From: Nate Begeman Date: Tue, 5 Apr 2005 08:51:15 +0000 (+0000) Subject: Remove 64 bit simple ISel, it never worked correctly X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=d3e6b94020ac6ed827fb1dfe1f4abe9ff39e4ec7;p=oota-llvm.git Remove 64 bit simple ISel, it never worked correctly Add initial (buggy) implementation of 64 bit pattern ISel git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@21096 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/Target/PowerPC/PPC.h b/lib/Target/PowerPC/PPC.h index bd1f2ac8ce5..ddda03c0f39 100644 --- a/lib/Target/PowerPC/PPC.h +++ b/lib/Target/PowerPC/PPC.h @@ -25,7 +25,7 @@ class TargetMachine; FunctionPass *createPPCBranchSelectionPass(); FunctionPass *createPPC32ISelSimple(TargetMachine &TM); FunctionPass *createPPC32ISelPattern(TargetMachine &TM); -FunctionPass *createPPC64ISelSimple(TargetMachine &TM); +FunctionPass *createPPC64ISelPattern(TargetMachine &TM); FunctionPass *createDarwinAsmPrinter(std::ostream &OS, TargetMachine &TM); FunctionPass *createAIXAsmPrinter(std::ostream &OS, TargetMachine &TM); diff --git a/lib/Target/PowerPC/PPC64ISelPattern.cpp b/lib/Target/PowerPC/PPC64ISelPattern.cpp new file mode 100644 index 00000000000..210acf580b4 --- /dev/null +++ b/lib/Target/PowerPC/PPC64ISelPattern.cpp @@ -0,0 +1,1767 @@ +//===-- 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. +// +//===----------------------------------------------------------------------===// + +#include "PowerPC.h" +#include "PowerPCInstrBuilder.h" +#include "PowerPCInstrInfo.h" +#include "PPC64RegisterInfo.h" +#include "llvm/Constants.h" // FIXME: REMOVE +#include "llvm/Function.h" +#include "llvm/CodeGen/MachineConstantPool.h" // FIXME: REMOVE +#include "llvm/CodeGen/MachineFunction.h" +#include "llvm/CodeGen/MachineFrameInfo.h" +#include "llvm/CodeGen/SelectionDAG.h" +#include "llvm/CodeGen/SelectionDAGISel.h" +#include "llvm/CodeGen/SSARegMap.h" +#include "llvm/Target/TargetData.h" +#include "llvm/Target/TargetLowering.h" +#include "llvm/Target/TargetOptions.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/MathExtras.h" +#include "llvm/ADT/Statistic.h" +#include +#include +using namespace llvm; + +//===----------------------------------------------------------------------===// +// PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface +namespace { + class PPC64TargetLowering : public TargetLowering { + int VarArgsFrameIndex; // FrameIndex for start of varargs area. + int ReturnAddrIndex; // FrameIndex for return slot. + public: + PPC64TargetLowering(TargetMachine &TM) : TargetLowering(TM) { + // Set up the register classes. + addRegisterClass(MVT::i64, PPC64::GPRCRegisterClass); + addRegisterClass(MVT::f32, PPC64::FPRCRegisterClass); + addRegisterClass(MVT::f64, PPC64::FPRCRegisterClass); + + // PowerPC has no intrinsics for these particular operations + setOperationAction(ISD::MEMMOVE, MVT::Other, Expand); + setOperationAction(ISD::MEMSET, MVT::Other, Expand); + setOperationAction(ISD::MEMCPY, MVT::Other, Expand); + + // PPC 64 has i16 and i32 but no i8 (or i1) SEXTLOAD + setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand); + setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand); + + setShiftAmountFlavor(Extend); // shl X, 32 == 0 + addLegalFPImmediate(+0.0); // Necessary for FSEL + addLegalFPImmediate(-0.0); // + + computeRegisterProperties(); + } + + /// LowerArguments - This hook must be implemented to indicate how we should + /// lower the arguments for the specified function, into the specified DAG. + virtual std::vector + LowerArguments(Function &F, SelectionDAG &DAG); + + /// LowerCallTo - This hook lowers an abstract call to a function into an + /// actual call. + virtual std::pair + LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, + SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG); + + virtual std::pair + LowerVAStart(SDOperand Chain, SelectionDAG &DAG); + + virtual std::pair + LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList, + const Type *ArgTy, SelectionDAG &DAG); + + virtual std::pair + LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth, + SelectionDAG &DAG); + }; +} + + +std::vector +PPC64TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) { + // + // add beautiful description of PPC stack frame format, or at least some docs + // + MachineFunction &MF = DAG.getMachineFunction(); + MachineFrameInfo *MFI = MF.getFrameInfo(); + MachineBasicBlock& BB = MF.front(); + std::vector ArgValues; + + // Due to the rather complicated nature of the PowerPC ABI, rather than a + // fixed size array of physical args, for the sake of simplicity let the STL + // handle tracking them for us. + std::vector argVR, argPR, argOp; + unsigned ArgOffset = 24; + unsigned GPR_remaining = 8; + unsigned FPR_remaining = 13; + 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 + }; + + // Add DAG nodes to load the arguments... On entry to a function on PPC, + // the arguments start at offset 24, although they are likely to be passed + // in registers. + for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) { + SDOperand newroot, argt; + unsigned ObjSize; + bool needsLoad = false; + MVT::ValueType ObjectVT = getValueType(I->getType()); + + switch (ObjectVT) { + default: assert(0 && "Unhandled argument type!"); + case MVT::i1: + case MVT::i8: + case MVT::i16: + case MVT::i32: + ObjSize = 4; + if (GPR_remaining > 0) { + BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]); + argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, + DAG.getRoot()); + if (ObjectVT != MVT::i32) + argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot); + } else { + needsLoad = true; + } + break; + case MVT::i64: ObjSize = 8; + // FIXME: can split 64b load between reg/mem if it is last arg in regs + if (GPR_remaining > 1) { + BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]); + BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx+1]); + // Copy the extracted halves into the virtual registers + SDOperand argHi = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, + DAG.getRoot()); + SDOperand argLo = DAG.getCopyFromReg(GPR[GPR_idx+1], MVT::i32, argHi); + // Build the outgoing arg thingy + argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi); + newroot = argLo; + } else { + needsLoad = true; + } + break; + case MVT::f32: ObjSize = 4; + case MVT::f64: ObjSize = 8; + if (FPR_remaining > 0) { + BuildMI(&BB, PPC::IMPLICIT_DEF, 0, FPR[FPR_idx]); + argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT, + DAG.getRoot()); + --FPR_remaining; + ++FPR_idx; + } else { + needsLoad = true; + } + break; + } + + // We need to load the argument to a virtual register if we determined above + // that we ran out of physical registers of the appropriate type + if (needsLoad) { + unsigned SubregOffset = 0; + if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3; + if (ObjectVT == MVT::i16) SubregOffset = 2; + int FI = MFI->CreateFixedObject(ObjSize, ArgOffset); + SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32); + FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, + DAG.getConstant(SubregOffset, MVT::i32)); + argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN); + } + + // Every 4 bytes of argument space consumes one of the GPRs available for + // argument passing. + if (GPR_remaining > 0) { + unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1; + GPR_remaining -= delta; + GPR_idx += delta; + } + ArgOffset += ObjSize; + + DAG.setRoot(newroot.getValue(1)); + ArgValues.push_back(argt); + } + + // If the function takes variable number of arguments, make a frame index for + // the start of the first vararg value... for expansion of llvm.va_start. + if (F.isVarArg()) { + VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset); + SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32); + // If this function is vararg, store any remaining integer argument regs + // to their spots on the stack so that they may be loaded by deferencing the + // result of va_next. + std::vector MemOps; + for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) { + BuildMI(&BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]); + SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot()); + SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1), + Val, FIN); + MemOps.push_back(Store); + // Increment the address by four for the next argument to store + SDOperand PtrOff = DAG.getConstant(4, getPointerTy()); + FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff); + } + DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps)); + } + + return ArgValues; +} + +std::pair +PPC64TargetLowering::LowerCallTo(SDOperand Chain, + const Type *RetTy, bool isVarArg, + SDOperand Callee, ArgListTy &Args, SelectionDAG &DAG) { + // args_to_use will accumulate outgoing args for the ISD::CALL case in + // SelectExpr to use to put the arguments in the appropriate registers. + std::vector args_to_use; + + // Count how many bytes are to be pushed on the stack, including the linkage + // area, and parameter passing area. + unsigned NumBytes = 24; + + if (Args.empty()) { + Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain, + DAG.getConstant(NumBytes, getPointerTy())); + } else { + for (unsigned i = 0, e = Args.size(); i != e; ++i) + switch (getValueType(Args[i].second)) { + default: assert(0 && "Unknown value type!"); + case MVT::i1: + case MVT::i8: + case MVT::i16: + case MVT::i32: + case MVT::f32: + NumBytes += 4; + break; + case MVT::i64: + case MVT::f64: + NumBytes += 8; + break; + } + + // Just to be safe, we'll always reserve the full 24 bytes of linkage area + // plus 32 bytes of argument space in case any called code gets funky on us. + if (NumBytes < 56) NumBytes = 56; + + // Adjust the stack pointer for the new arguments... + // These operations are automatically eliminated by the prolog/epilog pass + Chain = DAG.getNode(ISD::ADJCALLSTACKDOWN, MVT::Other, Chain, + DAG.getConstant(NumBytes, getPointerTy())); + + // Set up a copy of the stack pointer for use loading and storing any + // arguments that may not fit in the registers available for argument + // passing. + SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32, + DAG.getEntryNode()); + + // Figure out which arguments are going to go in registers, and which in + // memory. Also, if this is a vararg function, floating point operations + // must be stored to our stack, and loaded into integer regs as well, if + // any integer regs are available for argument passing. + unsigned ArgOffset = 24; + unsigned GPR_remaining = 8; + unsigned FPR_remaining = 13; + + std::vector MemOps; + for (unsigned i = 0, e = Args.size(); i != e; ++i) { + // PtrOff will be used to store the current argument to the stack if a + // register cannot be found for it. + SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy()); + PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff); + MVT::ValueType ArgVT = getValueType(Args[i].second); + + switch (ArgVT) { + default: assert(0 && "Unexpected ValueType for argument!"); + case MVT::i1: + case MVT::i8: + case MVT::i16: + // Promote the integer to 32 bits. If the input type is signed use a + // sign extend, otherwise use a zero extend. + if (Args[i].second->isSigned()) + Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first); + else + Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first); + // FALL THROUGH + case MVT::i32: + if (GPR_remaining > 0) { + args_to_use.push_back(Args[i].first); + --GPR_remaining; + } else { + MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain, + Args[i].first, PtrOff)); + } + ArgOffset += 4; + break; + case MVT::i64: + // If we have one free GPR left, we can place the upper half of the i64 + // in it, and store the other half to the stack. If we have two or more + // free GPRs, then we can pass both halves of the i64 in registers. + if (GPR_remaining > 0) { + SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, + Args[i].first, DAG.getConstant(1, MVT::i32)); + SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, + Args[i].first, DAG.getConstant(0, MVT::i32)); + args_to_use.push_back(Hi); + --GPR_remaining; + if (GPR_remaining > 0) { + args_to_use.push_back(Lo); + --GPR_remaining; + } else { + SDOperand ConstFour = DAG.getConstant(4, getPointerTy()); + PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour); + MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain, + Lo, PtrOff)); + } + } else { + MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain, + Args[i].first, PtrOff)); + } + ArgOffset += 8; + break; + case MVT::f32: + case MVT::f64: + if (FPR_remaining > 0) { + args_to_use.push_back(Args[i].first); + --FPR_remaining; + if (isVarArg) { + SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain, + Args[i].first, PtrOff); + MemOps.push_back(Store); + // Float varargs are always shadowed in available integer registers + if (GPR_remaining > 0) { + SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff); + MemOps.push_back(Load); + args_to_use.push_back(Load); + --GPR_remaining; + } + if (GPR_remaining > 0 && MVT::f64 == ArgVT) { + SDOperand ConstFour = DAG.getConstant(4, getPointerTy()); + PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour); + SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff); + MemOps.push_back(Load); + args_to_use.push_back(Load); + --GPR_remaining; + } + } else { + // If we have any FPRs remaining, we may also have GPRs remaining. + // Args passed in FPRs consume either 1 (f32) or 2 (f64) available + // GPRs. + if (GPR_remaining > 0) { + args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32)); + --GPR_remaining; + } + if (GPR_remaining > 0 && MVT::f64 == ArgVT) { + args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32)); + --GPR_remaining; + } + } + } else { + MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain, + Args[i].first, PtrOff)); + } + ArgOffset += (ArgVT == MVT::f32) ? 4 : 8; + break; + } + } + if (!MemOps.empty()) + Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps); + } + + std::vector RetVals; + MVT::ValueType RetTyVT = getValueType(RetTy); + if (RetTyVT != MVT::isVoid) + RetVals.push_back(RetTyVT); + RetVals.push_back(MVT::Other); + + SDOperand TheCall = SDOperand(DAG.getCall(RetVals, + Chain, Callee, args_to_use), 0); + Chain = TheCall.getValue(RetTyVT != MVT::isVoid); + Chain = DAG.getNode(ISD::ADJCALLSTACKUP, MVT::Other, Chain, + DAG.getConstant(NumBytes, getPointerTy())); + return std::make_pair(TheCall, Chain); +} + +std::pair +PPC64TargetLowering::LowerVAStart(SDOperand Chain, SelectionDAG &DAG) { + //vastart just returns the address of the VarArgsFrameIndex slot. + return std::make_pair(DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32), Chain); +} + +std::pair PPC64TargetLowering:: +LowerVAArgNext(bool isVANext, SDOperand Chain, SDOperand VAList, + const Type *ArgTy, SelectionDAG &DAG) { + MVT::ValueType ArgVT = getValueType(ArgTy); + SDOperand Result; + if (!isVANext) { + Result = DAG.getLoad(ArgVT, DAG.getEntryNode(), VAList); + } else { + unsigned Amt; + if (ArgVT == MVT::i32 || ArgVT == MVT::f32) + Amt = 4; + else { + assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) && + "Other types should have been promoted for varargs!"); + Amt = 8; + } + Result = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList, + DAG.getConstant(Amt, VAList.getValueType())); + } + return std::make_pair(Result, Chain); +} + + +std::pair PPC64TargetLowering:: +LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth, + SelectionDAG &DAG) { + assert(0 && "LowerFrameReturnAddress unimplemented"); + abort(); +} + +namespace { +Statistic<>NotLogic("ppc-codegen", "Number of inverted logical ops"); +Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations"); +//===--------------------------------------------------------------------===// +/// ISel - PPC32 specific code to select PPC32 machine instructions for +/// SelectionDAG operations. +//===--------------------------------------------------------------------===// +class ISel : public SelectionDAGISel { + + /// Comment Here. + PPC64TargetLowering PPC64Lowering; + + /// 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 ExprMap; + + unsigned GlobalBaseReg; + bool GlobalBaseInitialized; + +public: + ISel(TargetMachine &TM) : SelectionDAGISel(PPC64Lowering), PPC64Lowering(TM) + {} + + /// 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. + Select(DAG.getRoot()); + + // Clear state used for selection. + ExprMap.clear(); + } + + unsigned getGlobalBaseReg(); + unsigned getConstDouble(double floatVal, unsigned Result); + unsigned SelectSetCR0(SDOperand CC); + unsigned SelectExpr(SDOperand N); + unsigned SelectExprFP(SDOperand N, unsigned Result); + void Select(SDOperand N); + + bool SelectAddr(SDOperand N, unsigned& Reg, int& offset); + void SelectBranchCC(SDOperand N); +}; + +/// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It +/// returns zero when the input is not exactly a power of two. +static unsigned ExactLog2(unsigned Val) { + if (Val == 0 || (Val & (Val-1))) return 0; + unsigned Count = 0; + while (Val != 1) { + Val >>= 1; + ++Count; + } + return Count; +} + +/// getImmediateForOpcode - This method returns a value indicating whether +/// the ConstantSDNode N can be used as an immediate to Opcode. The return +/// values are either 0, 1 or 2. 0 indicates that either N is not a +/// ConstantSDNode, or is not suitable for use by that opcode. A return value +/// of 1 indicates that the constant may be used in normal immediate form. A +/// return value of 2 indicates that the constant may be used in shifted +/// immediate form. A return value of 3 indicates that log base 2 of the +/// constant may be used. +/// +static unsigned getImmediateForOpcode(SDOperand N, unsigned Opcode, + unsigned& Imm, bool U = false) { + if (N.getOpcode() != ISD::Constant) return 0; + + int v = (int)cast(N)->getSignExtended(); + + switch(Opcode) { + default: return 0; + case ISD::ADD: + if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; } + if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; } + break; + case ISD::AND: + case ISD::XOR: + case ISD::OR: + if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; } + if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; } + break; + case ISD::MUL: + case ISD::SUB: + if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; } + break; + case ISD::SETCC: + if (U && (v >= 0 && v <= 65535)) { Imm = v & 0xFFFF; return 1; } + if (!U && (v <= 32767 && v >= -32768)) { Imm = v & 0xFFFF; return 1; } + break; + case ISD::SDIV: + if ((Imm = ExactLog2(v))) { return 3; } + break; + } + return 0; +} + +/// getBCCForSetCC - Returns the PowerPC condition branch mnemonic corresponding +/// to Condition. If the Condition is unordered or unsigned, the bool argument +/// U is set to true, otherwise it is set to false. +static unsigned getBCCForSetCC(unsigned Condition, bool& U) { + U = false; + switch (Condition) { + default: assert(0 && "Unknown condition!"); abort(); + case ISD::SETEQ: return PPC::BEQ; + case ISD::SETNE: return PPC::BNE; + case ISD::SETULT: U = true; + case ISD::SETLT: return PPC::BLT; + case ISD::SETULE: U = true; + case ISD::SETLE: return PPC::BLE; + case ISD::SETUGT: U = true; + case ISD::SETGT: return PPC::BGT; + case ISD::SETUGE: U = true; + case ISD::SETGE: return PPC::BGE; + } + 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::STD: return PPC::STDX; + case PPC::LD: return PPC::LDX; 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 = MakeReg(MVT::i64); + BuildMI(FirstMBB, MBBI, PPC::MovePCtoLR, 0, PPC::LR); + BuildMI(FirstMBB, MBBI, PPC::MFLR, 1, GlobalBaseReg).addReg(PPC::LR); + GlobalBaseInitialized = true; + } + return GlobalBaseReg; +} + +/// getConstDouble - Loads a floating point value into a register, via the +/// Constant Pool. Optionally takes a register in which to load the value. +unsigned ISel::getConstDouble(double doubleVal, unsigned Result=0) { + unsigned Tmp1 = MakeReg(MVT::i32); + if (0 == Result) Result = MakeReg(MVT::f64); + MachineConstantPool *CP = BB->getParent()->getConstantPool(); + ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, doubleVal); + unsigned CPI = CP->getConstantPoolIndex(CFP); + BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg()) + .addConstantPoolIndex(CPI); + BuildMI(BB, PPC::LFD, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1); + return Result; +} + +unsigned ISel::SelectSetCR0(SDOperand CC) { + unsigned Opc, Tmp1, Tmp2; + static const unsigned CompareOpcodes[] = + { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW }; + + // If the first operand to the select is a SETCC node, then we can fold it + // into the branch that selects which value to return. + SetCCSDNode* SetCC = dyn_cast(CC.Val); + if (SetCC && CC.getOpcode() == ISD::SETCC) { + bool U; + Opc = getBCCForSetCC(SetCC->getCondition(), U); + Tmp1 = SelectExpr(SetCC->getOperand(0)); + + // Pass the optional argument U to getImmediateForOpcode for SETCC, + // so that it knows whether the SETCC immediate range is signed or not. + if (1 == getImmediateForOpcode(SetCC->getOperand(1), ISD::SETCC, + Tmp2, U)) { + if (U) + BuildMI(BB, PPC::CMPLWI, 2, PPC::CR0).addReg(Tmp1).addImm(Tmp2); + else + BuildMI(BB, PPC::CMPWI, 2, PPC::CR0).addReg(Tmp1).addSImm(Tmp2); + } else { + bool IsInteger = MVT::isInteger(SetCC->getOperand(0).getValueType()); + unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U]; + Tmp2 = SelectExpr(SetCC->getOperand(1)); + BuildMI(BB, CompareOpc, 2, PPC::CR0).addReg(Tmp1).addReg(Tmp2); + } + } else { + Tmp1 = SelectExpr(CC); + BuildMI(BB, PPC::CMPLWI, 2, PPC::CR0).addReg(Tmp1).addImm(0); + Opc = PPC::BNE; + } + return Opc; +} + +/// Check to see if the load is a constant offset from a base register +bool ISel::SelectAddr(SDOperand N, unsigned& Reg, int& offset) +{ + unsigned imm = 0, opcode = N.getOpcode(); + if (N.getOpcode() == ISD::ADD) { + Reg = SelectExpr(N.getOperand(0)); + if (1 == getImmediateForOpcode(N.getOperand(1), opcode, imm)) { + offset = imm; + return false; + } + offset = SelectExpr(N.getOperand(1)); + return true; + } + Reg = SelectExpr(N); + offset = 0; + return false; +} + +void ISel::SelectBranchCC(SDOperand N) +{ + assert(N.getOpcode() == ISD::BRCOND && "Not a BranchCC???"); + MachineBasicBlock *Dest = + cast(N.getOperand(2))->getBasicBlock(); + + // Get the MBB we will fall through to so that we can hand it off to the + // branch selection pass as an argument to the PPC::COND_BRANCH pseudo op. + //ilist::iterator It = BB; + //MachineBasicBlock *Fallthrough = ++It; + + Select(N.getOperand(0)); //chain + unsigned Opc = SelectSetCR0(N.getOperand(1)); + // FIXME: Use this once we have something approximating two-way branches + // We cannot currently use this in case the ISel hands us something like + // BRcc MBBx + // BR MBBy + // since the fallthrough basic block for the conditional branch does not start + // with the unconditional branch (it is skipped over). + //BuildMI(BB, PPC::COND_BRANCH, 4).addReg(PPC::CR0).addImm(Opc) + // .addMBB(Dest).addMBB(Fallthrough); + BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(Dest); + return; +} + +unsigned ISel::SelectExprFP(SDOperand N, unsigned Result) +{ + unsigned Tmp1, Tmp2, Tmp3; + unsigned Opc = 0; + SDNode *Node = N.Val; + MVT::ValueType DestType = N.getValueType(); + unsigned opcode = N.getOpcode(); + + switch (opcode) { + default: + Node->dump(); + assert(0 && "Node not handled!\n"); + + case ISD::SELECT: { + // Attempt to generate FSEL. We can do this whenever we have an FP result, + // and an FP comparison in the SetCC node. + SetCCSDNode* SetCC = dyn_cast(N.getOperand(0).Val); + if (SetCC && N.getOperand(0).getOpcode() == ISD::SETCC && + !MVT::isInteger(SetCC->getOperand(0).getValueType()) && + SetCC->getCondition() != ISD::SETEQ && + SetCC->getCondition() != ISD::SETNE) { + MVT::ValueType VT = SetCC->getOperand(0).getValueType(); + Tmp1 = SelectExpr(SetCC->getOperand(0)); // Val to compare against + unsigned TV = SelectExpr(N.getOperand(1)); // Use if TRUE + unsigned FV = SelectExpr(N.getOperand(2)); // Use if FALSE + + ConstantFPSDNode *CN = dyn_cast(SetCC->getOperand(1)); + if (CN && (CN->isExactlyValue(-0.0) || CN->isExactlyValue(0.0))) { + switch(SetCC->getCondition()) { + default: assert(0 && "Invalid FSEL condition"); abort(); + case ISD::SETULT: + case ISD::SETLT: + BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(FV).addReg(TV); + return Result; + case ISD::SETUGE: + case ISD::SETGE: + BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(TV).addReg(FV); + return Result; + case ISD::SETUGT: + case ISD::SETGT: { + Tmp2 = MakeReg(VT); + BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1); + BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(FV).addReg(TV); + return Result; + } + case ISD::SETULE: + case ISD::SETLE: { + Tmp2 = MakeReg(VT); + BuildMI(BB, PPC::FNEG, 1, Tmp2).addReg(Tmp1); + BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp2).addReg(TV).addReg(FV); + return Result; + } + } + } else { + Opc = (MVT::f64 == VT) ? PPC::FSUB : PPC::FSUBS; + Tmp2 = SelectExpr(SetCC->getOperand(1)); + Tmp3 = MakeReg(VT); + switch(SetCC->getCondition()) { + default: assert(0 && "Invalid FSEL condition"); abort(); + case ISD::SETULT: + case ISD::SETLT: + BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2); + BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV); + return Result; + case ISD::SETUGE: + case ISD::SETGE: + BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2); + BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV); + return Result; + case ISD::SETUGT: + case ISD::SETGT: + BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1); + BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(FV).addReg(TV); + return Result; + case ISD::SETULE: + case ISD::SETLE: + BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp2).addReg(Tmp1); + BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp3).addReg(TV).addReg(FV); + return Result; + } + } + assert(0 && "Should never get here"); + return 0; + } + + unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE + unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE + Opc = SelectSetCR0(N.getOperand(0)); + + // Create an iterator with which to insert the MBB for copying the false + // value and the MBB to hold the PHI instruction for this SetCC. + MachineBasicBlock *thisMBB = BB; + const BasicBlock *LLVM_BB = BB->getBasicBlock(); + ilist::iterator It = BB; + ++It; + + // thisMBB: + // ... + // TrueVal = ... + // cmpTY cr0, r1, r2 + // bCC copy1MBB + // fallthrough --> copy0MBB + MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB); + MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB); + BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB); + MachineFunction *F = BB->getParent(); + F->getBasicBlockList().insert(It, copy0MBB); + F->getBasicBlockList().insert(It, sinkMBB); + // Update machine-CFG edges + BB->addSuccessor(copy0MBB); + BB->addSuccessor(sinkMBB); + + // copy0MBB: + // %FalseValue = ... + // # fallthrough to sinkMBB + BB = copy0MBB; + // Update machine-CFG edges + BB->addSuccessor(sinkMBB); + + // sinkMBB: + // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] + // ... + BB = sinkMBB; + BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue) + .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB); + return Result; + } + + case ISD::FNEG: + if (!NoExcessFPPrecision && + ISD::ADD == N.getOperand(0).getOpcode() && + N.getOperand(0).Val->hasOneUse() && + ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() && + N.getOperand(0).getOperand(0).Val->hasOneUse()) { + ++FusedFP; // Statistic + Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1)); + Tmp3 = SelectExpr(N.getOperand(0).getOperand(1)); + Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS; + BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3); + } else if (!NoExcessFPPrecision && + ISD::SUB == N.getOperand(0).getOpcode() && + N.getOperand(0).Val->hasOneUse() && + ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() && + N.getOperand(0).getOperand(0).Val->hasOneUse()) { + ++FusedFP; // Statistic + Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1)); + Tmp3 = SelectExpr(N.getOperand(0).getOperand(1)); + Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS; + BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3); + } else if (ISD::FABS == N.getOperand(0).getOpcode()) { + Tmp1 = SelectExpr(N.getOperand(0).getOperand(0)); + BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1); + } else { + Tmp1 = SelectExpr(N.getOperand(0)); + BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1); + } + return Result; + + case ISD::FABS: + Tmp1 = SelectExpr(N.getOperand(0)); + BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1); + return Result; + + case ISD::FP_ROUND: + assert (DestType == MVT::f32 && + N.getOperand(0).getValueType() == MVT::f64 && + "only f64 to f32 conversion supported here"); + Tmp1 = SelectExpr(N.getOperand(0)); + BuildMI(BB, PPC::FRSP, 1, Result).addReg(Tmp1); + return Result; + + case ISD::FP_EXTEND: + assert (DestType == MVT::f64 && + N.getOperand(0).getValueType() == MVT::f32 && + "only f32 to f64 conversion supported here"); + Tmp1 = SelectExpr(N.getOperand(0)); + BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1); + return Result; + + case ISD::CopyFromReg: + if (Result == 1) + Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType()); + Tmp1 = dyn_cast(Node)->getReg(); + BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1); + return Result; + + case ISD::ConstantFP: { + ConstantFPSDNode *CN = cast(N); + Result = getConstDouble(CN->getValue(), Result); + return Result; + } + + case ISD::ADD: + if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL && + N.getOperand(0).Val->hasOneUse()) { + ++FusedFP; // Statistic + Tmp1 = SelectExpr(N.getOperand(0).getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(0).getOperand(1)); + Tmp3 = SelectExpr(N.getOperand(1)); + Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS; + BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3); + return Result; + } + Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS; + Tmp1 = SelectExpr(N.getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2); + return Result; + + case ISD::SUB: + if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL && + N.getOperand(0).Val->hasOneUse()) { + ++FusedFP; // Statistic + Tmp1 = SelectExpr(N.getOperand(0).getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(0).getOperand(1)); + Tmp3 = SelectExpr(N.getOperand(1)); + Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS; + BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3); + return Result; + } + Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS; + Tmp1 = SelectExpr(N.getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2); + return Result; + + case ISD::MUL: + case ISD::SDIV: + switch( opcode ) { + case ISD::MUL: Opc = DestType == MVT::f64 ? PPC::FMUL : PPC::FMULS; break; + case ISD::SDIV: Opc = DestType == MVT::f64 ? PPC::FDIV : PPC::FDIVS; break; + }; + Tmp1 = SelectExpr(N.getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2); + return Result; + + case ISD::UINT_TO_FP: + case ISD::SINT_TO_FP: { + bool IsUnsigned = (ISD::UINT_TO_FP == opcode); + Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register + Tmp2 = MakeReg(MVT::f64); // temp reg to load the integer value into + Tmp3 = MakeReg(MVT::i64); // temp reg to hold the conversion constant + unsigned ConstF = MakeReg(MVT::f64); // temp reg to hold the fp constant + + int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8); + MachineConstantPool *CP = BB->getParent()->getConstantPool(); + + // FIXME: pull this FP constant generation stuff out into something like + // the simple ISel's getReg. + if (IsUnsigned) { + addFrameReference(BuildMI(BB, PPC::STD, 3).addReg(Tmp1), FrameIdx); + addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx); + BuildMI(BB, PPC::FCFID, 1, Result).addReg(Tmp2); + } else { + ConstantFP *CFP = ConstantFP::get(Type::DoubleTy, 0x1.000008p52); + unsigned CPI = CP->getConstantPoolIndex(CFP); + // Load constant fp value + unsigned Tmp4 = MakeReg(MVT::i32); + unsigned TmpL = MakeReg(MVT::i32); + BuildMI(BB, PPC::LOADHiAddr, 2, Tmp4).addReg(getGlobalBaseReg()) + .addConstantPoolIndex(CPI); + BuildMI(BB, PPC::LFD, 2, ConstF).addConstantPoolIndex(CPI).addReg(Tmp4); + // Store the hi & low halves of the fp value, currently in int regs + BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330); + addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx); + BuildMI(BB, PPC::XORIS, 2, TmpL).addReg(Tmp1).addImm(0x8000); + addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(TmpL), FrameIdx, 4); + addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx); + // Generate the return value with a subtract + BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF); + } + return Result; + } + } + assert(0 && "Should never get here"); + return 0; +} + +unsigned ISel::SelectExpr(SDOperand N) { + unsigned Result; + unsigned Tmp1, Tmp2, Tmp3; + unsigned Opc = 0; + unsigned opcode = N.getOpcode(); + + SDNode *Node = N.Val; + MVT::ValueType DestType = N.getValueType(); + + 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::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; + } + + if (ISD::CopyFromReg == opcode) + DestType = N.getValue(0).getValueType(); + + if (DestType == MVT::f64 || DestType == MVT::f32) + if (ISD::LOAD != opcode && ISD::EXTLOAD != opcode && ISD::UNDEF != opcode) + return SelectExprFP(N, Result); + + switch (opcode) { + default: + Node->dump(); + assert(0 && "Node not handled!\n"); + case ISD::UNDEF: + BuildMI(BB, PPC::IMPLICIT_DEF, 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(N.getOperand(2)) || + cast(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::OR, 2, Result).addReg(PPC::R1).addReg(PPC::R1); + return Result; + + case ISD::ConstantPool: + Tmp1 = cast(N)->getIndex(); + Tmp2 = MakeReg(MVT::i64); + BuildMI(BB, PPC::LOADHiAddr, 2, Tmp2).addReg(getGlobalBaseReg()) + .addConstantPoolIndex(Tmp1); + BuildMI(BB, PPC::LA, 2, Result).addReg(Tmp2).addConstantPoolIndex(Tmp1); + return Result; + + case ISD::FrameIndex: + Tmp1 = cast(N)->getIndex(); + addFrameReference(BuildMI(BB, PPC::ADDI, 2, Result), (int)Tmp1, 0, false); + return Result; + + case ISD::GlobalAddress: { + GlobalValue *GV = cast(N)->getGlobal(); + Tmp1 = MakeReg(MVT::i64); + BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg()) + .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(Node)->getExtraValueType(); + bool sext = (ISD::SEXTLOAD == opcode); + bool byte = (MVT::i8 == TypeBeingLoaded); + + // 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 = sext ? PPC::LWA : PPC::LWZ; break; + case MVT::i64: Opc = PPC::LD; break; + case MVT::f32: Opc = PPC::LFS; break; + case MVT::f64: Opc = PPC::LFD; break; + } + + if (ConstantPoolSDNode *CP = dyn_cast(Address)) { + Tmp1 = MakeReg(MVT::i64); + int CPI = CP->getIndex(); + BuildMI(BB, PPC::LOADHiAddr, 2, Tmp1).addReg(getGlobalBaseReg()) + .addConstantPoolIndex(CPI); + BuildMI(BB, Opc, 2, Result).addConstantPoolIndex(CPI).addReg(Tmp1); + } + else if(Address.getOpcode() == ISD::FrameIndex) { + Tmp1 = cast(Address)->getIndex(); + addFrameReference(BuildMI(BB, Opc, 2, Result), (int)Tmp1); + } else { + int offset; + bool idx = SelectAddr(Address, Tmp1, offset); + if (idx) { + Opc = IndexedOpForOp(Opc); + BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(offset); + } else { + BuildMI(BB, Opc, 2, Result).addSImm(offset).addReg(Tmp1); + } + } + return Result; + } + + 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(N.getOperand(1))) { + CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(), + true); + } else if (ExternalSymbolSDNode *ESSDN = + dyn_cast(N.getOperand(1))) { + CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(), + true); + } else { + Tmp1 = SelectExpr(N.getOperand(1)); + BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1); + BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12); + CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0) + .addReg(PPC::R12); + } + + // Load the register args to virtual regs + std::vector 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::i1: + case MVT::i8: + case MVT::i16: + case MVT::i32: + case MVT::i64: + assert(GPR_idx < 8 && "Too many int args"); + if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) { + BuildMI(BB, PPC::OR,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, PPC::FMR, 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::i1: + case MVT::i8: + case MVT::i16: + case MVT::i32: + case MVT::i64: + BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3); + break; + case MVT::f32: + case MVT::f64: + BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1); + break; + } + return Result+N.ResNo; + } + + case ISD::SIGN_EXTEND: + case ISD::SIGN_EXTEND_INREG: + Tmp1 = SelectExpr(N.getOperand(0)); + switch(cast(Node)->getExtraValueType()) { + default: Node->dump(); assert(0 && "Unhandled SIGN_EXTEND type"); break; + case MVT::i32: + BuildMI(BB, PPC::EXTSW, 1, Result).addReg(Tmp1); + 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; + case MVT::i1: + BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp1).addSImm(0); + break; + } + return Result; + + case ISD::ZERO_EXTEND_INREG: + Tmp1 = SelectExpr(N.getOperand(0)); + switch(cast(Node)->getExtraValueType()) { + default: Node->dump(); assert(0 && "Unhandled ZERO_EXTEND type"); break; + case MVT::i16: Tmp2 = 16; break; + case MVT::i8: Tmp2 = 24; break; + case MVT::i1: Tmp2 = 31; break; + } + BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(0).addImm(Tmp2) + .addImm(31); + return Result; + + case ISD::CopyFromReg: + if (Result == 1) + Result = ExprMap[N.getValue(0)] = MakeReg(N.getValue(0).getValueType()); + Tmp1 = dyn_cast(Node)->getReg(); + BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1); + return Result; + + case ISD::SHL: + Tmp1 = SelectExpr(N.getOperand(0)); + if (ConstantSDNode *CN = dyn_cast(N.getOperand(1))) { + Tmp2 = CN->getValue() & 0x1F; + BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(Tmp2).addImm(0) + .addImm(31-Tmp2); + } else { + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, PPC::SLW, 2, Result).addReg(Tmp1).addReg(Tmp2); + } + return Result; + + case ISD::SRL: + Tmp1 = SelectExpr(N.getOperand(0)); + if (ConstantSDNode *CN = dyn_cast(N.getOperand(1))) { + Tmp2 = CN->getValue() & 0x1F; + BuildMI(BB, PPC::RLWINM, 4, Result).addReg(Tmp1).addImm(32-Tmp2) + .addImm(Tmp2).addImm(31); + } else { + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, PPC::SRW, 2, Result).addReg(Tmp1).addReg(Tmp2); + } + return Result; + + case ISD::SRA: + Tmp1 = SelectExpr(N.getOperand(0)); + if (ConstantSDNode *CN = dyn_cast(N.getOperand(1))) { + Tmp2 = CN->getValue() & 0x1F; + BuildMI(BB, PPC::SRAWI, 2, Result).addReg(Tmp1).addImm(Tmp2); + } else { + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, PPC::SRAW, 2, Result).addReg(Tmp1).addReg(Tmp2); + } + return Result; + + case ISD::ADD: + Tmp1 = SelectExpr(N.getOperand(0)); + switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) { + default: assert(0 && "unhandled result code"); + case 0: // No immediate + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, PPC::ADD, 2, Result).addReg(Tmp1).addReg(Tmp2); + break; + case 1: // Low immediate + BuildMI(BB, PPC::ADDI, 2, Result).addReg(Tmp1).addSImm(Tmp2); + break; + case 2: // Shifted immediate + BuildMI(BB, PPC::ADDIS, 2, Result).addReg(Tmp1).addSImm(Tmp2); + break; + } + return Result; + + case ISD::AND: + case ISD::OR: + Tmp1 = SelectExpr(N.getOperand(0)); + switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) { + default: assert(0 && "unhandled result code"); + case 0: // No immediate + Tmp2 = SelectExpr(N.getOperand(1)); + switch (opcode) { + case ISD::AND: Opc = PPC::AND; break; + case ISD::OR: Opc = PPC::OR; break; + } + BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2); + break; + case 1: // Low immediate + switch (opcode) { + case ISD::AND: Opc = PPC::ANDIo; break; + case ISD::OR: Opc = PPC::ORI; break; + } + BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(Tmp2); + break; + case 2: // Shifted immediate + switch (opcode) { + case ISD::AND: Opc = PPC::ANDISo; break; + case ISD::OR: Opc = PPC::ORIS; break; + } + BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addImm(Tmp2); + break; + } + return Result; + + case ISD::XOR: { + // Check for EQV: xor, (xor a, -1), b + if (N.getOperand(0).getOpcode() == ISD::XOR && + N.getOperand(0).getOperand(1).getOpcode() == ISD::Constant && + cast(N.getOperand(0).getOperand(1))->isAllOnesValue()) { + ++NotLogic; + Tmp1 = SelectExpr(N.getOperand(0).getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, PPC::EQV, 2, Result).addReg(Tmp1).addReg(Tmp2); + return Result; + } + // Check for NOT, NOR, and NAND: xor (copy, or, and), -1 + if (N.getOperand(1).getOpcode() == ISD::Constant && + cast(N.getOperand(1))->isAllOnesValue()) { + ++NotLogic; + switch(N.getOperand(0).getOpcode()) { + case ISD::OR: + Tmp1 = SelectExpr(N.getOperand(0).getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(0).getOperand(1)); + BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp2); + break; + case ISD::AND: + Tmp1 = SelectExpr(N.getOperand(0).getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(0).getOperand(1)); + BuildMI(BB, PPC::NAND, 2, Result).addReg(Tmp1).addReg(Tmp2); + break; + default: + Tmp1 = SelectExpr(N.getOperand(0)); + BuildMI(BB, PPC::NOR, 2, Result).addReg(Tmp1).addReg(Tmp1); + break; + } + return Result; + } + Tmp1 = SelectExpr(N.getOperand(0)); + switch(getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) { + default: assert(0 && "unhandled result code"); + case 0: // No immediate + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, PPC::XOR, 2, Result).addReg(Tmp1).addReg(Tmp2); + break; + case 1: // Low immediate + BuildMI(BB, PPC::XORI, 2, Result).addReg(Tmp1).addImm(Tmp2); + break; + case 2: // Shifted immediate + BuildMI(BB, PPC::XORIS, 2, Result).addReg(Tmp1).addImm(Tmp2); + break; + } + return Result; + } + + case ISD::SUB: + Tmp2 = SelectExpr(N.getOperand(1)); + if (1 == getImmediateForOpcode(N.getOperand(0), opcode, Tmp1)) + BuildMI(BB, PPC::SUBFIC, 2, Result).addReg(Tmp2).addSImm(Tmp1); + else { + Tmp1 = SelectExpr(N.getOperand(0)); + BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp2).addReg(Tmp1); + } + return Result; + + case ISD::MUL: + Tmp1 = SelectExpr(N.getOperand(0)); + if (1 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp2)) + BuildMI(BB, PPC::MULLI, 2, Result).addReg(Tmp1).addSImm(Tmp2); + else { + Tmp2 = SelectExpr(N.getOperand(1)); + BuildMI(BB, PPC::MULLD, 2, Result).addReg(Tmp1).addReg(Tmp2); + } + return Result; + + case ISD::SDIV: + case ISD::UDIV: + if (3 == getImmediateForOpcode(N.getOperand(1), opcode, Tmp3)) { + Tmp1 = MakeReg(MVT::i64); + Tmp2 = SelectExpr(N.getOperand(0)); + BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3); + BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1); + return Result; + } + Tmp1 = SelectExpr(N.getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(1)); + Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW; + BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2); + return Result; + + case ISD::UREM: + case ISD::SREM: { + Tmp1 = SelectExpr(N.getOperand(0)); + Tmp2 = SelectExpr(N.getOperand(1)); + Tmp3 = MakeReg(MVT::i64); + unsigned Tmp4 = MakeReg(MVT::i64); + Opc = (ISD::UREM == opcode) ? PPC::DIVDU : PPC::DIVD; + BuildMI(BB, Opc, 2, Tmp3).addReg(Tmp1).addReg(Tmp2); + BuildMI(BB, PPC::MULLD, 2, Tmp4).addReg(Tmp3).addReg(Tmp2); + BuildMI(BB, PPC::SUBF, 2, Result).addReg(Tmp4).addReg(Tmp1); + return Result; + } + + case ISD::FP_TO_UINT: + case ISD::FP_TO_SINT: { + bool U = (ISD::FP_TO_UINT == opcode); + Tmp1 = SelectExpr(N.getOperand(0)); + if (!U) { + Tmp2 = MakeReg(MVT::f64); + BuildMI(BB, PPC::FCTIWZ, 1, Tmp2).addReg(Tmp1); + int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8); + addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx); + addFrameReference(BuildMI(BB, PPC::LWZ, 2, Result), FrameIdx, 4); + return Result; + } else { + unsigned Zero = getConstDouble(0.0); + unsigned MaxInt = getConstDouble((1LL << 32) - 1); + unsigned Border = getConstDouble(1LL << 31); + unsigned UseZero = MakeReg(MVT::f64); + unsigned UseMaxInt = MakeReg(MVT::f64); + unsigned UseChoice = MakeReg(MVT::f64); + unsigned TmpReg = MakeReg(MVT::f64); + unsigned TmpReg2 = MakeReg(MVT::f64); + unsigned ConvReg = MakeReg(MVT::f64); + unsigned IntTmp = MakeReg(MVT::i32); + unsigned XorReg = MakeReg(MVT::i32); + MachineFunction *F = BB->getParent(); + int FrameIdx = F->getFrameInfo()->CreateStackObject(8, 8); + // Update machine-CFG edges + MachineBasicBlock *XorMBB = new MachineBasicBlock(BB->getBasicBlock()); + MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock()); + MachineBasicBlock *OldMBB = BB; + ilist::iterator It = BB; ++It; + F->getBasicBlockList().insert(It, XorMBB); + F->getBasicBlockList().insert(It, PhiMBB); + BB->addSuccessor(XorMBB); + BB->addSuccessor(PhiMBB); + // Convert from floating point to unsigned 32-bit value + // Use 0 if incoming value is < 0.0 + BuildMI(BB, PPC::FSEL, 3, UseZero).addReg(Tmp1).addReg(Tmp1).addReg(Zero); + // Use 2**32 - 1 if incoming value is >= 2**32 + BuildMI(BB, PPC::FSUB, 2, UseMaxInt).addReg(MaxInt).addReg(Tmp1); + BuildMI(BB, PPC::FSEL, 3, UseChoice).addReg(UseMaxInt).addReg(UseZero) + .addReg(MaxInt); + // Subtract 2**31 + BuildMI(BB, PPC::FSUB, 2, TmpReg).addReg(UseChoice).addReg(Border); + // Use difference if >= 2**31 + BuildMI(BB, PPC::FCMPU, 2, PPC::CR0).addReg(UseChoice).addReg(Border); + BuildMI(BB, PPC::FSEL, 3, TmpReg2).addReg(TmpReg).addReg(TmpReg) + .addReg(UseChoice); + // Convert to integer + BuildMI(BB, PPC::FCTIWZ, 1, ConvReg).addReg(TmpReg2); + addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(ConvReg), FrameIdx); + addFrameReference(BuildMI(BB, PPC::LWZ, 2, IntTmp), FrameIdx, 4); + BuildMI(BB, PPC::BLT, 2).addReg(PPC::CR0).addMBB(PhiMBB); + BuildMI(BB, PPC::B, 1).addMBB(XorMBB); + + // XorMBB: + // add 2**31 if input was >= 2**31 + BB = XorMBB; + BuildMI(BB, PPC::XORIS, 2, XorReg).addReg(IntTmp).addImm(0x8000); + XorMBB->addSuccessor(PhiMBB); + + // PhiMBB: + // DestReg = phi [ IntTmp, OldMBB ], [ XorReg, XorMBB ] + BB = PhiMBB; + BuildMI(BB, PPC::PHI, 4, Result).addReg(IntTmp).addMBB(OldMBB) + .addReg(XorReg).addMBB(XorMBB); + return Result; + } + assert(0 && "Should never get here"); + return 0; + } + + case ISD::SETCC: + if (SetCCSDNode *SetCC = dyn_cast(Node)) { + Opc = SelectSetCR0(N); + + unsigned TrueValue = MakeReg(MVT::i32); + BuildMI(BB, PPC::LI, 1, TrueValue).addSImm(1); + unsigned FalseValue = MakeReg(MVT::i32); + BuildMI(BB, PPC::LI, 1, FalseValue).addSImm(0); + + // Create an iterator with which to insert the MBB for copying the false + // value and the MBB to hold the PHI instruction for this SetCC. + MachineBasicBlock *thisMBB = BB; + const BasicBlock *LLVM_BB = BB->getBasicBlock(); + ilist::iterator It = BB; + ++It; + + // thisMBB: + // ... + // cmpTY cr0, r1, r2 + // %TrueValue = li 1 + // bCC sinkMBB + MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB); + MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB); + BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB); + MachineFunction *F = BB->getParent(); + F->getBasicBlockList().insert(It, copy0MBB); + F->getBasicBlockList().insert(It, sinkMBB); + // Update machine-CFG edges + BB->addSuccessor(copy0MBB); + BB->addSuccessor(sinkMBB); + + // copy0MBB: + // %FalseValue = li 0 + // fallthrough + BB = copy0MBB; + // Update machine-CFG edges + BB->addSuccessor(sinkMBB); + + // sinkMBB: + // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] + // ... + BB = sinkMBB; + BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue) + .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB); + return Result; + } + assert(0 && "Is this legal?"); + return 0; + + case ISD::SELECT: { + unsigned TrueValue = SelectExpr(N.getOperand(1)); //Use if TRUE + unsigned FalseValue = SelectExpr(N.getOperand(2)); //Use if FALSE + Opc = SelectSetCR0(N.getOperand(0)); + + // Create an iterator with which to insert the MBB for copying the false + // value and the MBB to hold the PHI instruction for this SetCC. + MachineBasicBlock *thisMBB = BB; + const BasicBlock *LLVM_BB = BB->getBasicBlock(); + ilist::iterator It = BB; + ++It; + + // thisMBB: + // ... + // TrueVal = ... + // cmpTY cr0, r1, r2 + // bCC copy1MBB + // fallthrough --> copy0MBB + MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB); + MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB); + BuildMI(BB, Opc, 2).addReg(PPC::CR0).addMBB(sinkMBB); + MachineFunction *F = BB->getParent(); + F->getBasicBlockList().insert(It, copy0MBB); + F->getBasicBlockList().insert(It, sinkMBB); + // Update machine-CFG edges + BB->addSuccessor(copy0MBB); + BB->addSuccessor(sinkMBB); + + // copy0MBB: + // %FalseValue = ... + // # fallthrough to sinkMBB + BB = copy0MBB; + // Update machine-CFG edges + BB->addSuccessor(sinkMBB); + + // sinkMBB: + // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] + // ... + BB = sinkMBB; + BuildMI(BB, PPC::PHI, 4, Result).addReg(FalseValue) + .addMBB(copy0MBB).addReg(TrueValue).addMBB(thisMBB); + + // FIXME: Select i64? + return Result; + } + + case ISD::Constant: + switch (N.getValueType()) { + default: assert(0 && "Cannot use constants of this type!"); + case MVT::i1: + BuildMI(BB, PPC::LI, 1, Result) + .addSImm(!cast(N)->isNullValue()); + break; + case MVT::i32: + { + int v = (int)cast(N)->getSignExtended(); + if (v < 32768 && v >= -32768) { + BuildMI(BB, PPC::LI, 1, Result).addSImm(v); + } else { + Tmp1 = MakeReg(MVT::i32); + BuildMI(BB, PPC::LIS, 1, Tmp1).addSImm(v >> 16); + BuildMI(BB, PPC::ORI, 2, Result).addReg(Tmp1).addImm(v & 0xFFFF); + } + } + } + return Result; + } + + return 0; +} + +void ISel::Select(SDOperand N) { + unsigned Tmp1, Tmp2, Opc; + unsigned opcode = N.getOpcode(); + + if (!ExprMap.insert(std::make_pair(N, 1)).second) + return; // Already selected. + + SDNode *Node = N.Val; + + switch (Node->getOpcode()) { + default: + Node->dump(); std::cerr << "\n"; + assert(0 && "Node not handled yet!"); + case ISD::EntryToken: return; // Noop + case ISD::TokenFactor: + for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) + Select(Node->getOperand(i)); + return; + case ISD::ADJCALLSTACKDOWN: + case ISD::ADJCALLSTACKUP: + Select(N.getOperand(0)); + Tmp1 = cast(N.getOperand(1))->getValue(); + Opc = N.getOpcode() == ISD::ADJCALLSTACKDOWN ? PPC::ADJCALLSTACKDOWN : + PPC::ADJCALLSTACKUP; + BuildMI(BB, Opc, 1).addImm(Tmp1); + return; + case ISD::BR: { + MachineBasicBlock *Dest = + cast(N.getOperand(1))->getBasicBlock(); + Select(N.getOperand(0)); + BuildMI(BB, PPC::B, 1).addMBB(Dest); + return; + } + case ISD::BRCOND: + SelectBranchCC(N); + return; + case ISD::CopyToReg: + Select(N.getOperand(0)); + Tmp1 = SelectExpr(N.getOperand(1)); + Tmp2 = cast(N)->getReg(); + + if (Tmp1 != Tmp2) { + if (N.getOperand(1).getValueType() == MVT::f64 || + N.getOperand(1).getValueType() == MVT::f32) + BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1); + else + BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1); + } + return; + case ISD::ImplicitDef: + Select(N.getOperand(0)); + BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast(N)->getReg()); + return; + case ISD::RET: + switch (N.getNumOperands()) { + default: + assert(0 && "Unknown return instruction!"); + case 3: + assert(N.getOperand(1).getValueType() == MVT::i32 && + N.getOperand(2).getValueType() == MVT::i32 && + "Unknown two-register value!"); + Select(N.getOperand(0)); + Tmp1 = SelectExpr(N.getOperand(1)); + Tmp2 = SelectExpr(N.getOperand(2)); + BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp2).addReg(Tmp2); + BuildMI(BB, PPC::OR, 2, PPC::R4).addReg(Tmp1).addReg(Tmp1); + break; + case 2: + Select(N.getOperand(0)); + Tmp1 = SelectExpr(N.getOperand(1)); + switch (N.getOperand(1).getValueType()) { + default: + assert(0 && "Unknown return type!"); + case MVT::f64: + case MVT::f32: + BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1); + break; + case MVT::i32: + BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1); + break; + } + case 1: + Select(N.getOperand(0)); + break; + } + BuildMI(BB, PPC::BLR, 0); // Just emit a 'ret' instruction + return; + case ISD::TRUNCSTORE: + case ISD::STORE: + { + SDOperand Chain = N.getOperand(0); + SDOperand Value = N.getOperand(1); + SDOperand Address = N.getOperand(2); + Select(Chain); + + Tmp1 = SelectExpr(Value); //value + + if (opcode == ISD::STORE) { + switch(Value.getValueType()) { + default: assert(0 && "unknown Type in store"); + case MVT::i64: Opc = PPC::STD; break; + case MVT::f64: Opc = PPC::STFD; break; + case MVT::f32: Opc = PPC::STFS; break; + } + } else { //ISD::TRUNCSTORE + switch(cast(Node)->getExtraValueType()) { + default: assert(0 && "unknown Type in store"); + case MVT::i1: //FIXME: DAG does not promote this load + case MVT::i8: Opc= PPC::STB; break; + case MVT::i16: Opc = PPC::STH; break; + case MVT::i32: Opc = PPC::STW; break; + } + } + + if(Address.getOpcode() == ISD::FrameIndex) + { + Tmp2 = cast(Address)->getIndex(); + addFrameReference(BuildMI(BB, Opc, 3).addReg(Tmp1), (int)Tmp2); + } + else + { + int offset; + bool idx = SelectAddr(Address, Tmp2, offset); + if (idx) { + Opc = IndexedOpForOp(Opc); + BuildMI(BB, Opc, 3).addReg(Tmp1).addReg(Tmp2).addReg(offset); + } else { + BuildMI(BB, Opc, 3).addReg(Tmp1).addImm(offset).addReg(Tmp2); + } + } + return; + } + case ISD::EXTLOAD: + case ISD::SEXTLOAD: + case ISD::ZEXTLOAD: + case ISD::LOAD: + case ISD::CopyFromReg: + case ISD::CALL: + case ISD::DYNAMIC_STACKALLOC: + ExprMap.erase(N); + SelectExpr(N); + return; + } + assert(0 && "Should not be reached!"); +} + + +/// createPPC32PatternInstructionSelector - This pass converts an LLVM function +/// into a machine code representation using pattern matching and a machine +/// description file. +/// +FunctionPass *llvm::createPPC64ISelPattern(TargetMachine &TM) { + return new ISel(TM); +} + diff --git a/lib/Target/PowerPC/PPC64ISelSimple.cpp b/lib/Target/PowerPC/PPC64ISelSimple.cpp deleted file mode 100644 index 30531b6eeef..00000000000 --- a/lib/Target/PowerPC/PPC64ISelSimple.cpp +++ /dev/null @@ -1,2946 +0,0 @@ -//===-- PPC64ISelSimple.cpp - A simple instruction selector for PowerPC ---===// -// -// The LLVM Compiler Infrastructure -// -// This file was developed by the LLVM research group and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. -// -//===----------------------------------------------------------------------===// - -#define DEBUG_TYPE "isel" -#include "PowerPC.h" -#include "PowerPCInstrBuilder.h" -#include "PowerPCInstrInfo.h" -#include "PPC64TargetMachine.h" -#include "llvm/Constants.h" -#include "llvm/DerivedTypes.h" -#include "llvm/Function.h" -#include "llvm/Instructions.h" -#include "llvm/Pass.h" -#include "llvm/CodeGen/IntrinsicLowering.h" -#include "llvm/CodeGen/MachineConstantPool.h" -#include "llvm/CodeGen/MachineFrameInfo.h" -#include "llvm/CodeGen/MachineFunction.h" -#include "llvm/CodeGen/SSARegMap.h" -#include "llvm/Target/MRegisterInfo.h" -#include "llvm/Target/TargetMachine.h" -#include "llvm/Support/GetElementPtrTypeIterator.h" -#include "llvm/Support/InstVisitor.h" -#include "llvm/Support/Debug.h" -#include "llvm/ADT/Statistic.h" -#include -using namespace llvm; - -namespace { - Statistic<> GEPFolds("ppc64-codegen", "Number of GEPs folded"); - - /// TypeClass - Used by the PowerPC backend to group LLVM types by their basic - /// PPC Representation. - /// - enum TypeClass { - cByte, cShort, cInt, cFP32, cFP64, cLong - }; -} - -/// getClass - Turn a primitive type into a "class" number which is based on the -/// size of the type, and whether or not it is floating point. -/// -static inline TypeClass getClass(const Type *Ty) { - switch (Ty->getTypeID()) { - case Type::SByteTyID: - case Type::UByteTyID: return cByte; // Byte operands are class #0 - case Type::ShortTyID: - case Type::UShortTyID: return cShort; // Short operands are class #1 - case Type::IntTyID: - case Type::UIntTyID: return cInt; // Ints are class #2 - - case Type::FloatTyID: return cFP32; // Single float is #3 - case Type::DoubleTyID: return cFP64; // Double Point is #4 - - case Type::PointerTyID: - case Type::LongTyID: - case Type::ULongTyID: return cLong; // Longs and pointers are class #5 - default: - assert(0 && "Invalid type to getClass!"); - return cByte; // not reached - } -} - -// getClassB - Just like getClass, but treat boolean values as ints. -static inline TypeClass getClassB(const Type *Ty) { - if (Ty == Type::BoolTy) return cInt; - return getClass(Ty); -} - -namespace { - struct PPC64ISel : public FunctionPass, InstVisitor { - PPC64TargetMachine &TM; - MachineFunction *F; // The function we are compiling into - MachineBasicBlock *BB; // The current MBB we are compiling - int VarArgsFrameIndex; // FrameIndex for start of varargs area - - std::map RegMap; // Mapping between Values and SSA Regs - - // External functions used in the Module - Function *fmodfFn, *fmodFn, *__cmpdi2Fn, *__fixsfdiFn, *__fixdfdiFn, - *__fixunssfdiFn, *__fixunsdfdiFn, *mallocFn, *freeFn; - - // MBBMap - Mapping between LLVM BB -> Machine BB - std::map MBBMap; - - // AllocaMap - Mapping from fixed sized alloca instructions to the - // FrameIndex for the alloca. - std::map AllocaMap; - - // Target configuration data - const unsigned ParameterSaveAreaOffset, MaxArgumentStackSpace; - - PPC64ISel(TargetMachine &tm):TM(reinterpret_cast(tm)), - F(0), BB(0), ParameterSaveAreaOffset(24), MaxArgumentStackSpace(32) {} - - bool doInitialization(Module &M) { - // Add external functions that we may call - Type *i = Type::IntTy; - Type *d = Type::DoubleTy; - Type *f = Type::FloatTy; - Type *l = Type::LongTy; - Type *ul = Type::ULongTy; - Type *voidPtr = PointerType::get(Type::SByteTy); - // float fmodf(float, float); - fmodfFn = M.getOrInsertFunction("fmodf", f, f, f, 0); - // double fmod(double, double); - fmodFn = M.getOrInsertFunction("fmod", d, d, d, 0); - // int __cmpdi2(long, long); - __cmpdi2Fn = M.getOrInsertFunction("__cmpdi2", i, l, l, 0); - // long __fixsfdi(float) - __fixsfdiFn = M.getOrInsertFunction("__fixsfdi", l, f, 0); - // long __fixdfdi(double) - __fixdfdiFn = M.getOrInsertFunction("__fixdfdi", l, d, 0); - // unsigned long __fixunssfdi(float) - __fixunssfdiFn = M.getOrInsertFunction("__fixunssfdi", ul, f, 0); - // unsigned long __fixunsdfdi(double) - __fixunsdfdiFn = M.getOrInsertFunction("__fixunsdfdi", ul, d, 0); - // void* malloc(size_t) - mallocFn = M.getOrInsertFunction("malloc", voidPtr, Type::UIntTy, 0); - // void free(void*) - freeFn = M.getOrInsertFunction("free", Type::VoidTy, voidPtr, 0); - return false; - } - - /// runOnFunction - Top level implementation of instruction selection for - /// the entire function. - /// - bool runOnFunction(Function &Fn) { - // First pass over the function, lower any unknown intrinsic functions - // with the IntrinsicLowering class. - LowerUnknownIntrinsicFunctionCalls(Fn); - - F = &MachineFunction::construct(&Fn, TM); - - // Create all of the machine basic blocks for the function... - for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) - F->getBasicBlockList().push_back(MBBMap[I] = new MachineBasicBlock(I)); - - BB = &F->front(); - - // Copy incoming arguments off of the stack... - LoadArgumentsToVirtualRegs(Fn); - - // Instruction select everything except PHI nodes - visit(Fn); - - // Select the PHI nodes - SelectPHINodes(); - - RegMap.clear(); - MBBMap.clear(); - AllocaMap.clear(); - F = 0; - // We always build a machine code representation for the function - return true; - } - - virtual const char *getPassName() const { - return "PowerPC Simple Instruction Selection"; - } - - /// visitBasicBlock - This method is called when we are visiting a new basic - /// block. This simply creates a new MachineBasicBlock to emit code into - /// and adds it to the current MachineFunction. Subsequent visit* for - /// instructions will be invoked for all instructions in the basic block. - /// - void visitBasicBlock(BasicBlock &LLVM_BB) { - BB = MBBMap[&LLVM_BB]; - } - - /// LowerUnknownIntrinsicFunctionCalls - This performs a prepass over the - /// function, lowering any calls to unknown intrinsic functions into the - /// equivalent LLVM code. - /// - void LowerUnknownIntrinsicFunctionCalls(Function &F); - - /// LoadArgumentsToVirtualRegs - Load all of the arguments to this function - /// from the stack into virtual registers. - /// - void LoadArgumentsToVirtualRegs(Function &F); - - /// SelectPHINodes - Insert machine code to generate phis. This is tricky - /// because we have to generate our sources into the source basic blocks, - /// not the current one. - /// - void SelectPHINodes(); - - // Visitation methods for various instructions. These methods simply emit - // fixed PowerPC code for each instruction. - - // Control flow operators - void visitReturnInst(ReturnInst &RI); - void visitBranchInst(BranchInst &BI); - - struct ValueRecord { - Value *Val; - unsigned Reg; - const Type *Ty; - ValueRecord(unsigned R, const Type *T) : Val(0), Reg(R), Ty(T) {} - ValueRecord(Value *V) : Val(V), Reg(0), Ty(V->getType()) {} - }; - - // This struct is for recording the necessary operations to emit the GEP - struct CollapsedGepOp { - bool isMul; - Value *index; - ConstantSInt *size; - CollapsedGepOp(bool mul, Value *i, ConstantSInt *s) : - isMul(mul), index(i), size(s) {} - }; - - void doCall(const ValueRecord &Ret, MachineInstr *CallMI, - const std::vector &Args, bool isVarArg); - void visitCallInst(CallInst &I); - void visitIntrinsicCall(Intrinsic::ID ID, CallInst &I); - - // Arithmetic operators - void visitSimpleBinary(BinaryOperator &B, unsigned OpcodeClass); - void visitAdd(BinaryOperator &B) { visitSimpleBinary(B, 0); } - void visitSub(BinaryOperator &B) { visitSimpleBinary(B, 1); } - void visitMul(BinaryOperator &B); - - void visitDiv(BinaryOperator &B) { visitDivRem(B); } - void visitRem(BinaryOperator &B) { visitDivRem(B); } - void visitDivRem(BinaryOperator &B); - - // Bitwise operators - void visitAnd(BinaryOperator &B) { visitSimpleBinary(B, 2); } - void visitOr (BinaryOperator &B) { visitSimpleBinary(B, 3); } - void visitXor(BinaryOperator &B) { visitSimpleBinary(B, 4); } - - // Comparison operators... - void visitSetCondInst(SetCondInst &I); - unsigned EmitComparison(unsigned OpNum, Value *Op0, Value *Op1, - MachineBasicBlock *MBB, - MachineBasicBlock::iterator MBBI); - void visitSelectInst(SelectInst &SI); - - - // Memory Instructions - void visitLoadInst(LoadInst &I); - void visitStoreInst(StoreInst &I); - void visitGetElementPtrInst(GetElementPtrInst &I); - void visitAllocaInst(AllocaInst &I); - void visitMallocInst(MallocInst &I); - void visitFreeInst(FreeInst &I); - - // Other operators - void visitShiftInst(ShiftInst &I); - void visitPHINode(PHINode &I) {} // PHI nodes handled by second pass - void visitCastInst(CastInst &I); - void visitVANextInst(VANextInst &I); - void visitVAArgInst(VAArgInst &I); - - void visitInstruction(Instruction &I) { - std::cerr << "Cannot instruction select: " << I; - abort(); - } - - /// promote32 - Make a value 32-bits wide, and put it somewhere. - /// - void promote32(unsigned targetReg, const ValueRecord &VR); - - /// emitGEPOperation - Common code shared between visitGetElementPtrInst and - /// constant expression GEP support. - /// - void emitGEPOperation(MachineBasicBlock *BB, MachineBasicBlock::iterator IP, - Value *Src, User::op_iterator IdxBegin, - User::op_iterator IdxEnd, unsigned TargetReg, - bool CollapseRemainder, ConstantSInt **Remainder, - unsigned *PendingAddReg); - - /// emitCastOperation - Common code shared between visitCastInst and - /// constant expression cast support. - /// - void emitCastOperation(MachineBasicBlock *BB,MachineBasicBlock::iterator IP, - Value *Src, const Type *DestTy, unsigned TargetReg); - - /// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary - /// and constant expression support. - /// - void emitSimpleBinaryOperation(MachineBasicBlock *BB, - MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, - unsigned OperatorClass, unsigned TargetReg); - - /// emitBinaryFPOperation - This method handles emission of floating point - /// Add (0), Sub (1), Mul (2), and Div (3) operations. - void emitBinaryFPOperation(MachineBasicBlock *BB, - MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, - unsigned OperatorClass, unsigned TargetReg); - - void emitMultiply(MachineBasicBlock *BB, MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, unsigned TargetReg); - - void doMultiply(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - unsigned DestReg, Value *Op0, Value *Op1); - - /// doMultiplyConst - This method will multiply the value in Op0Reg by the - /// value of the ContantInt *CI - void doMultiplyConst(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - unsigned DestReg, Value *Op0, ConstantInt *CI); - - void emitDivRemOperation(MachineBasicBlock *BB, - MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, bool isDiv, - unsigned TargetReg); - - /// emitSetCCOperation - Common code shared between visitSetCondInst and - /// constant expression support. - /// - void emitSetCCOperation(MachineBasicBlock *BB, - MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, unsigned Opcode, - unsigned TargetReg); - - /// emitShiftOperation - Common code shared between visitShiftInst and - /// constant expression support. - /// - void emitShiftOperation(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Value *Op, Value *ShiftAmount, bool isLeftShift, - const Type *ResultTy, unsigned DestReg); - - /// emitSelectOperation - Common code shared between visitSelectInst and the - /// constant expression support. - /// - void emitSelectOperation(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Value *Cond, Value *TrueVal, Value *FalseVal, - unsigned DestReg); - - /// copyConstantToRegister - Output the instructions required to put the - /// specified constant into the specified register. - /// - void copyConstantToRegister(MachineBasicBlock *MBB, - MachineBasicBlock::iterator MBBI, - Constant *C, unsigned Reg); - - void emitUCOM(MachineBasicBlock *MBB, MachineBasicBlock::iterator MBBI, - unsigned LHS, unsigned RHS); - - /// makeAnotherReg - This method returns the next register number we haven't - /// yet used. - /// - unsigned makeAnotherReg(const Type *Ty) { - assert(dynamic_cast(TM.getRegisterInfo()) && - "Current target doesn't have PPC reg info??"); - const PPC64RegisterInfo *PPCRI = - static_cast(TM.getRegisterInfo()); - // Add the mapping of regnumber => reg class to MachineFunction - const TargetRegisterClass *RC = PPCRI->getRegClassForType(Ty); - return F->getSSARegMap()->createVirtualRegister(RC); - } - - /// getReg - This method turns an LLVM value into a register number. - /// - unsigned getReg(Value &V) { return getReg(&V); } // Allow references - unsigned getReg(Value *V) { - // Just append to the end of the current bb. - MachineBasicBlock::iterator It = BB->end(); - return getReg(V, BB, It); - } - unsigned getReg(Value *V, MachineBasicBlock *MBB, - MachineBasicBlock::iterator IPt); - - /// canUseAsImmediateForOpcode - This method returns whether a ConstantInt - /// is okay to use as an immediate argument to a certain binary operation - bool canUseAsImmediateForOpcode(ConstantInt *CI, unsigned Opcode); - - /// getFixedSizedAllocaFI - Return the frame index for a fixed sized alloca - /// that is to be statically allocated with the initial stack frame - /// adjustment. - unsigned getFixedSizedAllocaFI(AllocaInst *AI); - }; -} - -/// dyn_castFixedAlloca - If the specified value is a fixed size alloca -/// instruction in the entry block, return it. Otherwise, return a null -/// pointer. -static AllocaInst *dyn_castFixedAlloca(Value *V) { - if (AllocaInst *AI = dyn_cast(V)) { - BasicBlock *BB = AI->getParent(); - if (isa(AI->getArraySize()) && BB ==&BB->getParent()->front()) - return AI; - } - return 0; -} - -/// getReg - This method turns an LLVM value into a register number. -/// -unsigned PPC64ISel::getReg(Value *V, MachineBasicBlock *MBB, - MachineBasicBlock::iterator IPt) { - if (Constant *C = dyn_cast(V)) { - unsigned Reg = makeAnotherReg(V->getType()); - copyConstantToRegister(MBB, IPt, C, Reg); - return Reg; - } else if (AllocaInst *AI = dyn_castFixedAlloca(V)) { - unsigned Reg = makeAnotherReg(V->getType()); - unsigned FI = getFixedSizedAllocaFI(AI); - addFrameReference(BuildMI(*MBB, IPt, PPC::ADDI, 2, Reg), FI, 0, false); - return Reg; - } - - unsigned &Reg = RegMap[V]; - if (Reg == 0) { - Reg = makeAnotherReg(V->getType()); - RegMap[V] = Reg; - } - - return Reg; -} - -/// canUseAsImmediateForOpcode - This method returns whether a ConstantInt -/// is okay to use as an immediate argument to a certain binary operator. -/// -/// Operator is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or, 4 for Xor. -bool PPC64ISel::canUseAsImmediateForOpcode(ConstantInt *CI, unsigned Operator) { - ConstantSInt *Op1Cs; - ConstantUInt *Op1Cu; - - // ADDI, Compare, and non-indexed Load take SIMM - bool cond1 = (Operator == 0) - && (Op1Cs = dyn_cast(CI)) - && (Op1Cs->getValue() <= 32767) - && (Op1Cs->getValue() >= -32768); - - // SUBI takes -SIMM since it is a mnemonic for ADDI - bool cond2 = (Operator == 1) - && (Op1Cs = dyn_cast(CI)) - && (Op1Cs->getValue() <= 32768) - && (Op1Cs->getValue() >= -32767); - - // ANDIo, ORI, and XORI take unsigned values - bool cond3 = (Operator >= 2) - && (Op1Cs = dyn_cast(CI)) - && (Op1Cs->getValue() >= 0) - && (Op1Cs->getValue() <= 32767); - - // ADDI and SUBI take SIMMs, so we have to make sure the UInt would fit - bool cond4 = (Operator < 2) - && (Op1Cu = dyn_cast(CI)) - && (Op1Cu->getValue() <= 32767); - - // ANDIo, ORI, and XORI take UIMMs, so they can be larger - bool cond5 = (Operator >= 2) - && (Op1Cu = dyn_cast(CI)) - && (Op1Cu->getValue() <= 65535); - - if (cond1 || cond2 || cond3 || cond4 || cond5) - return true; - - return false; -} - -/// getFixedSizedAllocaFI - Return the frame index for a fixed sized alloca -/// that is to be statically allocated with the initial stack frame -/// adjustment. -unsigned PPC64ISel::getFixedSizedAllocaFI(AllocaInst *AI) { - // Already computed this? - std::map::iterator I = AllocaMap.lower_bound(AI); - if (I != AllocaMap.end() && I->first == AI) return I->second; - - const Type *Ty = AI->getAllocatedType(); - ConstantUInt *CUI = cast(AI->getArraySize()); - unsigned TySize = TM.getTargetData().getTypeSize(Ty); - TySize *= CUI->getValue(); // Get total allocated size... - unsigned Alignment = TM.getTargetData().getTypeAlignment(Ty); - - // Create a new stack object using the frame manager... - int FrameIdx = F->getFrameInfo()->CreateStackObject(TySize, Alignment); - AllocaMap.insert(I, std::make_pair(AI, FrameIdx)); - return FrameIdx; -} - - -/// copyConstantToRegister - Output the instructions required to put the -/// specified constant into the specified register. -/// -void PPC64ISel::copyConstantToRegister(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Constant *C, unsigned R) { - if (C->getType()->isIntegral()) { - unsigned Class = getClassB(C->getType()); - - if (Class == cLong) { - if (ConstantUInt *CUI = dyn_cast(C)) { - uint64_t uval = CUI->getValue(); - if (uval < (1LL << 32)) { - ConstantUInt *CU = ConstantUInt::get(Type::UIntTy, uval); - copyConstantToRegister(MBB, IP, CU, R); - return; - } - } else if (ConstantSInt *CSI = dyn_cast(C)) { - int64_t val = CUI->getValue(); - if (val < (1LL << 31)) { - ConstantUInt *CU = ConstantUInt::get(Type::UIntTy, val); - copyConstantToRegister(MBB, IP, CU, R); - return; - } - } else { - std::cerr << "Unhandled long constant type!\n"; - abort(); - } - // Spill long to the constant pool and load it - MachineConstantPool *CP = F->getConstantPool(); - unsigned CPI = CP->getConstantPoolIndex(C); - BuildMI(*MBB, IP, PPC::LD, 1, R) - .addReg(PPC::R2).addConstantPoolIndex(CPI); - return; - } - - assert(Class <= cInt && "Type not handled yet!"); - - // Handle bool - if (C->getType() == Type::BoolTy) { - BuildMI(*MBB, IP, PPC::LI, 1, R).addSImm(C == ConstantBool::True); - return; - } - - // Handle int - if (ConstantUInt *CUI = dyn_cast(C)) { - unsigned uval = CUI->getValue(); - if (uval < 32768) { - BuildMI(*MBB, IP, PPC::LI, 1, R).addSImm(uval); - } else { - unsigned Temp = makeAnotherReg(Type::IntTy); - BuildMI(*MBB, IP, PPC::LIS, 1, Temp).addSImm(uval >> 16); - BuildMI(*MBB, IP, PPC::ORI, 2, R).addReg(Temp).addImm(uval); - } - return; - } else if (ConstantSInt *CSI = dyn_cast(C)) { - int sval = CSI->getValue(); - if (sval < 32768 && sval >= -32768) { - BuildMI(*MBB, IP, PPC::LI, 1, R).addSImm(sval); - } else { - unsigned Temp = makeAnotherReg(Type::IntTy); - BuildMI(*MBB, IP, PPC::LIS, 1, Temp).addSImm(sval >> 16); - BuildMI(*MBB, IP, PPC::ORI, 2, R).addReg(Temp).addImm(sval); - } - return; - } - std::cerr << "Unhandled integer constant!\n"; - abort(); - } else if (ConstantFP *CFP = dyn_cast(C)) { - // We need to spill the constant to memory... - MachineConstantPool *CP = F->getConstantPool(); - unsigned CPI = CP->getConstantPoolIndex(CFP); - const Type *Ty = CFP->getType(); - unsigned LoadOpcode = (Ty == Type::FloatTy) ? PPC::LFS : PPC::LFD; - BuildMI(*MBB,IP,LoadOpcode,2,R).addConstantPoolIndex(CPI).addReg(PPC::R2); - } else if (isa(C)) { - // Copy zero (null pointer) to the register. - BuildMI(*MBB, IP, PPC::LI, 1, R).addSImm(0); - } else if (GlobalValue *GV = dyn_cast(C)) { - static unsigned OpcodeTable[] = { - PPC::LBZ, PPC::LHZ, PPC::LWZ, PPC::LFS, PPC::LFD, PPC::LD - }; - unsigned Opcode = OpcodeTable[getClassB(GV->getType())]; - BuildMI(*MBB, IP, Opcode, 2, R).addGlobalAddress(GV).addReg(PPC::R2); - } else { - std::cerr << "Offending constant: " << *C << "\n"; - assert(0 && "Type not handled yet!"); - } -} - -/// LoadArgumentsToVirtualRegs - Load all of the arguments to this function from -/// the stack into virtual registers. -void PPC64ISel::LoadArgumentsToVirtualRegs(Function &Fn) { - unsigned ArgOffset = ParameterSaveAreaOffset; - unsigned GPR_remaining = 8; - unsigned FPR_remaining = 13; - 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 - }; - - MachineFrameInfo *MFI = F->getFrameInfo(); - - for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(); I != E; ++I) { - bool ArgLive = !I->use_empty(); - unsigned Reg = ArgLive ? getReg(*I) : 0; - int FI; // Frame object index - - switch (getClassB(I->getType())) { - case cByte: - if (ArgLive) { - FI = MFI->CreateFixedObject(4, ArgOffset); - if (GPR_remaining > 0) { - BuildMI(BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]); - BuildMI(BB, PPC::OR, 2, Reg).addReg(GPR[GPR_idx]) - .addReg(GPR[GPR_idx]); - } else { - addFrameReference(BuildMI(BB, PPC::LBZ, 2, Reg), FI); - } - } - break; - case cShort: - if (ArgLive) { - FI = MFI->CreateFixedObject(4, ArgOffset); - if (GPR_remaining > 0) { - BuildMI(BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]); - BuildMI(BB, PPC::OR, 2, Reg).addReg(GPR[GPR_idx]) - .addReg(GPR[GPR_idx]); - } else { - addFrameReference(BuildMI(BB, PPC::LHZ, 2, Reg), FI); - } - } - break; - case cInt: - if (ArgLive) { - FI = MFI->CreateFixedObject(4, ArgOffset); - if (GPR_remaining > 0) { - BuildMI(BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]); - BuildMI(BB, PPC::OR, 2, Reg).addReg(GPR[GPR_idx]) - .addReg(GPR[GPR_idx]); - } else { - addFrameReference(BuildMI(BB, PPC::LWZ, 2, Reg), FI); - } - } - break; - case cLong: - if (ArgLive) { - FI = MFI->CreateFixedObject(8, ArgOffset); - if (GPR_remaining > 1) { - BuildMI(BB, PPC::IMPLICIT_DEF, 0, GPR[GPR_idx]); - BuildMI(BB, PPC::OR, 2, Reg).addReg(GPR[GPR_idx]) - .addReg(GPR[GPR_idx]); - } else { - addFrameReference(BuildMI(BB, PPC::LD, 2, Reg), FI); - } - } - // longs require 4 additional bytes - ArgOffset += 4; - break; - case cFP32: - if (ArgLive) { - FI = MFI->CreateFixedObject(4, ArgOffset); - - if (FPR_remaining > 0) { - BuildMI(BB, PPC::IMPLICIT_DEF, 0, FPR[FPR_idx]); - BuildMI(BB, PPC::FMR, 1, Reg).addReg(FPR[FPR_idx]); - FPR_remaining--; - FPR_idx++; - } else { - addFrameReference(BuildMI(BB, PPC::LFS, 2, Reg), FI); - } - } - break; - case cFP64: - if (ArgLive) { - FI = MFI->CreateFixedObject(8, ArgOffset); - - if (FPR_remaining > 0) { - BuildMI(BB, PPC::IMPLICIT_DEF, 0, FPR[FPR_idx]); - BuildMI(BB, PPC::FMR, 1, Reg).addReg(FPR[FPR_idx]); - FPR_remaining--; - FPR_idx++; - } else { - addFrameReference(BuildMI(BB, PPC::LFD, 2, Reg), FI); - } - } - - // doubles require 4 additional bytes and use 2 GPRs of param space - ArgOffset += 4; - if (GPR_remaining > 0) { - GPR_remaining--; - GPR_idx++; - } - break; - default: - assert(0 && "Unhandled argument type!"); - } - ArgOffset += 4; // Each argument takes at least 4 bytes on the stack... - if (GPR_remaining > 0) { - GPR_remaining--; // uses up 2 GPRs - GPR_idx++; - } - } - - // If the function takes variable number of arguments, add a frame offset for - // the start of the first vararg value... this is used to expand - // llvm.va_start. - if (Fn.getFunctionType()->isVarArg()) - VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset); -} - - -/// SelectPHINodes - Insert machine code to generate phis. This is tricky -/// because we have to generate our sources into the source basic blocks, not -/// the current one. -/// -void PPC64ISel::SelectPHINodes() { - const TargetInstrInfo &TII = *TM.getInstrInfo(); - const Function &LF = *F->getFunction(); // The LLVM function... - for (Function::const_iterator I = LF.begin(), E = LF.end(); I != E; ++I) { - const BasicBlock *BB = I; - MachineBasicBlock &MBB = *MBBMap[I]; - - // Loop over all of the PHI nodes in the LLVM basic block... - MachineBasicBlock::iterator PHIInsertPoint = MBB.begin(); - for (BasicBlock::const_iterator I = BB->begin(); - PHINode *PN = const_cast(dyn_cast(I)); ++I) { - - // Create a new machine instr PHI node, and insert it. - unsigned PHIReg = getReg(*PN); - MachineInstr *PhiMI = BuildMI(MBB, PHIInsertPoint, - PPC::PHI, PN->getNumOperands(), PHIReg); - - // PHIValues - Map of blocks to incoming virtual registers. We use this - // so that we only initialize one incoming value for a particular block, - // even if the block has multiple entries in the PHI node. - // - std::map PHIValues; - - for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { - MachineBasicBlock *PredMBB = 0; - for (MachineBasicBlock::pred_iterator PI = MBB.pred_begin (), - PE = MBB.pred_end (); PI != PE; ++PI) - if (PN->getIncomingBlock(i) == (*PI)->getBasicBlock()) { - PredMBB = *PI; - break; - } - assert (PredMBB && "Couldn't find incoming machine-cfg edge for phi"); - - unsigned ValReg; - std::map::iterator EntryIt = - PHIValues.lower_bound(PredMBB); - - if (EntryIt != PHIValues.end() && EntryIt->first == PredMBB) { - // We already inserted an initialization of the register for this - // predecessor. Recycle it. - ValReg = EntryIt->second; - } else { - // Get the incoming value into a virtual register. - // - Value *Val = PN->getIncomingValue(i); - - // If this is a constant or GlobalValue, we may have to insert code - // into the basic block to compute it into a virtual register. - if ((isa(Val) && !isa(Val)) || - isa(Val)) { - // Simple constants get emitted at the end of the basic block, - // before any terminator instructions. We "know" that the code to - // move a constant into a register will never clobber any flags. - ValReg = getReg(Val, PredMBB, PredMBB->getFirstTerminator()); - } else { - // Because we don't want to clobber any values which might be in - // physical registers with the computation of this constant (which - // might be arbitrarily complex if it is a constant expression), - // just insert the computation at the top of the basic block. - MachineBasicBlock::iterator PI = PredMBB->begin(); - - // Skip over any PHI nodes though! - while (PI != PredMBB->end() && PI->getOpcode() == PPC::PHI) - ++PI; - - ValReg = getReg(Val, PredMBB, PI); - } - - // Remember that we inserted a value for this PHI for this predecessor - PHIValues.insert(EntryIt, std::make_pair(PredMBB, ValReg)); - } - - PhiMI->addRegOperand(ValReg); - PhiMI->addMachineBasicBlockOperand(PredMBB); - } - - // Now that we emitted all of the incoming values for the PHI node, make - // sure to reposition the InsertPoint after the PHI that we just added. - // This is needed because we might have inserted a constant into this - // block, right after the PHI's which is before the old insert point! - PHIInsertPoint = PhiMI; - ++PHIInsertPoint; - } - } -} - - -// canFoldSetCCIntoBranchOrSelect - Return the setcc instruction if we can fold -// it into the conditional branch or select instruction which is the only user -// of the cc instruction. This is the case if the conditional branch is the -// only user of the setcc, and if the setcc is in the same basic block as the -// conditional branch. -// -static SetCondInst *canFoldSetCCIntoBranchOrSelect(Value *V) { - if (SetCondInst *SCI = dyn_cast(V)) - if (SCI->hasOneUse()) { - Instruction *User = cast(SCI->use_back()); - if ((isa(User) || isa(User)) && - SCI->getParent() == User->getParent()) - return SCI; - } - return 0; -} - - -// canFoldGEPIntoLoadOrStore - Return the GEP instruction if we can fold it into -// the load or store instruction that is the only user of the GEP. -// -static GetElementPtrInst *canFoldGEPIntoLoadOrStore(Value *V) { - if (GetElementPtrInst *GEPI = dyn_cast(V)) - if (GEPI->hasOneUse()) { - Instruction *User = cast(GEPI->use_back()); - if (isa(User) && - GEPI->getParent() == User->getParent() && - User->getOperand(0) != GEPI && - User->getOperand(1) == GEPI) { - ++GEPFolds; - return GEPI; - } - if (isa(User) && - GEPI->getParent() == User->getParent() && - User->getOperand(0) == GEPI) { - ++GEPFolds; - return GEPI; - } - } - return 0; -} - - -// Return a fixed numbering for setcc instructions which does not depend on the -// order of the opcodes. -// -static unsigned getSetCCNumber(unsigned Opcode) { - switch (Opcode) { - default: assert(0 && "Unknown setcc instruction!"); - case Instruction::SetEQ: return 0; - case Instruction::SetNE: return 1; - case Instruction::SetLT: return 2; - case Instruction::SetGE: return 3; - case Instruction::SetGT: return 4; - case Instruction::SetLE: return 5; - } -} - -static unsigned getPPCOpcodeForSetCCNumber(unsigned Opcode) { - switch (Opcode) { - default: assert(0 && "Unknown setcc instruction!"); - case Instruction::SetEQ: return PPC::BEQ; - case Instruction::SetNE: return PPC::BNE; - case Instruction::SetLT: return PPC::BLT; - case Instruction::SetGE: return PPC::BGE; - case Instruction::SetGT: return PPC::BGT; - case Instruction::SetLE: return PPC::BLE; - } -} - -/// emitUCOM - emits an unordered FP compare. -void PPC64ISel::emitUCOM(MachineBasicBlock *MBB, MachineBasicBlock::iterator IP, - unsigned LHS, unsigned RHS) { - BuildMI(*MBB, IP, PPC::FCMPU, 2, PPC::CR0).addReg(LHS).addReg(RHS); -} - -/// EmitComparison - emits a comparison of the two operands, returning the -/// extended setcc code to use. The result is in CR0. -/// -unsigned PPC64ISel::EmitComparison(unsigned OpNum, Value *Op0, Value *Op1, - MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP) { - // The arguments are already supposed to be of the same type. - const Type *CompTy = Op0->getType(); - unsigned Class = getClassB(CompTy); - unsigned Op0r = getReg(Op0, MBB, IP); - - // Before we do a comparison, we have to make sure that we're truncating our - // registers appropriately. - if (Class == cByte) { - unsigned TmpReg = makeAnotherReg(CompTy); - if (CompTy->isSigned()) - BuildMI(*MBB, IP, PPC::EXTSB, 1, TmpReg).addReg(Op0r); - else - BuildMI(*MBB, IP, PPC::RLWINM, 4, TmpReg).addReg(Op0r).addImm(0) - .addImm(24).addImm(31); - Op0r = TmpReg; - } else if (Class == cShort) { - unsigned TmpReg = makeAnotherReg(CompTy); - if (CompTy->isSigned()) - BuildMI(*MBB, IP, PPC::EXTSH, 1, TmpReg).addReg(Op0r); - else - BuildMI(*MBB, IP, PPC::RLWINM, 4, TmpReg).addReg(Op0r).addImm(0) - .addImm(16).addImm(31); - Op0r = TmpReg; - } - - // Use crand for lt, gt and crandc for le, ge - unsigned CROpcode = (OpNum == 2 || OpNum == 4) ? PPC::CRAND : PPC::CRANDC; - unsigned Opcode = CompTy->isSigned() ? PPC::CMPW : PPC::CMPLW; - unsigned OpcodeImm = CompTy->isSigned() ? PPC::CMPWI : PPC::CMPLWI; - if (Class == cLong) { - Opcode = CompTy->isSigned() ? PPC::CMPD : PPC::CMPLD; - OpcodeImm = CompTy->isSigned() ? PPC::CMPDI : PPC::CMPLDI; - } - - // Special case handling of: cmp R, i - if (ConstantInt *CI = dyn_cast(Op1)) { - unsigned Op1v = CI->getRawValue() & 0xFFFF; - - // Treat compare like ADDI for the purposes of immediate suitability - if (canUseAsImmediateForOpcode(CI, 0)) { - BuildMI(*MBB, IP, OpcodeImm, 2, PPC::CR0).addReg(Op0r).addSImm(Op1v); - } else { - unsigned Op1r = getReg(Op1, MBB, IP); - BuildMI(*MBB, IP, Opcode, 2, PPC::CR0).addReg(Op0r).addReg(Op1r); - } - return OpNum; - } - - unsigned Op1r = getReg(Op1, MBB, IP); - - switch (Class) { - default: assert(0 && "Unknown type class!"); - case cByte: - case cShort: - case cInt: - case cLong: - BuildMI(*MBB, IP, Opcode, 2, PPC::CR0).addReg(Op0r).addReg(Op1r); - break; - - case cFP32: - case cFP64: - emitUCOM(MBB, IP, Op0r, Op1r); - break; - } - - return OpNum; -} - -/// visitSetCondInst - emit code to calculate the condition via -/// EmitComparison(), and possibly store a 0 or 1 to a register as a result -/// -void PPC64ISel::visitSetCondInst(SetCondInst &I) { - if (canFoldSetCCIntoBranchOrSelect(&I)) - return; - - unsigned DestReg = getReg(I); - unsigned OpNum = I.getOpcode(); - const Type *Ty = I.getOperand (0)->getType(); - - EmitComparison(OpNum, I.getOperand(0), I.getOperand(1), BB, BB->end()); - - unsigned Opcode = getPPCOpcodeForSetCCNumber(OpNum); - MachineBasicBlock *thisMBB = BB; - const BasicBlock *LLVM_BB = BB->getBasicBlock(); - ilist::iterator It = BB; - ++It; - - // thisMBB: - // ... - // cmpTY cr0, r1, r2 - // bCC copy1MBB - // b copy0MBB - - // FIXME: we wouldn't need copy0MBB (we could fold it into thisMBB) - // if we could insert other, non-terminator instructions after the - // bCC. But MBB->getFirstTerminator() can't understand this. - MachineBasicBlock *copy1MBB = new MachineBasicBlock(LLVM_BB); - F->getBasicBlockList().insert(It, copy1MBB); - BuildMI(BB, Opcode, 2).addReg(PPC::CR0).addMBB(copy1MBB); - MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB); - F->getBasicBlockList().insert(It, copy0MBB); - BuildMI(BB, PPC::B, 1).addMBB(copy0MBB); - MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB); - F->getBasicBlockList().insert(It, sinkMBB); - // Update machine-CFG edges - BB->addSuccessor(copy1MBB); - BB->addSuccessor(copy0MBB); - - // copy1MBB: - // %TrueValue = li 1 - // b sinkMBB - BB = copy1MBB; - unsigned TrueValue = makeAnotherReg(I.getType()); - BuildMI(BB, PPC::LI, 1, TrueValue).addSImm(1); - BuildMI(BB, PPC::B, 1).addMBB(sinkMBB); - // Update machine-CFG edges - BB->addSuccessor(sinkMBB); - - // copy0MBB: - // %FalseValue = li 0 - // fallthrough - BB = copy0MBB; - unsigned FalseValue = makeAnotherReg(I.getType()); - BuildMI(BB, PPC::LI, 1, FalseValue).addSImm(0); - // Update machine-CFG edges - BB->addSuccessor(sinkMBB); - - // sinkMBB: - // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, copy1MBB ] - // ... - BB = sinkMBB; - BuildMI(BB, PPC::PHI, 4, DestReg).addReg(FalseValue) - .addMBB(copy0MBB).addReg(TrueValue).addMBB(copy1MBB); -} - -void PPC64ISel::visitSelectInst(SelectInst &SI) { - unsigned DestReg = getReg(SI); - MachineBasicBlock::iterator MII = BB->end(); - emitSelectOperation(BB, MII, SI.getCondition(), SI.getTrueValue(), - SI.getFalseValue(), DestReg); -} - -/// emitSelect - Common code shared between visitSelectInst and the constant -/// expression support. -/// FIXME: this is most likely broken in one or more ways. Namely, PowerPC has -/// no select instruction. FSEL only works for comparisons against zero. -void PPC64ISel::emitSelectOperation(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Value *Cond, Value *TrueVal, - Value *FalseVal, unsigned DestReg) { - unsigned SelectClass = getClassB(TrueVal->getType()); - unsigned Opcode; - - // See if we can fold the setcc into the select instruction, or if we have - // to get the register of the Cond value - if (SetCondInst *SCI = canFoldSetCCIntoBranchOrSelect(Cond)) { - // We successfully folded the setcc into the select instruction. - unsigned OpNum = getSetCCNumber(SCI->getOpcode()); - OpNum = EmitComparison(OpNum, SCI->getOperand(0),SCI->getOperand(1),MBB,IP); - Opcode = getPPCOpcodeForSetCCNumber(SCI->getOpcode()); - } else { - unsigned CondReg = getReg(Cond, MBB, IP); - BuildMI(*MBB, IP, PPC::CMPI, 2, PPC::CR0).addReg(CondReg).addSImm(0); - Opcode = getPPCOpcodeForSetCCNumber(Instruction::SetNE); - } - - // thisMBB: - // ... - // cmpTY cr0, r1, r2 - // bCC copy1MBB - // b copy0MBB - - MachineBasicBlock *thisMBB = BB; - const BasicBlock *LLVM_BB = BB->getBasicBlock(); - ilist::iterator It = BB; - ++It; - - // FIXME: we wouldn't need copy0MBB (we could fold it into thisMBB) - // if we could insert other, non-terminator instructions after the - // bCC. But MBB->getFirstTerminator() can't understand this. - MachineBasicBlock *copy1MBB = new MachineBasicBlock(LLVM_BB); - F->getBasicBlockList().insert(It, copy1MBB); - BuildMI(BB, Opcode, 2).addReg(PPC::CR0).addMBB(copy1MBB); - MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB); - F->getBasicBlockList().insert(It, copy0MBB); - BuildMI(BB, PPC::B, 1).addMBB(copy0MBB); - MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB); - F->getBasicBlockList().insert(It, sinkMBB); - // Update machine-CFG edges - BB->addSuccessor(copy1MBB); - BB->addSuccessor(copy0MBB); - - // copy1MBB: - // %TrueValue = ... - // b sinkMBB - BB = copy1MBB; - unsigned TrueValue = getReg(TrueVal, BB, BB->begin()); - BuildMI(BB, PPC::B, 1).addMBB(sinkMBB); - // Update machine-CFG edges - BB->addSuccessor(sinkMBB); - - // copy0MBB: - // %FalseValue = ... - // fallthrough - BB = copy0MBB; - unsigned FalseValue = getReg(FalseVal, BB, BB->begin()); - // Update machine-CFG edges - BB->addSuccessor(sinkMBB); - - // sinkMBB: - // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, copy1MBB ] - // ... - BB = sinkMBB; - BuildMI(BB, PPC::PHI, 4, DestReg).addReg(FalseValue) - .addMBB(copy0MBB).addReg(TrueValue).addMBB(copy1MBB); - return; -} - - - -/// promote32 - Emit instructions to turn a narrow operand into a 32-bit-wide -/// operand, in the specified target register. -/// -void PPC64ISel::promote32(unsigned targetReg, const ValueRecord &VR) { - bool isUnsigned = VR.Ty->isUnsigned() || VR.Ty == Type::BoolTy; - - Value *Val = VR.Val; - const Type *Ty = VR.Ty; - if (Val) { - if (Constant *C = dyn_cast(Val)) { - Val = ConstantExpr::getCast(C, Type::IntTy); - if (isa(Val)) // Could not fold - Val = C; - else - Ty = Type::IntTy; // Folded! - } - - // If this is a simple constant, just emit a load directly to avoid the copy - if (ConstantInt *CI = dyn_cast(Val)) { - int TheVal = CI->getRawValue() & 0xFFFFFFFF; - - if (TheVal < 32768 && TheVal >= -32768) { - BuildMI(BB, PPC::LI, 1, targetReg).addSImm(TheVal); - } else { - unsigned TmpReg = makeAnotherReg(Type::IntTy); - BuildMI(BB, PPC::LIS, 1, TmpReg).addSImm(TheVal >> 16); - BuildMI(BB, PPC::ORI, 2, targetReg).addReg(TmpReg) - .addImm(TheVal & 0xFFFF); - } - return; - } - } - - // Make sure we have the register number for this value... - unsigned Reg = Val ? getReg(Val) : VR.Reg; - switch (getClassB(Ty)) { - case cByte: - // Extend value into target register (8->32) - if (isUnsigned) - BuildMI(BB, PPC::RLWINM, 4, targetReg).addReg(Reg).addZImm(0) - .addZImm(24).addZImm(31); - else - BuildMI(BB, PPC::EXTSB, 1, targetReg).addReg(Reg); - break; - case cShort: - // Extend value into target register (16->32) - if (isUnsigned) - BuildMI(BB, PPC::RLWINM, 4, targetReg).addReg(Reg).addZImm(0) - .addZImm(16).addZImm(31); - else - BuildMI(BB, PPC::EXTSH, 1, targetReg).addReg(Reg); - break; - case cInt: - case cLong: - // Move value into target register (32->32) - BuildMI(BB, PPC::OR, 2, targetReg).addReg(Reg).addReg(Reg); - break; - default: - assert(0 && "Unpromotable operand class in promote32"); - } -} - -/// visitReturnInst - implemented with BLR -/// -void PPC64ISel::visitReturnInst(ReturnInst &I) { - // Only do the processing if this is a non-void return - if (I.getNumOperands() > 0) { - Value *RetVal = I.getOperand(0); - switch (getClassB(RetVal->getType())) { - case cByte: // integral return values: extend or move into r3 and return - case cShort: - case cInt: - case cLong: - promote32(PPC::R3, ValueRecord(RetVal)); - break; - case cFP32: - case cFP64: { // Floats & Doubles: Return in f1 - unsigned RetReg = getReg(RetVal); - BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(RetReg); - break; - } - default: - visitInstruction(I); - } - } - BuildMI(BB, PPC::BLR, 1).addImm(1); -} - -// getBlockAfter - Return the basic block which occurs lexically after the -// specified one. -static inline BasicBlock *getBlockAfter(BasicBlock *BB) { - Function::iterator I = BB; ++I; // Get iterator to next block - return I != BB->getParent()->end() ? &*I : 0; -} - -/// visitBranchInst - Handle conditional and unconditional branches here. Note -/// that since code layout is frozen at this point, that if we are trying to -/// jump to a block that is the immediate successor of the current block, we can -/// just make a fall-through (but we don't currently). -/// -void PPC64ISel::visitBranchInst(BranchInst &BI) { - // Update machine-CFG edges - BB->addSuccessor(MBBMap[BI.getSuccessor(0)]); - if (BI.isConditional()) - BB->addSuccessor(MBBMap[BI.getSuccessor(1)]); - - BasicBlock *NextBB = getBlockAfter(BI.getParent()); // BB after current one - - if (!BI.isConditional()) { // Unconditional branch? - if (BI.getSuccessor(0) != NextBB) - BuildMI(BB, PPC::B, 1).addMBB(MBBMap[BI.getSuccessor(0)]); - return; - } - - // See if we can fold the setcc into the branch itself... - SetCondInst *SCI = canFoldSetCCIntoBranchOrSelect(BI.getCondition()); - if (SCI == 0) { - // Nope, cannot fold setcc into this branch. Emit a branch on a condition - // computed some other way... - unsigned condReg = getReg(BI.getCondition()); - BuildMI(BB, PPC::CMPLI, 3, PPC::CR0).addImm(0).addReg(condReg) - .addImm(0); - if (BI.getSuccessor(1) == NextBB) { - if (BI.getSuccessor(0) != NextBB) - BuildMI(BB, PPC::COND_BRANCH, 3).addReg(PPC::CR0).addImm(PPC::BNE) - .addMBB(MBBMap[BI.getSuccessor(0)]) - .addMBB(MBBMap[BI.getSuccessor(1)]); - } else { - BuildMI(BB, PPC::COND_BRANCH, 3).addReg(PPC::CR0).addImm(PPC::BEQ) - .addMBB(MBBMap[BI.getSuccessor(1)]) - .addMBB(MBBMap[BI.getSuccessor(0)]); - if (BI.getSuccessor(0) != NextBB) - BuildMI(BB, PPC::B, 1).addMBB(MBBMap[BI.getSuccessor(0)]); - } - return; - } - - unsigned OpNum = getSetCCNumber(SCI->getOpcode()); - unsigned Opcode = getPPCOpcodeForSetCCNumber(SCI->getOpcode()); - MachineBasicBlock::iterator MII = BB->end(); - OpNum = EmitComparison(OpNum, SCI->getOperand(0), SCI->getOperand(1), BB,MII); - - if (BI.getSuccessor(0) != NextBB) { - BuildMI(BB, PPC::COND_BRANCH, 3).addReg(PPC::CR0).addImm(Opcode) - .addMBB(MBBMap[BI.getSuccessor(0)]) - .addMBB(MBBMap[BI.getSuccessor(1)]); - if (BI.getSuccessor(1) != NextBB) - BuildMI(BB, PPC::B, 1).addMBB(MBBMap[BI.getSuccessor(1)]); - } else { - // Change to the inverse condition... - if (BI.getSuccessor(1) != NextBB) { - Opcode = PPC64InstrInfo::invertPPCBranchOpcode(Opcode); - BuildMI(BB, PPC::COND_BRANCH, 3).addReg(PPC::CR0).addImm(Opcode) - .addMBB(MBBMap[BI.getSuccessor(1)]) - .addMBB(MBBMap[BI.getSuccessor(0)]); - } - } -} - -/// doCall - This emits an abstract call instruction, setting up the arguments -/// and the return value as appropriate. For the actual function call itself, -/// it inserts the specified CallMI instruction into the stream. -/// -void PPC64ISel::doCall(const ValueRecord &Ret, MachineInstr *CallMI, - const std::vector &Args, bool isVarArg) { - // Count how many bytes are to be pushed on the stack, including the linkage - // area, and parameter passing area. - unsigned NumBytes = ParameterSaveAreaOffset; - unsigned ArgOffset = ParameterSaveAreaOffset; - - if (!Args.empty()) { - for (unsigned i = 0, e = Args.size(); i != e; ++i) - switch (getClassB(Args[i].Ty)) { - case cByte: case cShort: case cInt: - NumBytes += 4; break; - case cLong: - NumBytes += 8; break; - case cFP32: - NumBytes += 4; break; - case cFP64: - NumBytes += 8; break; - break; - default: assert(0 && "Unknown class!"); - } - - // Just to be safe, we'll always reserve the full argument passing space in - // case any called code gets funky on us. - if (NumBytes < ParameterSaveAreaOffset + MaxArgumentStackSpace) - NumBytes = ParameterSaveAreaOffset + MaxArgumentStackSpace; - - // Adjust the stack pointer for the new arguments... - // These functions are automatically eliminated by the prolog/epilog pass - BuildMI(BB, PPC::ADJCALLSTACKDOWN, 1).addImm(NumBytes); - - // Arguments go on the stack in reverse order, as specified by the ABI. - int GPR_remaining = 8, FPR_remaining = 13; - 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 - }; - - for (unsigned i = 0, e = Args.size(); i != e; ++i) { - unsigned ArgReg; - switch (getClassB(Args[i].Ty)) { - case cByte: - case cShort: - // Promote arg to 32 bits wide into a temporary register... - ArgReg = makeAnotherReg(Type::UIntTy); - promote32(ArgReg, Args[i]); - - // Reg or stack? - if (GPR_remaining > 0) { - BuildMI(BB, PPC::OR, 2, GPR[GPR_idx]).addReg(ArgReg) - .addReg(ArgReg); - CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use); - } - if (GPR_remaining <= 0 || isVarArg) { - BuildMI(BB, PPC::STW, 3).addReg(ArgReg).addSImm(ArgOffset) - .addReg(PPC::R1); - } - break; - case cInt: - ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg; - - // Reg or stack? - if (GPR_remaining > 0) { - BuildMI(BB, PPC::OR, 2, GPR[GPR_idx]).addReg(ArgReg) - .addReg(ArgReg); - CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use); - } - if (GPR_remaining <= 0 || isVarArg) { - BuildMI(BB, PPC::STW, 3).addReg(ArgReg).addSImm(ArgOffset) - .addReg(PPC::R1); - } - break; - case cLong: - ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg; - - // Reg or stack? - if (GPR_remaining > 0) { - BuildMI(BB, PPC::OR, 2, GPR[GPR_idx]).addReg(ArgReg) - .addReg(ArgReg); - CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use); - } - if (GPR_remaining <= 0 || isVarArg) { - BuildMI(BB, PPC::STD, 3).addReg(ArgReg).addSImm(ArgOffset) - .addReg(PPC::R1); - } - ArgOffset += 4; // 8 byte entry, not 4. - break; - case cFP32: - ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg; - // Reg or stack? - if (FPR_remaining > 0) { - BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgReg); - CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use); - FPR_remaining--; - FPR_idx++; - - // If this is a vararg function, and there are GPRs left, also - // pass the float in an int. Otherwise, put it on the stack. - if (isVarArg) { - BuildMI(BB, PPC::STFS, 3).addReg(ArgReg).addSImm(ArgOffset) - .addReg(PPC::R1); - if (GPR_remaining > 0) { - BuildMI(BB, PPC::LWZ, 2, GPR[GPR_idx]) - .addSImm(ArgOffset).addReg(ArgReg); - CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use); - } - } - } else { - BuildMI(BB, PPC::STFS, 3).addReg(ArgReg).addSImm(ArgOffset) - .addReg(PPC::R1); - } - break; - case cFP64: - ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg; - // Reg or stack? - if (FPR_remaining > 0) { - BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgReg); - CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use); - FPR_remaining--; - FPR_idx++; - // For vararg functions, must pass doubles via int regs as well - if (isVarArg) { - BuildMI(BB, PPC::STFD, 3).addReg(ArgReg).addSImm(ArgOffset) - .addReg(PPC::R1); - - if (GPR_remaining > 0) { - BuildMI(BB, PPC::LD, 2, GPR[GPR_idx]).addSImm(ArgOffset) - .addReg(PPC::R1); - CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use); - } - } - } else { - BuildMI(BB, PPC::STFD, 3).addReg(ArgReg).addSImm(ArgOffset) - .addReg(PPC::R1); - } - // Doubles use 8 bytes - ArgOffset += 4; - break; - - default: assert(0 && "Unknown class!"); - } - ArgOffset += 4; - GPR_remaining--; - GPR_idx++; - } - } else { - BuildMI(BB, PPC::ADJCALLSTACKDOWN, 1).addImm(0); - } - - BuildMI(BB, PPC::IMPLICIT_DEF, 0, PPC::LR); - BB->push_back(CallMI); - BuildMI(BB, PPC::NOP, 0); - - // These functions are automatically eliminated by the prolog/epilog pass - BuildMI(BB, PPC::ADJCALLSTACKUP, 1).addImm(NumBytes); - - // If there is a return value, scavenge the result from the location the call - // leaves it in... - // - if (Ret.Ty != Type::VoidTy) { - unsigned DestClass = getClassB(Ret.Ty); - switch (DestClass) { - case cByte: - case cShort: - case cInt: - case cLong: - // Integral results are in r3 - BuildMI(BB, PPC::OR, 2, Ret.Reg).addReg(PPC::R3).addReg(PPC::R3); - break; - case cFP32: // Floating-point return values live in f1 - case cFP64: - BuildMI(BB, PPC::FMR, 1, Ret.Reg).addReg(PPC::F1); - break; - default: assert(0 && "Unknown class!"); - } - } -} - - -/// visitCallInst - Push args on stack and do a procedure call instruction. -void PPC64ISel::visitCallInst(CallInst &CI) { - MachineInstr *TheCall; - Function *F = CI.getCalledFunction(); - if (F) { - // Is it an intrinsic function call? - if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) { - visitIntrinsicCall(ID, CI); // Special intrinsics are not handled here - return; - } - // Emit a CALL instruction with PC-relative displacement. - TheCall = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(F, true); - } else { // Emit an indirect call through the CTR - unsigned Reg = getReg(CI.getCalledValue()); - BuildMI(BB, PPC::MTCTR, 1).addReg(Reg); - TheCall = BuildMI(PPC::CALLindirect, 2).addZImm(20).addZImm(0); - } - - std::vector Args; - for (unsigned i = 1, e = CI.getNumOperands(); i != e; ++i) - Args.push_back(ValueRecord(CI.getOperand(i))); - - unsigned DestReg = CI.getType() != Type::VoidTy ? getReg(CI) : 0; - bool isVarArg = F ? F->getFunctionType()->isVarArg() : true; - doCall(ValueRecord(DestReg, CI.getType()), TheCall, Args, isVarArg); -} - - -/// dyncastIsNan - Return the operand of an isnan operation if this is an isnan. -/// -static Value *dyncastIsNan(Value *V) { - if (CallInst *CI = dyn_cast(V)) - if (Function *F = CI->getCalledFunction()) - if (F->getIntrinsicID() == Intrinsic::isunordered) - return CI->getOperand(1); - return 0; -} - -/// isOnlyUsedByUnorderedComparisons - Return true if this value is only used by -/// or's whos operands are all calls to the isnan predicate. -static bool isOnlyUsedByUnorderedComparisons(Value *V) { - assert(dyncastIsNan(V) && "The value isn't an isnan call!"); - - // Check all uses, which will be or's of isnans if this predicate is true. - for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ - Instruction *I = cast(*UI); - if (I->getOpcode() != Instruction::Or) return false; - if (I->getOperand(0) != V && !dyncastIsNan(I->getOperand(0))) return false; - if (I->getOperand(1) != V && !dyncastIsNan(I->getOperand(1))) return false; - } - - return true; -} - -/// LowerUnknownIntrinsicFunctionCalls - This performs a prepass over the -/// function, lowering any calls to unknown intrinsic functions into the -/// equivalent LLVM code. -/// -void PPC64ISel::LowerUnknownIntrinsicFunctionCalls(Function &F) { - for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) - for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) - if (CallInst *CI = dyn_cast(I++)) - if (Function *F = CI->getCalledFunction()) - switch (F->getIntrinsicID()) { - case Intrinsic::not_intrinsic: - case Intrinsic::vastart: - case Intrinsic::vacopy: - case Intrinsic::vaend: - case Intrinsic::returnaddress: - case Intrinsic::frameaddress: - // FIXME: should lower these ourselves - // case Intrinsic::isunordered: - // case Intrinsic::memcpy: -> doCall(). system memcpy almost - // guaranteed to be faster than anything we generate ourselves - // We directly implement these intrinsics - break; - case Intrinsic::readio: { - // On PPC, memory operations are in-order. Lower this intrinsic - // into a volatile load. - LoadInst * LI = new LoadInst(CI->getOperand(1), "", true, CI); - CI->replaceAllUsesWith(LI); - BB->getInstList().erase(CI); - break; - } - case Intrinsic::writeio: { - // On PPC, memory operations are in-order. Lower this intrinsic - // into a volatile store. - StoreInst *SI = new StoreInst(CI->getOperand(1), - CI->getOperand(2), true, CI); - CI->replaceAllUsesWith(SI); - BB->getInstList().erase(CI); - break; - } - default: - // All other intrinsic calls we must lower. - Instruction *Before = CI->getPrev(); - TM.getIntrinsicLowering().LowerIntrinsicCall(CI); - if (Before) { // Move iterator to instruction after call - I = Before; ++I; - } else { - I = BB->begin(); - } - } -} - -void PPC64ISel::visitIntrinsicCall(Intrinsic::ID ID, CallInst &CI) { - unsigned TmpReg1, TmpReg2, TmpReg3; - switch (ID) { - case Intrinsic::vastart: - // Get the address of the first vararg value... - TmpReg1 = getReg(CI); - addFrameReference(BuildMI(BB, PPC::ADDI, 2, TmpReg1), VarArgsFrameIndex, - 0, false); - return; - - case Intrinsic::vacopy: - TmpReg1 = getReg(CI); - TmpReg2 = getReg(CI.getOperand(1)); - BuildMI(BB, PPC::OR, 2, TmpReg1).addReg(TmpReg2).addReg(TmpReg2); - return; - case Intrinsic::vaend: return; - - case Intrinsic::returnaddress: - TmpReg1 = getReg(CI); - if (cast(CI.getOperand(1))->isNullValue()) { - MachineFrameInfo *MFI = F->getFrameInfo(); - unsigned NumBytes = MFI->getStackSize(); - - BuildMI(BB, PPC::LWZ, 2, TmpReg1).addSImm(NumBytes+8) - .addReg(PPC::R1); - } else { - // Values other than zero are not implemented yet. - BuildMI(BB, PPC::LI, 1, TmpReg1).addSImm(0); - } - return; - - case Intrinsic::frameaddress: - TmpReg1 = getReg(CI); - if (cast(CI.getOperand(1))->isNullValue()) { - BuildMI(BB, PPC::OR, 2, TmpReg1).addReg(PPC::R1).addReg(PPC::R1); - } else { - // Values other than zero are not implemented yet. - BuildMI(BB, PPC::LI, 1, TmpReg1).addSImm(0); - } - return; - -#if 0 - // This may be useful for supporting isunordered - case Intrinsic::isnan: - // If this is only used by 'isunordered' style comparisons, don't emit it. - if (isOnlyUsedByUnorderedComparisons(&CI)) return; - TmpReg1 = getReg(CI.getOperand(1)); - emitUCOM(BB, BB->end(), TmpReg1, TmpReg1); - TmpReg2 = makeAnotherReg(Type::IntTy); - BuildMI(BB, PPC::MFCR, TmpReg2); - TmpReg3 = getReg(CI); - BuildMI(BB, PPC::RLWINM, 4, TmpReg3).addReg(TmpReg2).addImm(4).addImm(31).addImm(31); - return; -#endif - - default: assert(0 && "Error: unknown intrinsics should have been lowered!"); - } -} - -/// visitSimpleBinary - Implement simple binary operators for integral types... -/// OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or, 4 for -/// Xor. -/// -void PPC64ISel::visitSimpleBinary(BinaryOperator &B, unsigned OperatorClass) { - unsigned DestReg = getReg(B); - MachineBasicBlock::iterator MI = BB->end(); - Value *Op0 = B.getOperand(0), *Op1 = B.getOperand(1); - unsigned Class = getClassB(B.getType()); - - emitSimpleBinaryOperation(BB, MI, Op0, Op1, OperatorClass, DestReg); -} - -/// emitBinaryFPOperation - This method handles emission of floating point -/// Add (0), Sub (1), Mul (2), and Div (3) operations. -void PPC64ISel::emitBinaryFPOperation(MachineBasicBlock *BB, - MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, - unsigned OperatorClass, unsigned DestReg){ - - static const unsigned OpcodeTab[][4] = { - { PPC::FADDS, PPC::FSUBS, PPC::FMULS, PPC::FDIVS }, // Float - { PPC::FADD, PPC::FSUB, PPC::FMUL, PPC::FDIV }, // Double - }; - - // Special case: R1 = op , R2 - if (ConstantFP *Op0C = dyn_cast(Op0)) - if (Op0C->isExactlyValue(-0.0) && OperatorClass == 1) { - // -0.0 - X === -X - unsigned op1Reg = getReg(Op1, BB, IP); - BuildMI(*BB, IP, PPC::FNEG, 1, DestReg).addReg(op1Reg); - return; - } - - unsigned Opcode = OpcodeTab[Op0->getType() == Type::DoubleTy][OperatorClass]; - unsigned Op0r = getReg(Op0, BB, IP); - unsigned Op1r = getReg(Op1, BB, IP); - BuildMI(*BB, IP, Opcode, 2, DestReg).addReg(Op0r).addReg(Op1r); -} - -/// emitSimpleBinaryOperation - Implement simple binary operators for integral -/// types... OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for -/// Or, 4 for Xor. -/// -/// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary -/// and constant expression support. -/// -void PPC64ISel::emitSimpleBinaryOperation(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, - unsigned OperatorClass, - unsigned DestReg) { - unsigned Class = getClassB(Op0->getType()); - - // Arithmetic and Bitwise operators - static const unsigned OpcodeTab[] = { - PPC::ADD, PPC::SUB, PPC::AND, PPC::OR, PPC::XOR - }; - // FIXME: Convert this to the version from PPC32ISel - static const unsigned ImmOpcodeTab[] = { - PPC::ADDI, PPC::ADDI, PPC::ANDIo, PPC::ORI, PPC::XORI - }; - static const unsigned RImmOpcodeTab[] = { - PPC::ADDI, PPC::SUBFIC, PPC::ANDIo, PPC::ORI, PPC::XORI - }; - - if (Class == cFP32 || Class == cFP64) { - assert(OperatorClass < 2 && "No logical ops for FP!"); - emitBinaryFPOperation(MBB, IP, Op0, Op1, OperatorClass, DestReg); - return; - } - - if (Op0->getType() == Type::BoolTy) { - if (OperatorClass == 3) - // If this is an or of two isnan's, emit an FP comparison directly instead - // of or'ing two isnan's together. - if (Value *LHS = dyncastIsNan(Op0)) - if (Value *RHS = dyncastIsNan(Op1)) { - unsigned Op0Reg = getReg(RHS, MBB, IP), Op1Reg = getReg(LHS, MBB, IP); - unsigned TmpReg = makeAnotherReg(Type::IntTy); - emitUCOM(MBB, IP, Op0Reg, Op1Reg); - BuildMI(*MBB, IP, PPC::MFCR, TmpReg); - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(TmpReg).addImm(4) - .addImm(31).addImm(31); - return; - } - } - - // Special case: op , Reg - if (ConstantInt *CI = dyn_cast(Op0)) { - // sub 0, X -> subfic - if (OperatorClass == 1 && canUseAsImmediateForOpcode(CI, 0)) { - unsigned Op1r = getReg(Op1, MBB, IP); - int imm = CI->getRawValue() & 0xFFFF; - BuildMI(*MBB, IP, PPC::SUBFIC, 2, DestReg).addReg(Op1r).addSImm(imm); - return; - } - - // If it is easy to do, swap the operands and emit an immediate op - if (Class != cLong && OperatorClass != 1 && - canUseAsImmediateForOpcode(CI, OperatorClass)) { - unsigned Op1r = getReg(Op1, MBB, IP); - int imm = CI->getRawValue() & 0xFFFF; - - if (OperatorClass < 2) - BuildMI(*MBB, IP, RImmOpcodeTab[OperatorClass], 2, DestReg).addReg(Op1r) - .addSImm(imm); - else - BuildMI(*MBB, IP, RImmOpcodeTab[OperatorClass], 2, DestReg).addReg(Op1r) - .addZImm(imm); - return; - } - } - - // Special case: op Reg, - if (ConstantInt *Op1C = dyn_cast(Op1)) { - unsigned Op0r = getReg(Op0, MBB, IP); - - // xor X, -1 -> not X - if (OperatorClass == 4 && Op1C->isAllOnesValue()) { - BuildMI(*MBB, IP, PPC::NOR, 2, DestReg).addReg(Op0r).addReg(Op0r); - return; - } - - if (canUseAsImmediateForOpcode(Op1C, OperatorClass)) { - int immediate = Op1C->getRawValue() & 0xFFFF; - - if (OperatorClass < 2) - BuildMI(*MBB, IP, ImmOpcodeTab[OperatorClass], 2,DestReg).addReg(Op0r) - .addSImm(immediate); - else - BuildMI(*MBB, IP, ImmOpcodeTab[OperatorClass], 2,DestReg).addReg(Op0r) - .addZImm(immediate); - } else { - unsigned Op1r = getReg(Op1, MBB, IP); - BuildMI(*MBB, IP, OpcodeTab[OperatorClass], 2, DestReg).addReg(Op0r) - .addReg(Op1r); - } - return; - } - - // We couldn't generate an immediate variant of the op, load both halves into - // registers and emit the appropriate opcode. - unsigned Op0r = getReg(Op0, MBB, IP); - unsigned Op1r = getReg(Op1, MBB, IP); - unsigned Opcode = OpcodeTab[OperatorClass]; - BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addReg(Op1r); -} - -// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It -// returns zero when the input is not exactly a power of two. -static unsigned ExactLog2(unsigned Val) { - if (Val == 0 || (Val & (Val-1))) return 0; - unsigned Count = 0; - while (Val != 1) { - Val >>= 1; - ++Count; - } - return Count; -} - -/// doMultiply - Emit appropriate instructions to multiply together the -/// Values Op0 and Op1, and put the result in DestReg. -/// -void PPC64ISel::doMultiply(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - unsigned DestReg, Value *Op0, Value *Op1) { - unsigned Class0 = getClass(Op0->getType()); - unsigned Class1 = getClass(Op1->getType()); - - unsigned Op0r = getReg(Op0, MBB, IP); - unsigned Op1r = getReg(Op1, MBB, IP); - - // 64 x 64 -> 64 - if (Class0 == cLong && Class1 == cLong) { - BuildMI(*MBB, IP, PPC::MULLD, 2, DestReg).addReg(Op0r).addReg(Op1r); - return; - } - - // 64 x 32 or less, promote 32 to 64 and do a 64 x 64 - if (Class0 == cLong && Class1 <= cInt) { - // FIXME: CLEAR or SIGN EXTEND Op1 - BuildMI(*MBB, IP, PPC::MULLD, 2, DestReg).addReg(Op0r).addReg(Op1r); - return; - } - - // 32 x 32 -> 32 - if (Class0 <= cInt && Class1 <= cInt) { - BuildMI(*MBB, IP, PPC::MULLW, 2, DestReg).addReg(Op0r).addReg(Op1r); - return; - } - - assert(0 && "doMultiply cannot operate on unknown type!"); -} - -/// doMultiplyConst - This method will multiply the value in Op0 by the -/// value of the ContantInt *CI -void PPC64ISel::doMultiplyConst(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - unsigned DestReg, Value *Op0, ConstantInt *CI) { - unsigned Class = getClass(Op0->getType()); - - // Mul op0, 0 ==> 0 - if (CI->isNullValue()) { - BuildMI(*MBB, IP, PPC::LI, 1, DestReg).addSImm(0); - return; - } - - // Mul op0, 1 ==> op0 - if (CI->equalsInt(1)) { - unsigned Op0r = getReg(Op0, MBB, IP); - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(Op0r).addReg(Op0r); - return; - } - - // If the element size is exactly a power of 2, use a shift to get it. - if (unsigned Shift = ExactLog2(CI->getRawValue())) { - ConstantUInt *ShiftCI = ConstantUInt::get(Type::UByteTy, Shift); - emitShiftOperation(MBB, IP, Op0, ShiftCI, true, Op0->getType(), DestReg); - return; - } - - // If 32 bits or less and immediate is in right range, emit mul by immediate - if (Class == cByte || Class == cShort || Class == cInt) { - if (canUseAsImmediateForOpcode(CI, 0)) { - unsigned Op0r = getReg(Op0, MBB, IP); - unsigned imm = CI->getRawValue() & 0xFFFF; - BuildMI(*MBB, IP, PPC::MULLI, 2, DestReg).addReg(Op0r).addSImm(imm); - return; - } - } - - doMultiply(MBB, IP, DestReg, Op0, CI); -} - -void PPC64ISel::visitMul(BinaryOperator &I) { - unsigned ResultReg = getReg(I); - - Value *Op0 = I.getOperand(0); - Value *Op1 = I.getOperand(1); - - MachineBasicBlock::iterator IP = BB->end(); - emitMultiply(BB, IP, Op0, Op1, ResultReg); -} - -void PPC64ISel::emitMultiply(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, unsigned DestReg) { - TypeClass Class = getClass(Op0->getType()); - - switch (Class) { - case cByte: - case cShort: - case cInt: - case cLong: - if (ConstantInt *CI = dyn_cast(Op1)) { - doMultiplyConst(MBB, IP, DestReg, Op0, CI); - } else { - doMultiply(MBB, IP, DestReg, Op0, Op1); - } - return; - case cFP32: - case cFP64: - emitBinaryFPOperation(MBB, IP, Op0, Op1, 2, DestReg); - return; - break; - } -} - - -/// visitDivRem - Handle division and remainder instructions... these -/// instruction both require the same instructions to be generated, they just -/// select the result from a different register. Note that both of these -/// instructions work differently for signed and unsigned operands. -/// -void PPC64ISel::visitDivRem(BinaryOperator &I) { - unsigned ResultReg = getReg(I); - Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); - - MachineBasicBlock::iterator IP = BB->end(); - emitDivRemOperation(BB, IP, Op0, Op1, I.getOpcode() == Instruction::Div, - ResultReg); -} - -void PPC64ISel::emitDivRemOperation(MachineBasicBlock *BB, - MachineBasicBlock::iterator IP, - Value *Op0, Value *Op1, bool isDiv, - unsigned ResultReg) { - const Type *Ty = Op0->getType(); - unsigned Class = getClass(Ty); - switch (Class) { - case cFP32: - if (isDiv) { - // Floating point divide... - emitBinaryFPOperation(BB, IP, Op0, Op1, 3, ResultReg); - return; - } else { - // Floating point remainder via fmodf(float x, float y); - unsigned Op0Reg = getReg(Op0, BB, IP); - unsigned Op1Reg = getReg(Op1, BB, IP); - MachineInstr *TheCall = - BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(fmodfFn, true); - std::vector Args; - Args.push_back(ValueRecord(Op0Reg, Type::FloatTy)); - Args.push_back(ValueRecord(Op1Reg, Type::FloatTy)); - doCall(ValueRecord(ResultReg, Type::FloatTy), TheCall, Args, false); - } - return; - case cFP64: - if (isDiv) { - // Floating point divide... - emitBinaryFPOperation(BB, IP, Op0, Op1, 3, ResultReg); - return; - } else { - // Floating point remainder via fmod(double x, double y); - unsigned Op0Reg = getReg(Op0, BB, IP); - unsigned Op1Reg = getReg(Op1, BB, IP); - MachineInstr *TheCall = - BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(fmodFn, true); - std::vector Args; - Args.push_back(ValueRecord(Op0Reg, Type::DoubleTy)); - Args.push_back(ValueRecord(Op1Reg, Type::DoubleTy)); - doCall(ValueRecord(ResultReg, Type::DoubleTy), TheCall, Args, false); - } - return; - case cLong: case cByte: case cShort: case cInt: - break; // Small integrals, handled below... - default: assert(0 && "Unknown class!"); - } - - // Special case signed division by power of 2. - if (isDiv) - if (ConstantSInt *CI = dyn_cast(Op1)) { - assert(Class != cLong && "This doesn't handle 64-bit divides!"); - int V = CI->getValue(); - - if (V == 1) { // X /s 1 => X - unsigned Op0Reg = getReg(Op0, BB, IP); - BuildMI(*BB, IP, PPC::OR, 2, ResultReg).addReg(Op0Reg).addReg(Op0Reg); - return; - } - - if (V == -1) { // X /s -1 => -X - unsigned Op0Reg = getReg(Op0, BB, IP); - BuildMI(*BB, IP, PPC::NEG, 1, ResultReg).addReg(Op0Reg); - return; - } - - unsigned log2V = ExactLog2(V); - if (log2V != 0 && Ty->isSigned()) { - unsigned Op0Reg = getReg(Op0, BB, IP); - unsigned TmpReg = makeAnotherReg(Op0->getType()); - unsigned Opcode = Class == cLong ? PPC::SRADI : PPC::SRAWI; - - BuildMI(*BB, IP, Opcode, 2, TmpReg).addReg(Op0Reg).addImm(log2V); - BuildMI(*BB, IP, PPC::ADDZE, 1, ResultReg).addReg(TmpReg); - return; - } - } - - static const unsigned DivOpcodes[] = - { PPC::DIVWU, PPC::DIVW, PPC::DIVDU, PPC::DIVD }; - - unsigned Op0Reg = getReg(Op0, BB, IP); - unsigned Op1Reg = getReg(Op1, BB, IP); - unsigned Opcode = DivOpcodes[2*(Class == cLong) + Ty->isSigned()]; - - if (isDiv) { - BuildMI(*BB, IP, Opcode, 2, ResultReg).addReg(Op0Reg).addReg(Op1Reg); - } else { // Remainder - unsigned TmpReg1 = makeAnotherReg(Op0->getType()); - unsigned TmpReg2 = makeAnotherReg(Op0->getType()); - unsigned MulOpcode = Class == cLong ? PPC::MULLD : PPC::MULLW; - - BuildMI(*BB, IP, Opcode, 2, TmpReg1).addReg(Op0Reg).addReg(Op1Reg); - BuildMI(*BB, IP, MulOpcode, 2, TmpReg2).addReg(TmpReg1).addReg(Op1Reg); - BuildMI(*BB, IP, PPC::SUBF, 2, ResultReg).addReg(TmpReg2).addReg(Op0Reg); - } -} - - -/// Shift instructions: 'shl', 'sar', 'shr' - Some special cases here -/// for constant immediate shift values, and for constant immediate -/// shift values equal to 1. Even the general case is sort of special, -/// because the shift amount has to be in CL, not just any old register. -/// -void PPC64ISel::visitShiftInst(ShiftInst &I) { - MachineBasicBlock::iterator IP = BB->end(); - emitShiftOperation(BB, IP, I.getOperand(0), I.getOperand(1), - I.getOpcode() == Instruction::Shl, I.getType(), - getReg(I)); -} - -/// emitShiftOperation - Common code shared between visitShiftInst and -/// constant expression support. -/// -void PPC64ISel::emitShiftOperation(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Value *Op, Value *ShiftAmount, - bool isLeftShift, const Type *ResultTy, - unsigned DestReg) { - unsigned SrcReg = getReg (Op, MBB, IP); - bool isSigned = ResultTy->isSigned (); - unsigned Class = getClass (ResultTy); - - // Longs, as usual, are handled specially... - if (Class == cLong) { - // If we have a constant shift, we can generate much more efficient code - // than otherwise... - // - if (ConstantUInt *CUI = dyn_cast(ShiftAmount)) { - unsigned Amount = CUI->getValue(); - assert(Amount < 64 && "Invalid immediate shift amount!"); - if (isLeftShift) { - BuildMI(*MBB, IP, PPC::RLDICR, 3, DestReg).addReg(SrcReg).addImm(Amount) - .addImm(63-Amount); - } else { - if (isSigned) { - BuildMI(*MBB, IP, PPC::SRADI, 2, DestReg).addReg(SrcReg) - .addImm(Amount); - } else { - BuildMI(*MBB, IP, PPC::RLDICL, 3, DestReg).addReg(SrcReg) - .addImm(64-Amount).addImm(Amount); - } - } - } else { - unsigned ShiftReg = getReg (ShiftAmount, MBB, IP); - - if (isLeftShift) { - BuildMI(*MBB, IP, PPC::SLD, 2, DestReg).addReg(SrcReg).addReg(ShiftReg); - } else { - unsigned Opcode = (isSigned) ? PPC::SRAD : PPC::SRD; - BuildMI(*MBB, IP, Opcode, DestReg).addReg(SrcReg).addReg(ShiftReg); - } - } - return; - } - - if (ConstantUInt *CUI = dyn_cast(ShiftAmount)) { - // The shift amount is constant, guaranteed to be a ubyte. Get its value. - assert(CUI->getType() == Type::UByteTy && "Shift amount not a ubyte?"); - unsigned Amount = CUI->getValue(); - - if (isLeftShift) { - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg) - .addImm(Amount).addImm(0).addImm(31-Amount); - } else { - if (isSigned) { - BuildMI(*MBB, IP, PPC::SRAWI,2,DestReg).addReg(SrcReg).addImm(Amount); - } else { - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg) - .addImm(32-Amount).addImm(Amount).addImm(31); - } - } - } else { // The shift amount is non-constant. - unsigned ShiftAmountReg = getReg(ShiftAmount, MBB, IP); - - if (isLeftShift) { - BuildMI(*MBB, IP, PPC::SLW, 2, DestReg).addReg(SrcReg) - .addReg(ShiftAmountReg); - } else { - BuildMI(*MBB, IP, isSigned ? PPC::SRAW : PPC::SRW, 2, DestReg) - .addReg(SrcReg).addReg(ShiftAmountReg); - } - } -} - - -/// visitLoadInst - Implement LLVM load instructions. Pretty straightforward -/// mapping of LLVM classes to PPC load instructions, with the exception of -/// signed byte loads, which need a sign extension following them. -/// -void PPC64ISel::visitLoadInst(LoadInst &I) { - // Immediate opcodes, for reg+imm addressing - static const unsigned ImmOpcodes[] = { - PPC::LBZ, PPC::LHZ, PPC::LWZ, - PPC::LFS, PPC::LFD, PPC::LWZ - }; - // Indexed opcodes, for reg+reg addressing - static const unsigned IdxOpcodes[] = { - PPC::LBZX, PPC::LHZX, PPC::LWZX, - PPC::LFSX, PPC::LFDX, PPC::LWZX - }; - - unsigned Class = getClassB(I.getType()); - unsigned ImmOpcode = ImmOpcodes[Class]; - unsigned IdxOpcode = IdxOpcodes[Class]; - unsigned DestReg = getReg(I); - Value *SourceAddr = I.getOperand(0); - - if (Class == cShort && I.getType()->isSigned()) ImmOpcode = PPC::LHA; - if (Class == cShort && I.getType()->isSigned()) IdxOpcode = PPC::LHAX; - - if (AllocaInst *AI = dyn_castFixedAlloca(SourceAddr)) { - unsigned FI = getFixedSizedAllocaFI(AI); - if (Class == cByte && I.getType()->isSigned()) { - unsigned TmpReg = makeAnotherReg(I.getType()); - addFrameReference(BuildMI(BB, ImmOpcode, 2, TmpReg), FI); - BuildMI(BB, PPC::EXTSB, 1, DestReg).addReg(TmpReg); - } else { - addFrameReference(BuildMI(BB, ImmOpcode, 2, DestReg), FI); - } - return; - } - - // If this load is the only use of the GEP instruction that is its address, - // then we can fold the GEP directly into the load instruction. - // emitGEPOperation with a second to last arg of 'true' will place the - // base register for the GEP into baseReg, and the constant offset from that - // into offset. If the offset fits in 16 bits, we can emit a reg+imm store - // otherwise, we copy the offset into another reg, and use reg+reg addressing. - if (GetElementPtrInst *GEPI = canFoldGEPIntoLoadOrStore(SourceAddr)) { - unsigned baseReg = getReg(GEPI); - unsigned pendingAdd; - ConstantSInt *offset; - - emitGEPOperation(BB, BB->end(), GEPI->getOperand(0), GEPI->op_begin()+1, - GEPI->op_end(), baseReg, true, &offset, &pendingAdd); - - if (pendingAdd == 0 && Class != cLong && - canUseAsImmediateForOpcode(offset, 0)) { - if (Class == cByte && I.getType()->isSigned()) { - unsigned TmpReg = makeAnotherReg(I.getType()); - BuildMI(BB, ImmOpcode, 2, TmpReg).addSImm(offset->getValue()) - .addReg(baseReg); - BuildMI(BB, PPC::EXTSB, 1, DestReg).addReg(TmpReg); - } else { - BuildMI(BB, ImmOpcode, 2, DestReg).addSImm(offset->getValue()) - .addReg(baseReg); - } - return; - } - - unsigned indexReg = (pendingAdd != 0) ? pendingAdd : getReg(offset); - - if (Class == cByte && I.getType()->isSigned()) { - unsigned TmpReg = makeAnotherReg(I.getType()); - BuildMI(BB, IdxOpcode, 2, TmpReg).addReg(indexReg).addReg(baseReg); - BuildMI(BB, PPC::EXTSB, 1, DestReg).addReg(TmpReg); - } else { - BuildMI(BB, IdxOpcode, 2, DestReg).addReg(indexReg).addReg(baseReg); - } - return; - } - - // The fallback case, where the load was from a source that could not be - // folded into the load instruction. - unsigned SrcAddrReg = getReg(SourceAddr); - - if (Class == cByte && I.getType()->isSigned()) { - unsigned TmpReg = makeAnotherReg(I.getType()); - BuildMI(BB, ImmOpcode, 2, TmpReg).addSImm(0).addReg(SrcAddrReg); - BuildMI(BB, PPC::EXTSB, 1, DestReg).addReg(TmpReg); - } else { - BuildMI(BB, ImmOpcode, 2, DestReg).addSImm(0).addReg(SrcAddrReg); - } -} - -/// visitStoreInst - Implement LLVM store instructions -/// -void PPC64ISel::visitStoreInst(StoreInst &I) { - // Immediate opcodes, for reg+imm addressing - static const unsigned ImmOpcodes[] = { - PPC::STB, PPC::STH, PPC::STW, - PPC::STFS, PPC::STFD, PPC::STW - }; - // Indexed opcodes, for reg+reg addressing - static const unsigned IdxOpcodes[] = { - PPC::STBX, PPC::STHX, PPC::STWX, - PPC::STFSX, PPC::STFDX, PPC::STWX - }; - - Value *SourceAddr = I.getOperand(1); - const Type *ValTy = I.getOperand(0)->getType(); - unsigned Class = getClassB(ValTy); - unsigned ImmOpcode = ImmOpcodes[Class]; - unsigned IdxOpcode = IdxOpcodes[Class]; - unsigned ValReg = getReg(I.getOperand(0)); - - // If this store is the only use of the GEP instruction that is its address, - // then we can fold the GEP directly into the store instruction. - // emitGEPOperation with a second to last arg of 'true' will place the - // base register for the GEP into baseReg, and the constant offset from that - // into offset. If the offset fits in 16 bits, we can emit a reg+imm store - // otherwise, we copy the offset into another reg, and use reg+reg addressing. - if (GetElementPtrInst *GEPI = canFoldGEPIntoLoadOrStore(SourceAddr)) { - unsigned baseReg = getReg(GEPI); - unsigned pendingAdd; - ConstantSInt *offset; - - emitGEPOperation(BB, BB->end(), GEPI->getOperand(0), GEPI->op_begin()+1, - GEPI->op_end(), baseReg, true, &offset, &pendingAdd); - - if (0 == pendingAdd && Class != cLong && - canUseAsImmediateForOpcode(offset, 0)) { - BuildMI(BB, ImmOpcode, 3).addReg(ValReg).addSImm(offset->getValue()) - .addReg(baseReg); - return; - } - - unsigned indexReg = (pendingAdd != 0) ? pendingAdd : getReg(offset); - BuildMI(BB, IdxOpcode, 3).addReg(ValReg).addReg(indexReg).addReg(baseReg); - return; - } - - // If the store address wasn't the only use of a GEP, we fall back to the - // standard path: store the ValReg at the value in AddressReg. - unsigned AddressReg = getReg(I.getOperand(1)); - BuildMI(BB, ImmOpcode, 3).addReg(ValReg).addSImm(0).addReg(AddressReg); -} - - -/// visitCastInst - Here we have various kinds of copying with or without sign -/// extension going on. -/// -void PPC64ISel::visitCastInst(CastInst &CI) { - Value *Op = CI.getOperand(0); - - unsigned SrcClass = getClassB(Op->getType()); - unsigned DestClass = getClassB(CI.getType()); - - // If this is a cast from a 32-bit integer to a Long type, and the only uses - // of the case are GEP instructions, then the cast does not need to be - // generated explicitly, it will be folded into the GEP. - if (DestClass == cLong && SrcClass == cInt) { - bool AllUsesAreGEPs = true; - for (Value::use_iterator I = CI.use_begin(), E = CI.use_end(); I != E; ++I) - if (!isa(*I)) { - AllUsesAreGEPs = false; - break; - } - - // No need to codegen this cast if all users are getelementptr instrs... - if (AllUsesAreGEPs) return; - } - - unsigned DestReg = getReg(CI); - MachineBasicBlock::iterator MI = BB->end(); - emitCastOperation(BB, MI, Op, CI.getType(), DestReg); -} - -/// emitCastOperation - Common code shared between visitCastInst and constant -/// expression cast support. -/// -void PPC64ISel::emitCastOperation(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Value *Src, const Type *DestTy, - unsigned DestReg) { - const Type *SrcTy = Src->getType(); - unsigned SrcClass = getClassB(SrcTy); - unsigned DestClass = getClassB(DestTy); - unsigned SrcReg = getReg(Src, MBB, IP); - - // Implement casts to bool by using compare on the operand followed by set if - // not zero on the result. - if (DestTy == Type::BoolTy) { - switch (SrcClass) { - case cByte: - case cShort: - case cInt: - case cLong: { - unsigned TmpReg = makeAnotherReg(Type::IntTy); - BuildMI(*MBB, IP, PPC::ADDIC, 2, TmpReg).addReg(SrcReg).addSImm(-1); - BuildMI(*MBB, IP, PPC::SUBFE, 2, DestReg).addReg(TmpReg).addReg(SrcReg); - break; - } - case cFP32: - case cFP64: - // FSEL perhaps? - std::cerr << "ERROR: Cast fp-to-bool not implemented!\n"; - abort(); - } - return; - } - - // Handle cast of Float -> Double - if (SrcClass == cFP32 && DestClass == cFP64) { - BuildMI(*MBB, IP, PPC::FMR, 1, DestReg).addReg(SrcReg); - return; - } - - // Handle cast of Double -> Float - if (SrcClass == cFP64 && DestClass == cFP32) { - BuildMI(*MBB, IP, PPC::FRSP, 1, DestReg).addReg(SrcReg); - return; - } - - // Handle casts from integer to floating point now... - if (DestClass == cFP32 || DestClass == cFP64) { - - // Spill the integer to memory and reload it from there. - unsigned TmpReg = makeAnotherReg(Type::DoubleTy); - int ValueFrameIdx = - F->getFrameInfo()->CreateStackObject(Type::DoubleTy, TM.getTargetData()); - - if (SrcClass == cLong) { - if (SrcTy->isSigned()) { - addFrameReference(BuildMI(*MBB, IP, PPC::STD, 3).addReg(SrcReg), - ValueFrameIdx); - addFrameReference(BuildMI(*MBB, IP, PPC::LFD, 2, TmpReg), - ValueFrameIdx); - BuildMI(*MBB, IP, PPC::FCFID, 1, DestReg).addReg(TmpReg); - } else { - unsigned Scale = getReg(ConstantFP::get(Type::DoubleTy, 0x1p32)); - unsigned TmpHi = makeAnotherReg(Type::IntTy); - unsigned TmpLo = makeAnotherReg(Type::IntTy); - unsigned FPLow = makeAnotherReg(Type::DoubleTy); - unsigned FPTmpHi = makeAnotherReg(Type::DoubleTy); - unsigned FPTmpLo = makeAnotherReg(Type::DoubleTy); - int OtherFrameIdx = F->getFrameInfo()->CreateStackObject(Type::DoubleTy, - TM.getTargetData()); - BuildMI(*MBB, IP, PPC::RLDICL, 3, TmpHi).addReg(SrcReg).addImm(32) - .addImm(32); - BuildMI(*MBB, IP, PPC::RLDICL, 3, TmpLo).addReg(SrcReg).addImm(0) - .addImm(32); - addFrameReference(BuildMI(*MBB, IP, PPC::STD, 3).addReg(TmpHi), - ValueFrameIdx); - addFrameReference(BuildMI(*MBB, IP, PPC::STD, 3).addReg(TmpLo), - OtherFrameIdx); - addFrameReference(BuildMI(*MBB, IP, PPC::LFD, 2, TmpReg), - ValueFrameIdx); - addFrameReference(BuildMI(*MBB, IP, PPC::LFD, 2, FPLow), - OtherFrameIdx); - BuildMI(*MBB, IP, PPC::FCFID, 1, FPTmpHi).addReg(TmpReg); - BuildMI(*MBB, IP, PPC::FCFID, 1, FPTmpLo).addReg(FPLow); - BuildMI(*MBB, IP, PPC::FMADD, 3, DestReg).addReg(Scale).addReg(FPTmpHi) - .addReg(FPTmpLo); - } - return; - } - - // FIXME: really want a promote64 - unsigned IntTmp = makeAnotherReg(Type::IntTy); - - if (SrcTy->isSigned()) - BuildMI(*MBB, IP, PPC::EXTSW, 1, IntTmp).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::RLDICL, 3, IntTmp).addReg(SrcReg).addImm(0) - .addImm(32); - addFrameReference(BuildMI(*MBB, IP, PPC::STD, 3).addReg(IntTmp), - ValueFrameIdx); - addFrameReference(BuildMI(*MBB, IP, PPC::LFD, 2, TmpReg), - ValueFrameIdx); - BuildMI(*MBB, IP, PPC::FCFID, 1, DestReg).addReg(TmpReg); - return; - } - - // Handle casts from floating point to integer now... - if (SrcClass == cFP32 || SrcClass == cFP64) { - static Function* const Funcs[] = - { __fixsfdiFn, __fixdfdiFn, __fixunssfdiFn, __fixunsdfdiFn }; - // emit library call - if (DestClass == cLong) { - bool isDouble = SrcClass == cFP64; - unsigned nameIndex = 2 * DestTy->isSigned() + isDouble; - std::vector Args; - Args.push_back(ValueRecord(SrcReg, SrcTy)); - Function *floatFn = Funcs[nameIndex]; - MachineInstr *TheCall = - BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(floatFn, true); - doCall(ValueRecord(DestReg, DestTy), TheCall, Args, false); - return; - } - - int ValueFrameIdx = - F->getFrameInfo()->CreateStackObject(SrcTy, TM.getTargetData()); - - if (DestTy->isSigned()) { - unsigned TempReg = makeAnotherReg(Type::DoubleTy); - - // Convert to integer in the FP reg and store it to a stack slot - BuildMI(*BB, IP, PPC::FCTIWZ, 1, TempReg).addReg(SrcReg); - addFrameReference(BuildMI(*BB, IP, PPC::STFD, 3) - .addReg(TempReg), ValueFrameIdx); - - // There is no load signed byte opcode, so we must emit a sign extend for - // that particular size. Make sure to source the new integer from the - // correct offset. - if (DestClass == cByte) { - unsigned TempReg2 = makeAnotherReg(DestTy); - addFrameReference(BuildMI(*BB, IP, PPC::LBZ, 2, TempReg2), - ValueFrameIdx, 7); - BuildMI(*MBB, IP, PPC::EXTSB, DestReg).addReg(TempReg2); - } else { - int offset = (DestClass == cShort) ? 6 : 4; - unsigned LoadOp = (DestClass == cShort) ? PPC::LHA : PPC::LWZ; - addFrameReference(BuildMI(*BB, IP, LoadOp, 2, DestReg), - ValueFrameIdx, offset); - } - } else { - unsigned Zero = getReg(ConstantFP::get(Type::DoubleTy, 0.0f)); - double maxInt = (1LL << 32) - 1; - unsigned MaxInt = getReg(ConstantFP::get(Type::DoubleTy, maxInt)); - double border = 1LL << 31; - unsigned Border = getReg(ConstantFP::get(Type::DoubleTy, border)); - unsigned UseZero = makeAnotherReg(Type::DoubleTy); - unsigned UseMaxInt = makeAnotherReg(Type::DoubleTy); - unsigned UseChoice = makeAnotherReg(Type::DoubleTy); - unsigned TmpReg = makeAnotherReg(Type::DoubleTy); - unsigned TmpReg2 = makeAnotherReg(Type::DoubleTy); - unsigned ConvReg = makeAnotherReg(Type::DoubleTy); - unsigned IntTmp = makeAnotherReg(Type::IntTy); - unsigned XorReg = makeAnotherReg(Type::IntTy); - int FrameIdx = - F->getFrameInfo()->CreateStackObject(SrcTy, TM.getTargetData()); - // Update machine-CFG edges - MachineBasicBlock *XorMBB = new MachineBasicBlock(BB->getBasicBlock()); - MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock()); - MachineBasicBlock *OldMBB = BB; - ilist::iterator It = BB; ++It; - F->getBasicBlockList().insert(It, XorMBB); - F->getBasicBlockList().insert(It, PhiMBB); - BB->addSuccessor(XorMBB); - BB->addSuccessor(PhiMBB); - - // Convert from floating point to unsigned 32-bit value - // Use 0 if incoming value is < 0.0 - BuildMI(*BB, IP, PPC::FSEL, 3, UseZero).addReg(SrcReg).addReg(SrcReg) - .addReg(Zero); - // Use 2**32 - 1 if incoming value is >= 2**32 - BuildMI(*BB, IP, PPC::FSUB, 2, UseMaxInt).addReg(MaxInt).addReg(SrcReg); - BuildMI(*BB, IP, PPC::FSEL, 3, UseChoice).addReg(UseMaxInt) - .addReg(UseZero).addReg(MaxInt); - // Subtract 2**31 - BuildMI(*BB, IP, PPC::FSUB, 2, TmpReg).addReg(UseChoice).addReg(Border); - // Use difference if >= 2**31 - BuildMI(*BB, IP, PPC::FCMPU, 2, PPC::CR0).addReg(UseChoice) - .addReg(Border); - BuildMI(*BB, IP, PPC::FSEL, 3, TmpReg2).addReg(TmpReg).addReg(TmpReg) - .addReg(UseChoice); - // Convert to integer - BuildMI(*BB, IP, PPC::FCTIWZ, 1, ConvReg).addReg(TmpReg2); - addFrameReference(BuildMI(*BB, IP, PPC::STFD, 3).addReg(ConvReg), - FrameIdx); - if (DestClass == cByte) { - addFrameReference(BuildMI(*BB, IP, PPC::LBZ, 2, DestReg), - FrameIdx, 7); - } else if (DestClass == cShort) { - addFrameReference(BuildMI(*BB, IP, PPC::LHZ, 2, DestReg), - FrameIdx, 6); - } if (DestClass == cInt) { - addFrameReference(BuildMI(*BB, IP, PPC::LWZ, 2, IntTmp), - FrameIdx, 4); - BuildMI(*BB, IP, PPC::BLT, 2).addReg(PPC::CR0).addMBB(PhiMBB); - BuildMI(*BB, IP, PPC::B, 1).addMBB(XorMBB); - - // XorMBB: - // add 2**31 if input was >= 2**31 - BB = XorMBB; - BuildMI(BB, PPC::XORIS, 2, XorReg).addReg(IntTmp).addImm(0x8000); - XorMBB->addSuccessor(PhiMBB); - - // PhiMBB: - // DestReg = phi [ IntTmp, OldMBB ], [ XorReg, XorMBB ] - BB = PhiMBB; - BuildMI(BB, PPC::PHI, 4, DestReg).addReg(IntTmp).addMBB(OldMBB) - .addReg(XorReg).addMBB(XorMBB); - } - } - return; - } - - // Check our invariants - assert((SrcClass <= cInt || SrcClass == cLong) && - "Unhandled source class for cast operation!"); - assert((DestClass <= cInt || DestClass == cLong) && - "Unhandled destination class for cast operation!"); - - bool sourceUnsigned = SrcTy->isUnsigned() || SrcTy == Type::BoolTy; - bool destUnsigned = DestTy->isUnsigned(); - - // Unsigned -> Unsigned, clear if larger - if (sourceUnsigned && destUnsigned) { - // handle long dest class now to keep switch clean - if (DestClass == cLong) { - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - return; - } - - // handle u{ byte, short, int } x u{ byte, short, int } - unsigned clearBits = (SrcClass == cByte || DestClass == cByte) ? 24 : 16; - switch (SrcClass) { - case cByte: - case cShort: - if (SrcClass == DestClass) - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg) - .addImm(0).addImm(clearBits).addImm(31); - break; - case cInt: - case cLong: - if (DestClass == cInt) - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg) - .addImm(0).addImm(clearBits).addImm(31); - break; - } - return; - } - - // Signed -> Signed - if (!sourceUnsigned && !destUnsigned) { - // handle long dest class now to keep switch clean - if (DestClass == cLong) { - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - return; - } - - // handle { byte, short, int } x { byte, short, int } - switch (SrcClass) { - case cByte: - if (DestClass == cByte) - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::EXTSB, 1, DestReg).addReg(SrcReg); - break; - case cShort: - if (DestClass == cByte) - BuildMI(*MBB, IP, PPC::EXTSB, 1, DestReg).addReg(SrcReg); - else if (DestClass == cShort) - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::EXTSH, 1, DestReg).addReg(SrcReg); - break; - case cInt: - case cLong: - if (DestClass == cByte) - BuildMI(*MBB, IP, PPC::EXTSB, 1, DestReg).addReg(SrcReg); - else if (DestClass == cShort) - BuildMI(*MBB, IP, PPC::EXTSH, 1, DestReg).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - break; - } - return; - } - - // Unsigned -> Signed - if (sourceUnsigned && !destUnsigned) { - // handle long dest class now to keep switch clean - if (DestClass == cLong) { - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - return; - } - - // handle u{ byte, short, int } -> { byte, short, int } - switch (SrcClass) { - case cByte: - if (DestClass == cByte) - // uByte 255 -> signed byte == -1 - BuildMI(*MBB, IP, PPC::EXTSB, 1, DestReg).addReg(SrcReg); - else - // uByte 255 -> signed short/int == 255 - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg).addImm(0) - .addImm(24).addImm(31); - break; - case cShort: - if (DestClass == cByte) - BuildMI(*MBB, IP, PPC::EXTSB, 1, DestReg).addReg(SrcReg); - else if (DestClass == cShort) - BuildMI(*MBB, IP, PPC::EXTSH, 1, DestReg).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg).addImm(0) - .addImm(16).addImm(31); - break; - case cInt: - case cLong: - if (DestClass == cByte) - BuildMI(*MBB, IP, PPC::EXTSB, 1, DestReg).addReg(SrcReg); - else if (DestClass == cShort) - BuildMI(*MBB, IP, PPC::EXTSH, 1, DestReg).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - break; - } - return; - } - - // Signed -> Unsigned - if (!sourceUnsigned && destUnsigned) { - // handle long dest class now to keep switch clean - if (DestClass == cLong) { - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - return; - } - - // handle { byte, short, int } -> u{ byte, short, int } - unsigned clearBits = (DestClass == cByte) ? 24 : 16; - switch (SrcClass) { - case cByte: - case cShort: - if (DestClass == cByte || DestClass == cShort) - // sbyte -1 -> ubyte 0x000000FF - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg) - .addImm(0).addImm(clearBits).addImm(31); - else - // sbyte -1 -> ubyte 0xFFFFFFFF - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - break; - case cInt: - case cLong: - if (DestClass == cInt) - BuildMI(*MBB, IP, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg); - else - BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg) - .addImm(0).addImm(clearBits).addImm(31); - break; - } - return; - } - - // Anything we haven't handled already, we can't (yet) handle at all. - std::cerr << "Unhandled cast from " << SrcTy->getDescription() - << "to " << DestTy->getDescription() << '\n'; - abort(); -} - -/// visitVANextInst - Implement the va_next instruction... -/// -void PPC64ISel::visitVANextInst(VANextInst &I) { - unsigned VAList = getReg(I.getOperand(0)); - unsigned DestReg = getReg(I); - - unsigned Size; - switch (I.getArgType()->getTypeID()) { - default: - std::cerr << I; - assert(0 && "Error: bad type for va_next instruction!"); - return; - case Type::PointerTyID: - case Type::UIntTyID: - case Type::IntTyID: - Size = 4; - break; - case Type::ULongTyID: - case Type::LongTyID: - case Type::DoubleTyID: - Size = 8; - break; - } - - // Increment the VAList pointer... - BuildMI(BB, PPC::ADDI, 2, DestReg).addReg(VAList).addSImm(Size); -} - -void PPC64ISel::visitVAArgInst(VAArgInst &I) { - unsigned VAList = getReg(I.getOperand(0)); - unsigned DestReg = getReg(I); - - switch (I.getType()->getTypeID()) { - default: - std::cerr << I; - assert(0 && "Error: bad type for va_next instruction!"); - return; - case Type::PointerTyID: - case Type::UIntTyID: - case Type::IntTyID: - BuildMI(BB, PPC::LWZ, 2, DestReg).addSImm(0).addReg(VAList); - break; - case Type::ULongTyID: - case Type::LongTyID: - BuildMI(BB, PPC::LD, 2, DestReg).addSImm(0).addReg(VAList); - break; - case Type::FloatTyID: - BuildMI(BB, PPC::LFS, 2, DestReg).addSImm(0).addReg(VAList); - break; - case Type::DoubleTyID: - BuildMI(BB, PPC::LFD, 2, DestReg).addSImm(0).addReg(VAList); - break; - } -} - -/// visitGetElementPtrInst - instruction-select GEP instructions -/// -void PPC64ISel::visitGetElementPtrInst(GetElementPtrInst &I) { - if (canFoldGEPIntoLoadOrStore(&I)) - return; - - unsigned outputReg = getReg(I); - emitGEPOperation(BB, BB->end(), I.getOperand(0), I.op_begin()+1, I.op_end(), - outputReg, false, 0, 0); -} - -/// emitGEPOperation - Common code shared between visitGetElementPtrInst and -/// constant expression GEP support. -/// -void PPC64ISel::emitGEPOperation(MachineBasicBlock *MBB, - MachineBasicBlock::iterator IP, - Value *Src, User::op_iterator IdxBegin, - User::op_iterator IdxEnd, unsigned TargetReg, - bool GEPIsFolded, ConstantSInt **RemainderPtr, - unsigned *PendingAddReg) { - const TargetData &TD = TM.getTargetData(); - const Type *Ty = Src->getType(); - unsigned basePtrReg = getReg(Src, MBB, IP); - int64_t constValue = 0; - - // Record the operations to emit the GEP in a vector so that we can emit them - // after having analyzed the entire instruction. - std::vector ops; - - // GEPs have zero or more indices; we must perform a struct access - // or array access for each one. - for (GetElementPtrInst::op_iterator oi = IdxBegin, oe = IdxEnd; oi != oe; - ++oi) { - Value *idx = *oi; - if (const StructType *StTy = dyn_cast(Ty)) { - // It's a struct access. idx is the index into the structure, - // which names the field. Use the TargetData structure to - // pick out what the layout of the structure is in memory. - // Use the (constant) structure index's value to find the - // right byte offset from the StructLayout class's list of - // structure member offsets. - unsigned fieldIndex = cast(idx)->getValue(); - unsigned memberOffset = - TD.getStructLayout(StTy)->MemberOffsets[fieldIndex]; - - // StructType member offsets are always constant values. Add it to the - // running total. - constValue += memberOffset; - - // The next type is the member of the structure selected by the - // index. - Ty = StTy->getElementType (fieldIndex); - } else if (const SequentialType *SqTy = dyn_cast (Ty)) { - // Many GEP instructions use a [cast (int/uint) to LongTy] as their - // operand. Handle this case directly now... - if (CastInst *CI = dyn_cast(idx)) - if (CI->getOperand(0)->getType() == Type::IntTy || - CI->getOperand(0)->getType() == Type::UIntTy) - idx = CI->getOperand(0); - - // It's an array or pointer access: [ArraySize x ElementType]. - // We want to add basePtrReg to (idxReg * sizeof ElementType). First, we - // must find the size of the pointed-to type (Not coincidentally, the next - // type is the type of the elements in the array). - Ty = SqTy->getElementType(); - unsigned elementSize = TD.getTypeSize(Ty); - - if (ConstantInt *C = dyn_cast(idx)) { - if (ConstantSInt *CS = dyn_cast(C)) - constValue += CS->getValue() * elementSize; - else if (ConstantUInt *CU = dyn_cast(C)) - constValue += CU->getValue() * elementSize; - else - assert(0 && "Invalid ConstantInt GEP index type!"); - } else { - // Push current gep state to this point as an add - ops.push_back(CollapsedGepOp(false, 0, - ConstantSInt::get(Type::IntTy,constValue))); - - // Push multiply gep op and reset constant value - ops.push_back(CollapsedGepOp(true, idx, - ConstantSInt::get(Type::IntTy, elementSize))); - - constValue = 0; - } - } - } - // Emit instructions for all the collapsed ops - bool pendingAdd = false; - unsigned pendingAddReg = 0; - - for(std::vector::iterator cgo_i = ops.begin(), - cgo_e = ops.end(); cgo_i != cgo_e; ++cgo_i) { - CollapsedGepOp& cgo = *cgo_i; - unsigned nextBasePtrReg = makeAnotherReg(Type::IntTy); - - // If we didn't emit an add last time through the loop, we need to now so - // that the base reg is updated appropriately. - if (pendingAdd) { - assert(pendingAddReg != 0 && "Uninitialized register in pending add!"); - BuildMI(*MBB, IP, PPC::ADD, 2, nextBasePtrReg).addReg(basePtrReg) - .addReg(pendingAddReg); - basePtrReg = nextBasePtrReg; - nextBasePtrReg = makeAnotherReg(Type::IntTy); - pendingAddReg = 0; - pendingAdd = false; - } - - if (cgo.isMul) { - // We know the elementSize is a constant, so we can emit a constant mul - unsigned TmpReg = makeAnotherReg(Type::IntTy); - doMultiplyConst(MBB, IP, nextBasePtrReg, cgo.index, cgo.size); - pendingAddReg = basePtrReg; - pendingAdd = true; - } else { - // Try and generate an immediate addition if possible - if (cgo.size->isNullValue()) { - BuildMI(*MBB, IP, PPC::OR, 2, nextBasePtrReg).addReg(basePtrReg) - .addReg(basePtrReg); - } else if (canUseAsImmediateForOpcode(cgo.size, 0)) { - BuildMI(*MBB, IP, PPC::ADDI, 2, nextBasePtrReg).addReg(basePtrReg) - .addSImm(cgo.size->getValue()); - } else { - unsigned Op1r = getReg(cgo.size, MBB, IP); - BuildMI(*MBB, IP, PPC::ADD, 2, nextBasePtrReg).addReg(basePtrReg) - .addReg(Op1r); - } - } - - basePtrReg = nextBasePtrReg; - } - // Add the current base register plus any accumulated constant value - ConstantSInt *remainder = ConstantSInt::get(Type::IntTy, constValue); - - // If we are emitting this during a fold, copy the current base register to - // the target, and save the current constant offset so the folding load or - // store can try and use it as an immediate. - if (GEPIsFolded) { - // If this is a folded GEP and the last element was an index, then we need - // to do some extra work to turn a shift/add/stw into a shift/stwx - if (pendingAdd && 0 == remainder->getValue()) { - assert(pendingAddReg != 0 && "Uninitialized register in pending add!"); - *PendingAddReg = pendingAddReg; - } else { - *PendingAddReg = 0; - if (pendingAdd) { - unsigned nextBasePtrReg = makeAnotherReg(Type::IntTy); - assert(pendingAddReg != 0 && "Uninitialized register in pending add!"); - BuildMI(*MBB, IP, PPC::ADD, 2, nextBasePtrReg).addReg(basePtrReg) - .addReg(pendingAddReg); - basePtrReg = nextBasePtrReg; - } - } - BuildMI (*MBB, IP, PPC::OR, 2, TargetReg).addReg(basePtrReg) - .addReg(basePtrReg); - *RemainderPtr = remainder; - return; - } - - // If we still have a pending add at this point, emit it now - if (pendingAdd) { - unsigned TmpReg = makeAnotherReg(Type::IntTy); - BuildMI(*MBB, IP, PPC::ADD, 2, TmpReg).addReg(pendingAddReg) - .addReg(basePtrReg); - basePtrReg = TmpReg; - } - - // After we have processed all the indices, the result is left in - // basePtrReg. Move it to the register where we were expected to - // put the answer. - if (remainder->isNullValue()) { - BuildMI (*MBB, IP, PPC::OR, 2, TargetReg).addReg(basePtrReg) - .addReg(basePtrReg); - } else if (canUseAsImmediateForOpcode(remainder, 0)) { - BuildMI(*MBB, IP, PPC::ADDI, 2, TargetReg).addReg(basePtrReg) - .addSImm(remainder->getValue()); - } else { - unsigned Op1r = getReg(remainder, MBB, IP); - BuildMI(*MBB, IP, PPC::ADD, 2, TargetReg).addReg(basePtrReg).addReg(Op1r); - } -} - -/// visitAllocaInst - If this is a fixed size alloca, allocate space from the -/// frame manager, otherwise do it the hard way. -/// -void PPC64ISel::visitAllocaInst(AllocaInst &I) { - // If this is a fixed size alloca in the entry block for the function, we - // statically stack allocate the space, so we don't need to do anything here. - // - if (dyn_castFixedAlloca(&I)) return; - - // Find the data size of the alloca inst's getAllocatedType. - const Type *Ty = I.getAllocatedType(); - unsigned TySize = TM.getTargetData().getTypeSize(Ty); - - // Create a register to hold the temporary result of multiplying the type size - // constant by the variable amount. - unsigned TotalSizeReg = makeAnotherReg(Type::UIntTy); - - // TotalSizeReg = mul , - MachineBasicBlock::iterator MBBI = BB->end(); - ConstantUInt *CUI = ConstantUInt::get(Type::UIntTy, TySize); - doMultiplyConst(BB, MBBI, TotalSizeReg, I.getArraySize(), CUI); - - // AddedSize = add , 15 - unsigned AddedSizeReg = makeAnotherReg(Type::UIntTy); - BuildMI(BB, PPC::ADDI, 2, AddedSizeReg).addReg(TotalSizeReg).addSImm(15); - - // AlignedSize = and , ~15 - unsigned AlignedSize = makeAnotherReg(Type::UIntTy); - BuildMI(BB, PPC::RLWINM, 4, AlignedSize).addReg(AddedSizeReg).addImm(0) - .addImm(0).addImm(27); - - // Subtract size from stack pointer, thereby allocating some space. - BuildMI(BB, PPC::SUB, 2, PPC::R1).addReg(PPC::R1).addReg(AlignedSize); - - // Put a pointer to the space into the result register, by copying - // the stack pointer. - BuildMI(BB, PPC::OR, 2, getReg(I)).addReg(PPC::R1).addReg(PPC::R1); - - // Inform the Frame Information that we have just allocated a variable-sized - // object. - F->getFrameInfo()->CreateVariableSizedObject(); -} - -/// visitMallocInst - Malloc instructions are code generated into direct calls -/// to the library malloc. -/// -void PPC64ISel::visitMallocInst(MallocInst &I) { - unsigned AllocSize = TM.getTargetData().getTypeSize(I.getAllocatedType()); - unsigned Arg; - - if (ConstantUInt *C = dyn_cast(I.getOperand(0))) { - Arg = getReg(ConstantUInt::get(Type::UIntTy, C->getValue() * AllocSize)); - } else { - Arg = makeAnotherReg(Type::UIntTy); - MachineBasicBlock::iterator MBBI = BB->end(); - ConstantUInt *CUI = ConstantUInt::get(Type::UIntTy, AllocSize); - doMultiplyConst(BB, MBBI, Arg, I.getOperand(0), CUI); - } - - std::vector Args; - Args.push_back(ValueRecord(Arg, Type::UIntTy)); - MachineInstr *TheCall = - BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(mallocFn, true); - doCall(ValueRecord(getReg(I), I.getType()), TheCall, Args, false); -} - - -/// visitFreeInst - Free instructions are code gen'd to call the free libc -/// function. -/// -void PPC64ISel::visitFreeInst(FreeInst &I) { - std::vector Args; - Args.push_back(ValueRecord(I.getOperand(0))); - MachineInstr *TheCall = - BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(freeFn, true); - doCall(ValueRecord(0, Type::VoidTy), TheCall, Args, false); -} - -/// createPPC64ISelSimple - This pass converts an LLVM function into a machine -/// code representation is a very simple peep-hole fashion. -/// -FunctionPass *llvm::createPPC64ISelSimple(TargetMachine &TM) { - return new PPC64ISel(TM); -} diff --git a/lib/Target/PowerPC/PPCTargetMachine.cpp b/lib/Target/PowerPC/PPCTargetMachine.cpp index e7ab04ff0cc..3d27c98e6af 100644 --- a/lib/Target/PowerPC/PPCTargetMachine.cpp +++ b/lib/Target/PowerPC/PPCTargetMachine.cpp @@ -97,7 +97,7 @@ bool PowerPCTargetMachine::addPassesToEmitAssembly(PassManager &PM, PM.add(createUnreachableBlockEliminationPass()); if (LP64) - PM.add(createPPC64ISelSimple(*this)); + PM.add(createPPC64ISelPattern(*this)); else if (EnablePatternISel) PM.add(createPPC32ISelPattern(*this)); else