diff options
author | Chris Lattner <sabre@nondot.org> | 2006-01-12 01:46:07 +0000 |
---|---|---|
committer | Chris Lattner <sabre@nondot.org> | 2006-01-12 01:46:07 +0000 |
commit | 05f1fe8d448013531f86a24b5173fe57330291ef (patch) | |
tree | 24f58fba2ad290ad82d6d671c8ecb78f6dbe1114 /lib/Target/PowerPC/PPCISelPattern.cpp | |
parent | 271426a99a6afb5a884b6dcf2e50a6a9680ec839 (diff) |
Goodbye PPC pattern isel. You have served us well, but it is now time for
you to ride off into the sunset.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@25236 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Target/PowerPC/PPCISelPattern.cpp')
-rw-r--r-- | lib/Target/PowerPC/PPCISelPattern.cpp | 1713 |
1 files changed, 0 insertions, 1713 deletions
diff --git a/lib/Target/PowerPC/PPCISelPattern.cpp b/lib/Target/PowerPC/PPCISelPattern.cpp deleted file mode 100644 index bae19f8fb7..0000000000 --- a/lib/Target/PowerPC/PPCISelPattern.cpp +++ /dev/null @@ -1,1713 +0,0 @@ -//===-- PPC32ISelPattern.cpp - A pattern matching inst selector for PPC32 -===// -// -// The LLVM Compiler Infrastructure -// -// This file was developed by Nate Begeman and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. -// -//===----------------------------------------------------------------------===// -// -// This file defines a pattern matching instruction selector for 32 bit PowerPC. -// Magic number generation for integer divide from the PowerPC Compiler Writer's -// Guide, section 3.2.3.5 -// -//===----------------------------------------------------------------------===// - -#include "PPC.h" -#include "PPCInstrBuilder.h" -#include "PPCTargetMachine.h" -#include "PPCISelLowering.h" -#include "llvm/Constants.h" -#include "llvm/Function.h" -#include "llvm/CodeGen/MachineConstantPool.h" -#include "llvm/CodeGen/MachineFunction.h" -#include "llvm/CodeGen/SelectionDAG.h" -#include "llvm/CodeGen/SelectionDAGISel.h" -#include "llvm/CodeGen/SSARegMap.h" -#include "llvm/Target/TargetData.h" -#include "llvm/Target/TargetOptions.h" -#include "llvm/Support/Debug.h" -#include "llvm/Support/MathExtras.h" -#include "llvm/ADT/Statistic.h" -#include <set> -#include <algorithm> -using namespace llvm; - -namespace { -Statistic<> Recorded("ppc-codegen", "Number of recording ops emitted"); -Statistic<> FusedFP ("ppc-codegen", "Number of fused fp operations"); -Statistic<> FrameOff("ppc-codegen", "Number of frame idx offsets collapsed"); - -//===--------------------------------------------------------------------===// -// ISel - PPC32 specific code to select PPC32 machine instructions for -// SelectionDAG operations. -//===--------------------------------------------------------------------===// - -class ISel : public SelectionDAGISel { - PPCTargetLowering PPCLowering; - SelectionDAG *ISelDAG; // Hack to support us having a dag->dag transform - // for sdiv and udiv until it is put into the future - // dag combiner. - - /// ExprMap - As shared expressions are codegen'd, we keep track of which - /// vreg the value is produced in, so we only emit one copy of each compiled - /// tree. - std::map<SDOperand, unsigned> ExprMap; - - unsigned GlobalBaseReg; - bool GlobalBaseInitialized; - bool RecordSuccess; -public: - ISel(TargetMachine &TM) : SelectionDAGISel(PPCLowering), PPCLowering(TM), - ISelDAG(0) {} - - /// runOnFunction - Override this function in order to reset our per-function - /// variables. - virtual bool runOnFunction(Function &Fn) { - // Make sure we re-emit a set of the global base reg if necessary - GlobalBaseInitialized = false; - return SelectionDAGISel::runOnFunction(Fn); - } - - /// InstructionSelectBasicBlock - This callback is invoked by - /// SelectionDAGISel when it has created a SelectionDAG for us to codegen. - virtual void InstructionSelectBasicBlock(SelectionDAG &DAG) { - DEBUG(BB->dump()); - // Codegen the basic block. - ISelDAG = &DAG; - Select(DAG.getRoot()); - - // Clear state used for selection. - ExprMap.clear(); - ISelDAG = 0; - } - - // convenience functions for virtual register creation - inline unsigned MakeIntReg() { - return RegMap->createVirtualRegister(PPC::GPRCRegisterClass); - } - - // dag -> dag expanders for integer divide by constant - SDOperand BuildSDIVSequence(SDOperand N); - SDOperand BuildUDIVSequence(SDOperand N); - - unsigned getGlobalBaseReg(); - void MoveCRtoGPR(unsigned CCReg, ISD::CondCode CC, unsigned Result); - bool SelectBitfieldInsert(SDOperand OR, unsigned Result); - unsigned FoldIfWideZeroExtend(SDOperand N); - unsigned SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC); - bool SelectIntImmediateExpr(SDOperand N, unsigned Result, - unsigned OCHi, unsigned OCLo, - bool IsArithmetic = false, bool Negate = false); - unsigned SelectExpr(SDOperand N, bool Recording=false); - void Select(SDOperand N); - - unsigned SelectAddr(SDOperand N, unsigned& Reg, int& offset); - void SelectBranchCC(SDOperand N); - - virtual const char *getPassName() const { - return "PowerPC Pattern Instruction Selection"; - } -}; - -// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with -// any number of 0s on either side. The 1s are allowed to wrap from LSB to -// MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is -// not, since all 1s are not contiguous. -static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) { - if (isShiftedMask_32(Val)) { - // look for the first non-zero bit - MB = CountLeadingZeros_32(Val); - // look for the first zero bit after the run of ones - ME = CountLeadingZeros_32((Val - 1) ^ Val); - return true; - } else if (isShiftedMask_32(Val = ~Val)) { // invert mask - // effectively look for the first zero bit - ME = CountLeadingZeros_32(Val) - 1; - // effectively look for the first one bit after the run of zeros - MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1; - return true; - } - // no run present - return false; -} - -// isRotateAndMask - Returns true if Mask and Shift can be folded in to a rotate -// and mask opcode and mask operation. -static bool isRotateAndMask(unsigned Opcode, unsigned Shift, unsigned Mask, - bool IsShiftMask, - unsigned &SH, unsigned &MB, unsigned &ME) { - if (Shift > 31) return false; - unsigned Indeterminant = ~0; // bit mask marking indeterminant results - - if (Opcode == ISD::SHL) { // shift left - // apply shift to mask if it comes first - if (IsShiftMask) Mask = Mask << Shift; - // determine which bits are made indeterminant by shift - Indeterminant = ~(0xFFFFFFFFu << Shift); - } else if (Opcode == ISD::SRL) { // shift rights - // apply shift to mask if it comes first - if (IsShiftMask) Mask = Mask >> Shift; - // determine which bits are made indeterminant by shift - Indeterminant = ~(0xFFFFFFFFu >> Shift); - // adjust for the left rotate - Shift = 32 - Shift; - } - - // if the mask doesn't intersect any Indeterminant bits - if (Mask && !(Mask & Indeterminant)) { - SH = Shift; - // make sure the mask is still a mask (wrap arounds may not be) - return isRunOfOnes(Mask, MB, ME); - } - - // can't do it - return false; -} - -// isIntImmediate - This method tests to see if a constant operand. -// If so Imm will receive the 32 bit value. -static bool isIntImmediate(SDOperand N, unsigned& Imm) { - // test for constant - if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) { - // retrieve value - Imm = (unsigned)CN->getValue(); - // passes muster - return true; - } - // not a constant - return false; -} - -// isOpcWithIntImmediate - This method tests to see if the node is a specific -// opcode and that it has a immediate integer right operand. -// If so Imm will receive the 32 bit value. -static bool isOpcWithIntImmediate(SDOperand N, unsigned Opc, unsigned& Imm) { - return N.getOpcode() == Opc && isIntImmediate(N.getOperand(1), Imm); -} - -// isOprShiftImm - Returns true if the specified operand is a shift opcode with -// a immediate shift count less than 32. -static bool isOprShiftImm(SDOperand N, unsigned& Opc, unsigned& SH) { - Opc = N.getOpcode(); - return (Opc == ISD::SHL || Opc == ISD::SRL || Opc == ISD::SRA) && - isIntImmediate(N.getOperand(1), SH) && SH < 32; -} - -// isOprNot - Returns true if the specified operand is an xor with immediate -1. -static bool isOprNot(SDOperand N) { - unsigned Imm; - return isOpcWithIntImmediate(N, ISD::XOR, Imm) && (signed)Imm == -1; -} - -// Immediate constant composers. -// Lo16 - grabs the lo 16 bits from a 32 bit constant. -// Hi16 - grabs the hi 16 bits from a 32 bit constant. -// HA16 - computes the hi bits required if the lo bits are add/subtracted in -// arithmethically. -static unsigned Lo16(unsigned x) { return x & 0x0000FFFF; } -static unsigned Hi16(unsigned x) { return Lo16(x >> 16); } -static unsigned HA16(unsigned x) { return Hi16((signed)x - (signed short)x); } - -/// NodeHasRecordingVariant - If SelectExpr can always produce code for -/// NodeOpcode that also sets CR0 as a side effect, return true. Otherwise, -/// return false. -static bool NodeHasRecordingVariant(unsigned NodeOpcode) { - switch(NodeOpcode) { - default: return false; - case ISD::AND: - case ISD::OR: - return true; - } -} - -/// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding -/// to Condition. -static unsigned getBCCForSetCC(ISD::CondCode CC) { - switch (CC) { - default: assert(0 && "Unknown condition!"); abort(); - case ISD::SETEQ: return PPC::BEQ; - case ISD::SETNE: return PPC::BNE; - case ISD::SETULT: - case ISD::SETLT: return PPC::BLT; - case ISD::SETULE: - case ISD::SETLE: return PPC::BLE; - case ISD::SETUGT: - case ISD::SETGT: return PPC::BGT; - case ISD::SETUGE: - case ISD::SETGE: return PPC::BGE; - } - return 0; -} - -/// getCRIdxForSetCC - Return the index of the condition register field -/// associated with the SetCC condition, and whether or not the field is -/// treated as inverted. That is, lt = 0; ge = 0 inverted. -static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool& Inv) { - switch (CC) { - default: assert(0 && "Unknown condition!"); abort(); - case ISD::SETULT: - case ISD::SETLT: Inv = false; return 0; - case ISD::SETUGE: - case ISD::SETGE: Inv = true; return 0; - case ISD::SETUGT: - case ISD::SETGT: Inv = false; return 1; - case ISD::SETULE: - case ISD::SETLE: Inv = true; return 1; - case ISD::SETEQ: Inv = false; return 2; - case ISD::SETNE: Inv = true; return 2; - } - return 0; -} - -/// IndexedOpForOp - Return the indexed variant for each of the PowerPC load -/// and store immediate instructions. -static unsigned IndexedOpForOp(unsigned Opcode) { - switch(Opcode) { - default: assert(0 && "Unknown opcode!"); abort(); - case PPC::LBZ: return PPC::LBZX; case PPC::STB: return PPC::STBX; - case PPC::LHZ: return PPC::LHZX; case PPC::STH: return PPC::STHX; - case PPC::LHA: return PPC::LHAX; case PPC::STW: return PPC::STWX; - case PPC::LWZ: return PPC::LWZX; case PPC::STFS: return PPC::STFSX; - case PPC::LFS: return PPC::LFSX; case PPC::STFD: return PPC::STFDX; - case PPC::LFD: return PPC::LFDX; - } - return 0; -} -} - -/// getGlobalBaseReg - Output the instructions required to put the -/// base address to use for accessing globals into a register. -/// -unsigned ISel::getGlobalBaseReg() { - if (!GlobalBaseInitialized) { - // Insert the set of GlobalBaseReg into the first MBB of the function - MachineBasicBlock &FirstMBB = BB->getParent()->front(); - MachineBasicBlock::iterator MBBI = FirstMBB.begin(); - GlobalBaseReg = MakeIntReg(); - BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR); - BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg); - GlobalBaseInitialized = true; - } - return GlobalBaseReg; -} - -/// MoveCRtoGPR - Move CCReg[Idx] to the least significant bit of Result. If -/// Inv is true, then invert the result. -void ISel::MoveCRtoGPR(unsigned CCReg, ISD::CondCode CC, unsigned Result){ - bool Inv; - unsigned IntCR = MakeIntReg(); - unsigned Idx = getCRIdxForSetCC(CC, Inv); - BuildMI(BB, PPC::MCRF, 1, PPC::CR7).addReg(CCReg); - bool GPOpt = - TLI.getTargetMachine().getSubtarget<PPCSubtarget>().isGigaProcessor(); - if (GPOpt) - BuildMI(BB, PPC::MFOCRF, 1, IntCR).addReg(PPC::CR7); - else - BuildMI(BB, PPC::MFCR, 0, IntCR); - if (Inv) { - unsigned Tmp1 = MakeIntReg(); - BuildMI(BB, PPC::RLWINM, 4, Tmp1).addReg(IntCR).addImm(32-(3-Idx)) - .addImm(31).addImm(31); - BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(1); - } else { - BuildMI(BB, PPC::RLWINM, 4, Result).addReg(IntCR).addImm(32-(3-Idx)) - .addImm(31).addImm(31); - } -} - -/// SelectBitfieldInsert - turn an or of two masked values into -/// the rotate left word immediate then mask insert (rlwimi) instruction. -/// Returns true on success, false if the caller still needs to select OR. -/// -/// Patterns matched: -/// 1. or shl, and 5. or and, and -/// 2. or and, shl 6. or shl, shr -/// 3. or shr, and 7. or shr, shl -/// 4. or and, shr -bool ISel::SelectBitfieldInsert(SDOperand OR, unsigned Result) { - bool IsRotate = false; - unsigned TgtMask = 0xFFFFFFFF, InsMask = 0xFFFFFFFF, Amount = 0; - unsigned Value; - - SDOperand Op0 = OR.getOperand(0); - SDOperand Op1 = OR.getOperand(1); - - unsigned Op0Opc = Op0.getOpcode(); - unsigned Op1Opc = Op1.getOpcode(); - - // Verify that we have the correct opcodes - if (ISD::SHL != Op0Opc && ISD::SRL != Op0Opc && ISD::AND != Op0Opc) - return false; - if (ISD::SHL != Op1Opc && ISD::SRL != Op1Opc && ISD::AND != Op1Opc) - return false; - - // Generate Mask value for Target - if (isIntImmediate(Op0.getOperand(1), Value)) { - switch(Op0Opc) { - case ISD::SHL: TgtMask <<= Value; break; - case ISD::SRL: TgtMask >>= Value; break; - case ISD::AND: TgtMask &= Value; break; - } - } else { - return false; - } - - // Generate Mask value for Insert - if (isIntImmediate(Op1.getOperand(1), Value)) { - switch(Op1Opc) { - case ISD::SHL: - Amount = Value; - InsMask <<= Amount; - if (Op0Opc == ISD::SRL) IsRotate = true; - break; - case ISD::SRL: - Amount = Value; - InsMask >>= Amount; - Amount = 32-Amount; - if (Op0Opc == ISD::SHL) IsRotate = true; - break; - case ISD::AND: - InsMask &= Value; - break; - } - } else { - return false; - } - - unsigned Tmp3 = 0; - - // If both of the inputs are ANDs and one of them has a logical shift by - // constant as its input, make that the inserted value so that we can combine - // the shift into the rotate part of the rlwimi instruction - if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) { - if (Op1.getOperand(0).getOpcode() == ISD::SHL || - Op1.getOperand(0).getOpcode() == ISD::SRL) { - if (isIntImmediate(Op1.getOperand(0).getOperand(1), Value)) { - Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ? - Value : 32 - Value; - Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0)); - } - } else if (Op0.getOperand(0).getOpcode() == ISD::SHL || - Op0.getOperand(0).getOpcode() == ISD::SRL) { - if (isIntImmediate(Op0.getOperand(0).getOperand(1), Value)) { - std::swap(Op0, Op1); - std::swap(TgtMask, InsMask); - Amount = Op1.getOperand(0).getOpcode() == ISD::SHL ? - Value : 32 - Value; - Tmp3 = SelectExpr(Op1.getOperand(0).getOperand(0)); - } - } - } - - // Verify that the Target mask and Insert mask together form a full word mask - // and that the Insert mask is a run of set bits (which implies both are runs - // of set bits). Given that, Select the arguments and generate the rlwimi - // instruction. - unsigned MB, ME; - if (((TgtMask & InsMask) == 0) && isRunOfOnes(InsMask, MB, ME)) { - unsigned Tmp1, Tmp2; - bool fullMask = (TgtMask ^ InsMask) == 0xFFFFFFFF; - // Check for rotlwi / rotrwi here, a special case of bitfield insert - // where both bitfield halves are sourced from the same value. - if (IsRotate && fullMask && - OR.getOperand(0).getOperand(0) == OR.getOperand(1).getOperand(0)) { - Tmp1 = SelectExpr(OR.getOperand(0).getOperand(0)); - BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Amount) - .addImm(0).addImm(31); - return true; - } - if (Op0Opc == ISD::AND && fullMask) - Tmp1 = SelectExpr(Op0.getOperand(0)); - else - Tmp1 = SelectExpr(Op0); - Tmp2 = Tmp3 ? Tmp3 : SelectExpr(Op1.getOperand(0)); - BuildMI(BB, PPC::RLWIMI, 5, Result).addReg(Tmp1).addReg(Tmp2) - .addImm(Amount).addImm(MB).addImm(ME); - return true; - } - return false; -} - -/// FoldIfWideZeroExtend - 32 bit PowerPC implicit masks shift amounts to the -/// low six bits. If the shift amount is an ISD::AND node with a mask that is -/// wider than the implicit mask, then we can get rid of the AND and let the -/// shift do the mask. -unsigned ISel::FoldIfWideZeroExtend(SDOperand N) { - unsigned C; - if (isOpcWithIntImmediate(N, ISD::AND, C) && isMask_32(C) && C > 63) - return SelectExpr(N.getOperand(0)); - else - return SelectExpr(N); -} - -unsigned ISel::SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC) { - unsigned Result, Tmp1, Tmp2; - bool AlreadySelected = false; - - // Allocate a condition register for this expression - Result = RegMap->createVirtualRegister(PPC::CRRCRegisterClass); - - // Use U to determine whether the SETCC immediate range is signed or not. - bool U = ISD::isUnsignedIntSetCC(CC); - if (isIntImmediate(RHS, Tmp2) && - ((U && isUInt16(Tmp2)) || (!U && isInt16(Tmp2)))) { - Tmp2 = Lo16(Tmp2); - // For comparisons against zero, we can implicity set CR0 if a recording - // variant (e.g. 'or.' instead of 'or') of the instruction that defines - // operand zero of the SetCC node is available. - if (Tmp2 == 0 && - NodeHasRecordingVariant(LHS.getOpcode()) && LHS.Val->hasOneUse()) { - RecordSuccess = false; - Tmp1 = SelectExpr(LHS, true); - if (RecordSuccess) { - ++Recorded; - BuildMI(BB, PPC::MCRF, 1, Result).addReg(PPC::CR0); - return Result; - } - AlreadySelected = true; - } - // If we could not implicitly set CR0, then emit a compare immediate - // instead. - if (!AlreadySelected) Tmp1 = SelectExpr(LHS); - if (U) - BuildMI(BB, PPC::CMPLWI, 2, Result).addReg(Tmp1).addImm(Tmp2); - else - BuildMI(BB, PPC::CMPWI, 2, Result).addReg(Tmp1).addSImm(Tmp2); - } else { - unsigned CompareOpc; - if (MVT::isInteger(LHS.getValueType())) - CompareOpc = U ? PPC::CMPLW : PPC::CMPW; - else if (LHS.getValueType() == MVT::f32) - CompareOpc = PPC::FCMPUS; - else - CompareOpc = PPC::FCMPUD; - Tmp1 = SelectExpr(LHS); - Tmp2 = SelectExpr(RHS); - BuildMI(BB, CompareOpc, 2, Result).addReg(Tmp1).addReg(Tmp2); - } - return Result; -} - -/// Check to see if the load is a constant offset from a base register. -unsigned ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset) -{ - unsigned imm = 0, opcode = N.getOpcode(); - if (N.getOpcode() == ISD::ADD) { - bool isFrame = N.getOperand(0).getOpcode() == ISD::FrameIndex; - if (isIntImmediate(N.getOperand(1), imm) && isInt16(imm)) { - offset = Lo16(imm); - if (isFrame) { - ++FrameOff; - Reg = cast<FrameIndexSDNode>(N.getOperand(0))->getIndex(); - return 1; - } else { - Reg = SelectExpr(N.getOperand(0)); - return 0; - } - } else { - Reg = SelectExpr(N.getOperand(0)); - offset = SelectExpr(N.getOperand(1)); - return 2; - } - } - // Now check if we're dealing with a global, and whether or not we should emit - // an optimized load or store for statics. - if(GlobalAddressSDNode *GN = dyn_cast<GlobalAddressSDNode>(N)) { - GlobalValue *GV = GN->getGlobal(); - if (!GV->hasWeakLinkage() && !GV->isExternal()) { - unsigned GlobalHi = MakeIntReg(); - if (PICEnabled) - BuildMI(BB, PPC::ADDIS, 2, GlobalHi).addReg(getGlobalBaseReg()) - .addGlobalAddress(GV); - else - BuildMI(BB, PPC::LIS, 1, GlobalHi).addGlobalAddress(GV); - Reg = GlobalHi; - offset = 0; - return 3; - } - } - Reg = SelectExpr(N); - offset = 0; - return 0; -} - -void ISel::SelectBranchCC(SDOperand N) -{ - MachineBasicBlock *Dest = - cast<BasicBlockSDNode>(N.getOperand(4))->getBasicBlock(); - - Select(N.getOperand(0)); //chain - ISD::CondCode CC = cast<CondCodeSDNode>(N.getOperand(1))->get(); - unsigned CCReg = SelectCC(N.getOperand(2), N.getOperand(3), CC); - unsigned Opc = getBCCForSetCC(CC); - - // If this is a two way branch, then grab the fallthrough basic block argument - // and build a PowerPC branch pseudo-op, suitable for long branch conversion - // if necessary by the branch selection pass. Otherwise, emit a standard - // conditional branch. - if (N.getOpcode() == ISD::BRTWOWAY_CC) { - MachineBasicBlock *Fallthrough = - cast<BasicBlockSDNode>(N.getOperand(5))->getBasicBlock(); - BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc) - .addMBB(Dest).addMBB(Fallthrough); - BuildMI(BB, PPC::B, 1).addMBB(Fallthrough); - } else { - // Iterate to the next basic block - ilist<MachineBasicBlock>::iterator It = BB; - ++It; - - // If the fallthrough path is off the end of the function, which would be - // undefined behavior, set it to be the same as the current block because - // we have nothing better to set it to, and leaving it alone will cause the - // PowerPC Branch Selection pass to crash. - if (It == BB->getParent()->end()) It = Dest; - BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc) - .addMBB(Dest).addMBB(It); - } - return; -} - -// SelectIntImmediateExpr - Choose code for opcodes with immediate value. -bool ISel::SelectIntImmediateExpr(SDOperand N, unsigned Result, - unsigned OCHi, unsigned OCLo, - bool IsArithmetic, bool Negate) { - // check constant - ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1)); - // exit if not a constant - if (!CN) return false; - // extract immediate - unsigned C = (unsigned)CN->getValue(); - // negate if required (ISD::SUB) - if (Negate) C = -C; - // get the hi and lo portions of constant - unsigned Hi = IsArithmetic ? HA16(C) : Hi16(C); - unsigned Lo = Lo16(C); - // assume no intermediate result from lo instruction (same as final result) - unsigned Tmp = Result; - // check if two instructions are needed - if (Hi && Lo) { - // exit if usage indicates it would be better to load immediate into a - // register - if (CN->use_size() > 2) return false; - // need intermediate result for two instructions - Tmp = MakeIntReg(); - } - // get first operand - unsigned Opr0 = SelectExpr(N.getOperand(0)); - // is a lo instruction needed - if (Lo) { - // generate instruction for lo portion - BuildMI(BB, OCLo, 2, Tmp).addReg(Opr0).addImm(Lo); - // need to switch out first operand for hi instruction - Opr0 = Tmp; - } - // is a hi instruction needed - if (Hi) { - // generate instruction for hi portion - BuildMI(BB, OCHi, 2, Result).addReg(Opr0).addImm(Hi); - } - return true; -} - -unsigned ISel::SelectExpr(SDOperand N, bool Recording) { - unsigned Result; - unsigned Tmp1, Tmp2, Tmp3; - unsigned Opc = 0; - unsigned opcode = N.getOpcode(); - - SDNode *Node = N.Val; - MVT::ValueType DestType = N.getValueType(); - - if (Node->getOpcode() == ISD::CopyFromReg) { - unsigned Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg(); - // Just use the specified register as our input. - if (MRegisterInfo::isVirtualRegister(Reg) || Reg == PPC::R1) - return Reg; - } - - unsigned &Reg = ExprMap[N]; - if (Reg) return Reg; - - switch (N.getOpcode()) { - default: - Reg = Result = (N.getValueType() != MVT::Other) ? - MakeReg(N.getValueType()) : 1; - break; - case ISD::AssertSext: - case ISD::AssertZext: - // Don't allocate a vreg for these nodes. - return Reg = SelectExpr(N.getOperand(0)); - case ISD::TAILCALL: - case ISD::CALL: - // If this is a call instruction, make sure to prepare ALL of the result - // values as well as the chain. - if (Node->getNumValues() == 1) - Reg = Result = 1; // Void call, just a chain. - else { - Result = MakeReg(Node->getValueType(0)); - ExprMap[N.getValue(0)] = Result; - for (unsigned i = 1, e = N.Val->getNumValues()-1; i != e; ++i) - ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i)); - ExprMap[SDOperand(Node, Node->getNumValues()-1)] = 1; - } - break; - case ISD::ADD_PARTS: - case ISD::SUB_PARTS: - Result = MakeReg(Node->getValueType(0)); - ExprMap[N.getValue(0)] = Result; - for (unsigned i = 1, e = N.Val->getNumValues(); i != e; ++i) - ExprMap[N.getValue(i)] = MakeReg(Node->getValueType(i)); - break; - } - - switch (opcode) { - default: - Node->dump(); std::cerr << '\n'; - assert(0 && "Node not handled!\n"); - case PPCISD::FSEL: - Tmp1 = SelectExpr(N.getOperand(0)); - Tmp2 = SelectExpr(N.getOperand(1)); - Tmp3 = SelectExpr(N.getOperand(2)); - - // Extend the comparison to 64-bits if needed. - if (N.getOperand(0).getValueType() == MVT::f32) { - unsigned Tmp1New = MakeReg(MVT::f64); - BuildMI(BB, PPC::FMRSD, 1, Tmp1New).addReg(Tmp1); - Tmp1 = Tmp1New; - } - - Opc = N.Val->getValueType(0) == MVT::f32 ? PPC::FSELS : PPC::FSELD; - BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3); - return Result; - case PPCISD::FCFID: - Tmp1 = SelectExpr(N.getOperand(0)); - BuildMI(BB, PPC::FCFID, 1, Result).addReg(Tmp1); - return Result; - case PPCISD::FCTIDZ: - Tmp1 = SelectExpr(N.getOperand(0)); - BuildMI(BB, PPC::FCTIDZ, 1, Result).addReg(Tmp1); - return Result; - case PPCISD::FCTIWZ: - Tmp1 = SelectExpr(N.getOperand(0)); - BuildMI(BB, PPC::FCTIWZ, 1, Result).addReg(Tmp1); - return Result; - case ISD::UNDEF: - if (Node->getValueType(0) == MVT::i32) - BuildMI(BB, PPC::IMPLICIT_DEF_GPR, 0, Result); - else if (Node->getValueType(0) == MVT::f32) - BuildMI(BB, PPC::IMPLICIT_DEF_F4, 0, Result); - else - BuildMI(BB, PPC::IMPLICIT_DEF_F8, 0, Result); - return Result; - case ISD::DYNAMIC_STACKALLOC: - // Generate both result values. FIXME: Need a better commment here? - if (Result != 1) - ExprMap[N.getValue(1)] = 1; - else - Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType()); - - // FIXME: We are currently ignoring the requested alignment for handling - // greater than the stack alignment. This will need to be revisited at some - // point. Align = N.getOperand(2); - if (!isa<ConstantSDNode>(N.getOperand(2)) || - cast<ConstantSDNode>(N.getOperand(2))->getValue() != 0) { - std::cerr << "Cannot allocate stack object with greater alignment than" - << " the stack alignment yet!"; - abort(); - } - Select(N.getOperand(0)); - Tmp1 = SelectExpr(N.getOperand(1)); - // Subtract size from stack pointer, thereby allocating some space. - BuildMI(BB, PPC::SUBF, 2, PPC::R1).addReg(Tmp1).addReg(PPC::R1); - // Put a pointer to the space into the result register by copying the SP - BuildMI(BB, PPC::OR4, 2, Result).addReg(PPC::R1).addReg(PPC::R1); - return Result; - - case ISD::ConstantPool: - Tmp1 = BB->getParent()->getConstantPool()-> - getConstantPoolIndex(cast<ConstantPoolSDNode>(N)->get()); - Tmp2 = MakeIntReg(); - if (PICEnabled) - BuildMI(BB, PPC::ADDIS, 2, Tmp2).addReg(getGlobalBaseReg()) - .addConstantPoolIndex(Tmp1); - else - BuildMI(BB, PPC::LIS, 1, Tmp2).addConstantPoolIndex(Tmp1); - BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1); - return Result; - - case ISD::FrameIndex: - Tmp1 = cast<FrameIndexSDNode>(N)->getIndex(); - addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false); - return Result; - - case ISD::GlobalAddress: { - GlobalValue *GV = cast<GlobalAddressSDNode>(N)->getGlobal(); - Tmp1 = MakeIntReg(); - if (PICEnabled) - BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg()) - .addGlobalAddress(GV); - else - BuildMI(BB, PPC::LIS, 1, Tmp1).addGlobalAddress(GV); - if (GV->hasWeakLinkage() || GV->isExternal()) { - BuildMI(BB, PPC::LWZ, 2, Result).addGlobalAddress(GV).addReg(Tmp1); - } else { - BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp1).addGlobalAddress(GV); - } - return Result; - } - - case ISD::LOAD: - case ISD::EXTLOAD: - case ISD::ZEXTLOAD: - case ISD::SEXTLOAD: { - MVT::ValueType TypeBeingLoaded = (ISD::LOAD == opcode) ? - Node->getValueType(0) : cast<VTSDNode>(Node->getOperand(3))->getVT(); - bool sext = (ISD::SEXTLOAD == opcode); - - // Make sure we generate both values. - if (Result != 1) - ExprMap[N.getValue(1)] = 1; // Generate the token - else - Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType()); - - SDOperand Chain = N.getOperand(0); - SDOperand Address = N.getOperand(1); - Select(Chain); - - switch (TypeBeingLoaded) { - default: Node->dump(); assert(0 && "Cannot load this type!"); - case MVT::i1: Opc = PPC::LBZ; break; - case MVT::i8: Opc = PPC::LBZ; break; - case MVT::i16: Opc = sext ? PPC::LHA : PPC::LHZ; break; - case MVT::i32: Opc = PPC::LWZ; break; - case MVT::f32: Opc = PPC::LFS; break; - case MVT::f64: Opc = PPC::LFD; break; - } - - if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Address)) { - Tmp1 = MakeIntReg(); - unsigned CPI = BB->getParent()->getConstantPool()-> - getConstantPoolIndex(CP->get()); - if (PICEnabled) - BuildMI(BB, PPC::ADDIS, 2, Tmp1).addReg(getGlobalBaseReg()) - .addConstantPoolIndex(CPI); - else - BuildMI(BB, PPC::LIS, 1, Tmp1).addConstantPoolIndex(CPI); - BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1); - } else if (Address.getOpcode() == ISD::FrameIndex) { - Tmp1 = cast<FrameIndexSDNode>(Address)->getIndex(); - addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1); - } else { - int offset; - switch(SelectAddr(Address, Tmp1, offset)) { - default: assert(0 && "Unhandled return value from SelectAddr"); - case 0: // imm offset, no frame, no index - BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1); - break; - case 1: // imm offset + frame index - addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1, offset); - break; - case 2: // base+index addressing - Opc = IndexedOpForOp(Opc); - BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset); - break; - case 3: { - GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Address); - GlobalValue *GV = GN->getGlobal(); - BuildMI(BB, Opc, 2, Result).addGlobalAddress(GV).addReg(Tmp1); - } - } - } - return Result; - } - - case ISD::TAILCALL: - case ISD::CALL: { - unsigned GPR_idx = 0, FPR_idx = 0; - static const unsigned GPR[] = { - PPC::R3, PPC::R4, PPC::R5, PPC::R6, - PPC::R7, PPC::R8, PPC::R9, PPC::R10, - }; - static const unsigned FPR[] = { - PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7, - PPC::F8, PPC::F9, PPC::F10, PPC::F11, PPC::F12, PPC::F13 - }; - - // Lower the chain for this call. - Select(N.getOperand(0)); - ExprMap[N.getValue(Node->getNumValues()-1)] = 1; - - MachineInstr *CallMI; - // Emit the correct call instruction based on the type of symbol called. - if (GlobalAddressSDNode *GASD = - dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) { - CallMI = BuildMI(PPC::BL, 1).addGlobalAddress(GASD->getGlobal(), true); - } else if (ExternalSymbolSDNode *ESSDN = - dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) { - CallMI = BuildMI(PPC::BL, 1).addExternalSymbol(ESSDN->getSymbol(), true); - } else { - Tmp1 = SelectExpr(N.getOperand(1)); - BuildMI(BB, PPC::MTCTR, 1).addReg(Tmp1); - BuildMI(BB, PPC::OR4, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1); - CallMI = BuildMI(PPC::BCTRL, 1).addReg(PPC::R12); - } - - // Load the register args to virtual regs - std::vector<unsigned> ArgVR; - for(int i = 2, e = Node->getNumOperands(); i < e; ++i) - ArgVR.push_back(SelectExpr(N.getOperand(i))); - - // Copy the virtual registers into the appropriate argument register - for(int i = 0, e = ArgVR.size(); i < e; ++i) { - switch(N.getOperand(i+2).getValueType()) { - default: Node->dump(); assert(0 && "Unknown value type for call"); - case MVT::i32: - assert(GPR_idx < 8 && "Too many int args"); - if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) { - BuildMI(BB, PPC::OR4,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]); - CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use); - } - ++GPR_idx; - break; - case MVT::f64: - case MVT::f32: - assert(FPR_idx < 13 && "Too many fp args"); - BuildMI(BB, N.getOperand(i+2).getValueType() == MVT::f32 ? PPC::FMRS : - PPC::FMRD, 1, FPR[FPR_idx]).addReg(ArgVR[i]); - CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use); - ++FPR_idx; - break; - } - } - - // Put the call instruction in the correct place in the MachineBasicBlock - BB->push_back(CallMI); - - switch (Node->getValueType(0)) { - default: assert(0 && "Unknown value type for call result!"); - case MVT::Other: return 1; - case MVT::i32: - if (Node->getValueType(1) == MVT::i32) { - BuildMI(BB, PPC::OR4, 2, Result+1).addReg(PPC::R3).addReg(PPC::R3); - BuildMI(BB, PPC::OR4, 2, Result).addReg(PPC::R4).addReg(PPC::R4); - } else { - BuildMI(BB, PPC::OR4, 2, Result).addReg(PPC::R3).addReg(PPC::R3); - } - break; - case MVT::f32: - BuildMI(BB, PPC::FMRS, 1, Result).addReg(PPC::F1); - break; - case MVT::f64: - BuildMI(BB, PPC::FMRD, 1, Result).addReg(PPC::F1); - break; - } - return Result+N.ResNo; - } - - case ISD::SIGN_EXTEND_INREG: - Tmp1 = SelectExpr(N.getOperand(0)); - switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) { - default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break; - case MVT::i16: - BuildMI(BB, PPC::EXTSH, 1, Result).addReg(Tmp1); - break; - case MVT::i8: - BuildMI(BB, PPC::EXTSB, 1, Result).addReg(Tmp1); - break; - } - return Result; - - case ISD::CopyFromReg: - DestType = N.getValue(0).getValueType(); - if (Result == 1) - Result = ExprMap[N.getValue(0)] = MakeReg(DestType); - else - ExprMap[N.getValue(1)] = 1; - Tmp1 = dyn_cast<RegisterSDNode>(Node->getOperand(1))->getReg(); - if (MVT::isInteger(DestType)) - BuildMI(BB, PPC::OR4, 2, Result).addReg(Tmp1).addReg(Tmp1); - else if (DestType == MVT::f32) - BuildMI(BB, PPC::FMRS, 1, Result).addReg(Tmp1); - else - BuildMI(BB, PPC::FMRD, 1, Result).addReg(Tmp1); - return Result; - - case ISD::SHL: - if (isIntImmediate(N.getOperand(1), Tmp2)) { - unsigned SH, MB, ME; - if (isOpcWithIntImmediate(N.getOperand(0), ISD::AND, Tmp3) && - isRotateAndMask(ISD::SHL, Tmp2, Tmp3, true, SH, MB, ME)) { - Tmp1 = SelectExpr(N.getOperand(0).getOperand(0)); - BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(SH) |