != e; ++i) + if (RegisterSDNode *R = dyn_cast(F->getOperand(i))) { + unsigned Reg = R->getReg(); + if (TargetRegisterInfo::isPhysicalRegister(Reg)) + UsedRegs.push_back(Reg); + } } } - return; + + // Finally mark unused registers as dead. + if (!UsedRegs.empty() || II.getImplicitDefs()) + MI->setPhysRegsDeadExcept(UsedRegs, *TRI); + + // Run post-isel target hook to adjust this instruction if needed. +#ifdef NDEBUG + if (II.hasPostISelHook()) +#endif + TLI->AdjustInstrPostInstrSelection(MI, Node); } /// EmitSpecialNode - Generate machine code for a target-independent node and @@ -646,10 +839,8 @@ EmitSpecialNode(SDNode *Node, bool IsClone, bool IsCloned, Node->dump(); #endif llvm_unreachable("This target-independent node should have been selected!"); - break; case ISD::EntryToken: llvm_unreachable("EntryToken should have been excluded from the schedule!"); - break; case ISD::MERGE_VALUES: case ISD::TokenFactor: // fall thru break; @@ -660,28 +851,13 @@ EmitSpecialNode(SDNode *Node, bool IsClone, bool IsCloned, SrcReg = R->getReg(); else SrcReg = getVR(SrcVal, VRBaseMap); - + unsigned DestReg = cast(Node->getOperand(1))->getReg(); if (SrcReg == DestReg) // Coalesced away the copy? Ignore. break; - - const TargetRegisterClass *SrcTRC = 0, *DstTRC = 0; - // Get the register classes of the src/dst. - if (TargetRegisterInfo::isVirtualRegister(SrcReg)) - SrcTRC = MRI->getRegClass(SrcReg); - else - SrcTRC = TRI->getPhysicalRegisterRegClass(SrcReg,SrcVal.getValueType()); - - if (TargetRegisterInfo::isVirtualRegister(DestReg)) - DstTRC = MRI->getRegClass(DestReg); - else - DstTRC = TRI->getPhysicalRegisterRegClass(DestReg, - Node->getOperand(1).getValueType()); - bool Emitted = TII->copyRegToReg(*MBB, InsertPos, DestReg, SrcReg, - DstTRC, SrcTRC); - assert(Emitted && "Unable to issue a copy instruction!\n"); - (void) Emitted; + BuildMI(*MBB, InsertPos, Node->getDebugLoc(), TII->get(TargetOpcode::COPY), + DestReg).addReg(SrcReg); break; } case ISD::CopyFromReg: { @@ -695,55 +871,104 @@ EmitSpecialNode(SDNode *Node, bool IsClone, bool IsCloned, TII->get(TargetOpcode::EH_LABEL)).addSym(S); break; } - + + case ISD::LIFETIME_START: + case ISD::LIFETIME_END: { + unsigned TarOp = (Node->getOpcode() == ISD::LIFETIME_START) ? + TargetOpcode::LIFETIME_START : TargetOpcode::LIFETIME_END; + + FrameIndexSDNode *FI = dyn_cast(Node->getOperand(1)); + BuildMI(*MBB, InsertPos, Node->getDebugLoc(), TII->get(TarOp)) + .addFrameIndex(FI->getIndex()); + break; + } + case ISD::INLINEASM: { unsigned NumOps = Node->getNumOperands(); - if (Node->getOperand(NumOps-1).getValueType() == MVT::Flag) - --NumOps; // Ignore the flag operand. - + if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue) + --NumOps; // Ignore the glue operand. + // Create the inline asm machine instruction. MachineInstr *MI = BuildMI(*MF, Node->getDebugLoc(), TII->get(TargetOpcode::INLINEASM)); // Add the asm string as an external symbol operand. - const char *AsmStr = - cast(Node->getOperand(1))->getSymbol(); + SDValue AsmStrV = Node->getOperand(InlineAsm::Op_AsmString); + const char *AsmStr = cast(AsmStrV)->getSymbol(); MI->addOperand(MachineOperand::CreateES(AsmStr)); - + + // Add the HasSideEffect and isAlignStack bits. + int64_t ExtraInfo = + cast(Node->getOperand(InlineAsm::Op_ExtraInfo))-> + getZExtValue(); + MI->addOperand(MachineOperand::CreateImm(ExtraInfo)); + + // Remember to operand index of the group flags. + SmallVector GroupIdx; + // Add all of the operand registers to the instruction. - for (unsigned i = 2; i != NumOps;) { + for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) { unsigned Flags = cast(Node->getOperand(i))->getZExtValue(); - unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags); - + const unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags); + + GroupIdx.push_back(MI->getNumOperands()); MI->addOperand(MachineOperand::CreateImm(Flags)); ++i; // Skip the ID value. - - switch (Flags & 7) { + + switch (InlineAsm::getKind(Flags)) { default: llvm_unreachable("Bad flags!"); - case 2: // Def of register. - for (; NumVals; --NumVals, ++i) { + case InlineAsm::Kind_RegDef: + for (unsigned j = 0; j != NumVals; ++j, ++i) { unsigned Reg = cast(Node->getOperand(i))->getReg(); - MI->addOperand(MachineOperand::CreateReg(Reg, true)); + // FIXME: Add dead flags for physical and virtual registers defined. + // For now, mark physical register defs as implicit to help fast + // regalloc. This makes inline asm look a lot like calls. + MI->addOperand(MachineOperand::CreateReg(Reg, true, + /*isImp=*/ TargetRegisterInfo::isPhysicalRegister(Reg))); } break; - case 6: // Def of earlyclobber register. - for (; NumVals; --NumVals, ++i) { + case InlineAsm::Kind_RegDefEarlyClobber: + case InlineAsm::Kind_Clobber: + for (unsigned j = 0; j != NumVals; ++j, ++i) { unsigned Reg = cast(Node->getOperand(i))->getReg(); - MI->addOperand(MachineOperand::CreateReg(Reg, true, false, false, - false, false, true)); + MI->addOperand(MachineOperand::CreateReg(Reg, /*isDef=*/ true, + /*isImp=*/ TargetRegisterInfo::isPhysicalRegister(Reg), + /*isKill=*/ false, + /*isDead=*/ false, + /*isUndef=*/false, + /*isEarlyClobber=*/ true)); } break; - case 1: // Use of register. - case 3: // Immediate. - case 4: // Addressing mode. + case InlineAsm::Kind_RegUse: // Use of register. + case InlineAsm::Kind_Imm: // Immediate. + case InlineAsm::Kind_Mem: // Addressing mode. // The addressing mode has been selected, just add all of the // operands to the machine instruction. - for (; NumVals; --NumVals, ++i) - AddOperand(MI, Node->getOperand(i), 0, 0, VRBaseMap); + for (unsigned j = 0; j != NumVals; ++j, ++i) + AddOperand(MI, Node->getOperand(i), 0, 0, VRBaseMap, + /*IsDebug=*/false, IsClone, IsCloned); + + // Manually set isTied bits. + if (InlineAsm::getKind(Flags) == InlineAsm::Kind_RegUse) { + unsigned DefGroup = 0; + if (InlineAsm::isUseOperandTiedToDef(Flags, DefGroup)) { + unsigned DefIdx = GroupIdx[DefGroup] + 1; + unsigned UseIdx = GroupIdx.back() + 1; + for (unsigned j = 0; j != NumVals; ++j) + MI->tieOperands(DefIdx + j, UseIdx + j); + } + } break; } } + + // Get the mdnode from the asm if it exists and add it to the instruction. + SDValue MDV = Node->getOperand(InlineAsm::Op_MDNode); + const MDNode *MD = cast(MDV)->getMD(); + if (MD) + MI->addOperand(MachineOperand::CreateMetadata(MD)); + MBB->insert(InsertPos, MI); break; }