diff options
Diffstat (limited to 'contrib/llvm/lib/Target/X86/X86FrameLowering.cpp')
| -rw-r--r-- | contrib/llvm/lib/Target/X86/X86FrameLowering.cpp | 529 |
1 files changed, 401 insertions, 128 deletions
diff --git a/contrib/llvm/lib/Target/X86/X86FrameLowering.cpp b/contrib/llvm/lib/Target/X86/X86FrameLowering.cpp index f5ffe0cf7e88..03d925692adf 100644 --- a/contrib/llvm/lib/Target/X86/X86FrameLowering.cpp +++ b/contrib/llvm/lib/Target/X86/X86FrameLowering.cpp @@ -159,6 +159,8 @@ static unsigned findDeadCallerSavedReg(MachineBasicBlock &MBB, unsigned Opc = MBBI->getOpcode(); switch (Opc) { default: return 0; + case TargetOpcode::PATCHABLE_RET: + case X86::RET: case X86::RETL: case X86::RETQ: case X86::RETIL: @@ -314,8 +316,8 @@ void X86FrameLowering::emitSPUpdate(MachineBasicBlock &MBB, } MachineInstrBuilder X86FrameLowering::BuildStackAdjustment( - MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, DebugLoc DL, - int64_t Offset, bool InEpilogue) const { + MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, + const DebugLoc &DL, int64_t Offset, bool InEpilogue) const { assert(Offset != 0 && "zero offset stack adjustment requested"); // On Atom, using LEA to adjust SP is preferred, but using it in the epilogue @@ -374,16 +376,33 @@ int X86FrameLowering::mergeSPUpdates(MachineBasicBlock &MBB, unsigned Opc = PI->getOpcode(); int Offset = 0; + if (!doMergeWithPrevious && NI != MBB.end() && + NI->getOpcode() == TargetOpcode::CFI_INSTRUCTION) { + // Don't merge with the next instruction if it has CFI. + return Offset; + } + if ((Opc == X86::ADD64ri32 || Opc == X86::ADD64ri8 || - Opc == X86::ADD32ri || Opc == X86::ADD32ri8 || - Opc == X86::LEA32r || Opc == X86::LEA64_32r) && + Opc == X86::ADD32ri || Opc == X86::ADD32ri8) && PI->getOperand(0).getReg() == StackPtr){ + assert(PI->getOperand(1).getReg() == StackPtr); Offset += PI->getOperand(2).getImm(); MBB.erase(PI); if (!doMergeWithPrevious) MBBI = NI; + } else if ((Opc == X86::LEA32r || Opc == X86::LEA64_32r) && + PI->getOperand(0).getReg() == StackPtr && + PI->getOperand(1).getReg() == StackPtr && + PI->getOperand(2).getImm() == 1 && + PI->getOperand(3).getReg() == X86::NoRegister && + PI->getOperand(5).getReg() == X86::NoRegister) { + // For LEAs we have: def = lea SP, FI, noreg, Offset, noreg. + Offset += PI->getOperand(4).getImm(); + MBB.erase(PI); + if (!doMergeWithPrevious) MBBI = NI; } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB64ri8 || Opc == X86::SUB32ri || Opc == X86::SUB32ri8) && PI->getOperand(0).getReg() == StackPtr) { + assert(PI->getOperand(1).getReg() == StackPtr); Offset -= PI->getOperand(2).getImm(); MBB.erase(PI); if (!doMergeWithPrevious) MBBI = NI; @@ -393,18 +412,18 @@ int X86FrameLowering::mergeSPUpdates(MachineBasicBlock &MBB, } void X86FrameLowering::BuildCFI(MachineBasicBlock &MBB, - MachineBasicBlock::iterator MBBI, DebugLoc DL, - MCCFIInstruction CFIInst) const { + MachineBasicBlock::iterator MBBI, + const DebugLoc &DL, + const MCCFIInstruction &CFIInst) const { MachineFunction &MF = *MBB.getParent(); unsigned CFIIndex = MF.getMMI().addFrameInst(CFIInst); BuildMI(MBB, MBBI, DL, TII.get(TargetOpcode::CFI_INSTRUCTION)) .addCFIIndex(CFIIndex); } -void -X86FrameLowering::emitCalleeSavedFrameMoves(MachineBasicBlock &MBB, - MachineBasicBlock::iterator MBBI, - DebugLoc DL) const { +void X86FrameLowering::emitCalleeSavedFrameMoves( + MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, + const DebugLoc &DL) const { MachineFunction &MF = *MBB.getParent(); MachineFrameInfo *MFI = MF.getFrameInfo(); MachineModuleInfo &MMI = MF.getMMI(); @@ -429,7 +448,7 @@ X86FrameLowering::emitCalleeSavedFrameMoves(MachineBasicBlock &MBB, MachineInstr *X86FrameLowering::emitStackProbe(MachineFunction &MF, MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, - DebugLoc DL, + const DebugLoc &DL, bool InProlog) const { const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>(); if (STI.isTargetWindowsCoreCLR()) { @@ -457,6 +476,8 @@ void X86FrameLowering::inlineStackProbe(MachineFunction &MF, } if (ChkStkStub != nullptr) { + assert(!ChkStkStub->isBundled() && + "Not expecting bundled instructions here"); MachineBasicBlock::iterator MBBI = std::next(ChkStkStub->getIterator()); assert(std::prev(MBBI).operator==(ChkStkStub) && "MBBI expected after __chkstk_stub."); @@ -467,8 +488,8 @@ void X86FrameLowering::inlineStackProbe(MachineFunction &MF, } MachineInstr *X86FrameLowering::emitStackProbeInline( - MachineFunction &MF, MachineBasicBlock &MBB, - MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const { + MachineFunction &MF, MachineBasicBlock &MBB, + MachineBasicBlock::iterator MBBI, const DebugLoc &DL, bool InProlog) const { const X86Subtarget &STI = MF.getSubtarget<X86Subtarget>(); assert(STI.is64Bit() && "different expansion needed for 32 bit"); assert(STI.isTargetWindowsCoreCLR() && "custom expansion expects CoreCLR"); @@ -679,12 +700,12 @@ MachineInstr *X86FrameLowering::emitStackProbeInline( // Possible TODO: physreg liveness for InProlog case. - return ContinueMBBI; + return &*ContinueMBBI; } MachineInstr *X86FrameLowering::emitStackProbeCall( MachineFunction &MF, MachineBasicBlock &MBB, - MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const { + MachineBasicBlock::iterator MBBI, const DebugLoc &DL, bool InProlog) const { bool IsLargeCodeModel = MF.getTarget().getCodeModel() == CodeModel::Large; unsigned CallOp; @@ -743,19 +764,19 @@ MachineInstr *X86FrameLowering::emitStackProbeCall( ExpansionMBBI->setFlag(MachineInstr::FrameSetup); } - return MBBI; + return &*MBBI; } MachineInstr *X86FrameLowering::emitStackProbeInlineStub( MachineFunction &MF, MachineBasicBlock &MBB, - MachineBasicBlock::iterator MBBI, DebugLoc DL, bool InProlog) const { + MachineBasicBlock::iterator MBBI, const DebugLoc &DL, bool InProlog) const { assert(InProlog && "ChkStkStub called outside prolog!"); BuildMI(MBB, MBBI, DL, TII.get(X86::CALLpcrel32)) .addExternalSymbol("__chkstk_stub"); - return MBBI; + return &*MBBI; } static unsigned calculateSetFPREG(uint64_t SPAdjust) { @@ -786,7 +807,7 @@ uint64_t X86FrameLowering::calculateMaxStackAlign(const MachineFunction &MF) con void X86FrameLowering::BuildStackAlignAND(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, - DebugLoc DL, unsigned Reg, + const DebugLoc &DL, unsigned Reg, uint64_t MaxAlign) const { uint64_t Val = -MaxAlign; unsigned AndOp = getANDriOpcode(Uses64BitFramePtr, Val); @@ -950,6 +971,7 @@ void X86FrameLowering::emitPrologue(MachineFunction &MF, !MF.shouldSplitStack()) { // Regular stack uint64_t MinSize = X86FI->getCalleeSavedFrameSize(); if (HasFP) MinSize += SlotSize; + X86FI->setUsesRedZone(MinSize > 0 || StackSize > 0); StackSize = std::max(MinSize, StackSize > 128 ? StackSize - 128 : 0); MFI->setStackSize(StackSize); } @@ -1009,7 +1031,7 @@ void X86FrameLowering::emitPrologue(MachineFunction &MF, // Callee-saved registers are pushed on stack before the stack is realigned. if (TRI->needsStackRealignment(MF) && !IsWin64Prologue) - NumBytes = RoundUpToAlignment(NumBytes, MaxAlign); + NumBytes = alignTo(NumBytes, MaxAlign); // Get the offset of the stack slot for the EBP register, which is // guaranteed to be the last slot by processFunctionBeforeFrameFinalized. @@ -1130,7 +1152,7 @@ void X86FrameLowering::emitPrologue(MachineFunction &MF, // virtual memory manager are allocated in correct sequence. uint64_t AlignedNumBytes = NumBytes; if (IsWin64Prologue && !IsFunclet && TRI->needsStackRealignment(MF)) - AlignedNumBytes = RoundUpToAlignment(AlignedNumBytes, MaxAlign); + AlignedNumBytes = alignTo(AlignedNumBytes, MaxAlign); if (AlignedNumBytes >= StackProbeSize && UseStackProbe) { // Check whether EAX is livein for this block. bool isEAXAlive = isEAXLiveIn(MBB); @@ -1260,7 +1282,7 @@ void X86FrameLowering::emitPrologue(MachineFunction &MF, } while (MBBI != MBB.end() && MBBI->getFlag(MachineInstr::FrameSetup)) { - const MachineInstr *FrameInstr = &*MBBI; + const MachineInstr &FrameInstr = *MBBI; ++MBBI; if (NeedsWinCFI) { @@ -1360,6 +1382,18 @@ void X86FrameLowering::emitPrologue(MachineFunction &MF, if (PushedRegs) emitCalleeSavedFrameMoves(MBB, MBBI, DL); } + + // X86 Interrupt handling function cannot assume anything about the direction + // flag (DF in EFLAGS register). Clear this flag by creating "cld" instruction + // in each prologue of interrupt handler function. + // + // FIXME: Create "cld" instruction only in these cases: + // 1. The interrupt handling function uses any of the "rep" instructions. + // 2. Interrupt handling function calls another function. + // + if (Fn->getCallingConv() == CallingConv::X86_INTR) + BuildMI(MBB, MBBI, DL, TII.get(X86::CLD)) + .setMIFlag(MachineInstr::FrameSetup); } bool X86FrameLowering::canUseLEAForSPInEpilogue( @@ -1373,8 +1407,8 @@ bool X86FrameLowering::canUseLEAForSPInEpilogue( return !MF.getTarget().getMCAsmInfo()->usesWindowsCFI() || hasFP(MF); } -static bool isFuncletReturnInstr(MachineInstr *MI) { - switch (MI->getOpcode()) { +static bool isFuncletReturnInstr(MachineInstr &MI) { + switch (MI.getOpcode()) { case X86::CATCHRET: case X86::CLEANUPRET: return true; @@ -1400,11 +1434,10 @@ static bool isFuncletReturnInstr(MachineInstr *MI) { unsigned X86FrameLowering::getPSPSlotOffsetFromSP(const MachineFunction &MF) const { const WinEHFuncInfo &Info = *MF.getWinEHFuncInfo(); - // getFrameIndexReferenceFromSP has an out ref parameter for the stack - // pointer register; pass a dummy that we ignore unsigned SPReg; - int Offset = getFrameIndexReferenceFromSP(MF, Info.PSPSymFrameIdx, SPReg); - assert(Offset >= 0); + int Offset = getFrameIndexReferencePreferSP(MF, Info.PSPSymFrameIdx, SPReg, + /*IgnoreSPUpdates*/ true); + assert(Offset >= 0 && SPReg == TRI->getStackRegister()); return static_cast<unsigned>(Offset); } @@ -1429,18 +1462,25 @@ X86FrameLowering::getWinEHFuncletFrameSize(const MachineFunction &MF) const { // RBP is not included in the callee saved register block. After pushing RBP, // everything is 16 byte aligned. Everything we allocate before an outgoing // call must also be 16 byte aligned. - unsigned FrameSizeMinusRBP = - RoundUpToAlignment(CSSize + UsedSize, getStackAlignment()); + unsigned FrameSizeMinusRBP = alignTo(CSSize + UsedSize, getStackAlignment()); // Subtract out the size of the callee saved registers. This is how much stack // each funclet will allocate. return FrameSizeMinusRBP - CSSize; } +static bool isTailCallOpcode(unsigned Opc) { + return Opc == X86::TCRETURNri || Opc == X86::TCRETURNdi || + Opc == X86::TCRETURNmi || + Opc == X86::TCRETURNri64 || Opc == X86::TCRETURNdi64 || + Opc == X86::TCRETURNmi64; +} + void X86FrameLowering::emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const { const MachineFrameInfo *MFI = MF.getFrameInfo(); X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>(); MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator(); + unsigned RetOpcode = MBBI->getOpcode(); DebugLoc DL; if (MBBI != MBB.end()) DL = MBBI->getDebugLoc(); @@ -1453,7 +1493,7 @@ void X86FrameLowering::emitEpilogue(MachineFunction &MF, bool IsWin64Prologue = MF.getTarget().getMCAsmInfo()->usesWindowsCFI(); bool NeedsWinCFI = IsWin64Prologue && MF.getFunction()->needsUnwindTableEntry(); - bool IsFunclet = isFuncletReturnInstr(MBBI); + bool IsFunclet = isFuncletReturnInstr(*MBBI); MachineBasicBlock *TargetMBB = nullptr; // Get the number of bytes to allocate from the FrameInfo. @@ -1490,7 +1530,7 @@ void X86FrameLowering::emitEpilogue(MachineFunction &MF, // Callee-saved registers were pushed on stack before the stack was // realigned. if (TRI->needsStackRealignment(MF) && !IsWin64Prologue) - NumBytes = RoundUpToAlignment(FrameSize, MaxAlign); + NumBytes = alignTo(FrameSize, MaxAlign); // Pop EBP. BuildMI(MBB, MBBI, DL, @@ -1589,15 +1629,17 @@ void X86FrameLowering::emitEpilogue(MachineFunction &MF, if (NeedsWinCFI) BuildMI(MBB, MBBI, DL, TII.get(X86::SEH_Epilogue)); - // Add the return addr area delta back since we are not tail calling. - int Offset = -1 * X86FI->getTCReturnAddrDelta(); - assert(Offset >= 0 && "TCDelta should never be positive"); - if (Offset) { - MBBI = MBB.getFirstTerminator(); + if (!isTailCallOpcode(RetOpcode)) { + // Add the return addr area delta back since we are not tail calling. + int Offset = -1 * X86FI->getTCReturnAddrDelta(); + assert(Offset >= 0 && "TCDelta should never be positive"); + if (Offset) { + MBBI = MBB.getFirstTerminator(); - // Check for possible merge with preceding ADD instruction. - Offset += mergeSPUpdates(MBB, MBBI, true); - emitSPUpdate(MBB, MBBI, Offset, /*InEpilogue=*/true); + // Check for possible merge with preceding ADD instruction. + Offset += mergeSPUpdates(MBB, MBBI, true); + emitSPUpdate(MBB, MBBI, Offset, /*InEpilogue=*/true); + } } } @@ -1689,58 +1731,61 @@ int X86FrameLowering::getFrameIndexReference(const MachineFunction &MF, int FI, return Offset + FPDelta; } -// Simplified from getFrameIndexReference keeping only StackPointer cases -int X86FrameLowering::getFrameIndexReferenceFromSP(const MachineFunction &MF, - int FI, - unsigned &FrameReg) const { +int +X86FrameLowering::getFrameIndexReferencePreferSP(const MachineFunction &MF, + int FI, unsigned &FrameReg, + bool IgnoreSPUpdates) const { + const MachineFrameInfo *MFI = MF.getFrameInfo(); // Does not include any dynamic realign. const uint64_t StackSize = MFI->getStackSize(); - { -#ifndef NDEBUG - // LLVM arranges the stack as follows: - // ... - // ARG2 - // ARG1 - // RETADDR - // PUSH RBP <-- RBP points here - // PUSH CSRs - // ~~~~~~~ <-- possible stack realignment (non-win64) - // ... - // STACK OBJECTS - // ... <-- RSP after prologue points here - // ~~~~~~~ <-- possible stack realignment (win64) - // - // if (hasVarSizedObjects()): - // ... <-- "base pointer" (ESI/RBX) points here - // DYNAMIC ALLOCAS - // ... <-- RSP points here - // - // Case 1: In the simple case of no stack realignment and no dynamic - // allocas, both "fixed" stack objects (arguments and CSRs) are addressable - // with fixed offsets from RSP. - // - // Case 2: In the case of stack realignment with no dynamic allocas, fixed - // stack objects are addressed with RBP and regular stack objects with RSP. - // - // Case 3: In the case of dynamic allocas and stack realignment, RSP is used - // to address stack arguments for outgoing calls and nothing else. The "base - // pointer" points to local variables, and RBP points to fixed objects. - // - // In cases 2 and 3, we can only answer for non-fixed stack objects, and the - // answer we give is relative to the SP after the prologue, and not the - // SP in the middle of the function. - - assert((!MFI->isFixedObjectIndex(FI) || !TRI->needsStackRealignment(MF) || - STI.isTargetWin64()) && - "offset from fixed object to SP is not static"); - - // We don't handle tail calls, and shouldn't be seeing them either. - int TailCallReturnAddrDelta = - MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta(); - assert(!(TailCallReturnAddrDelta < 0) && "we don't handle this case!"); -#endif - } + // LLVM arranges the stack as follows: + // ... + // ARG2 + // ARG1 + // RETADDR + // PUSH RBP <-- RBP points here + // PUSH CSRs + // ~~~~~~~ <-- possible stack realignment (non-win64) + // ... + // STACK OBJECTS + // ... <-- RSP after prologue points here + // ~~~~~~~ <-- possible stack realignment (win64) + // + // if (hasVarSizedObjects()): + // ... <-- "base pointer" (ESI/RBX) points here + // DYNAMIC ALLOCAS + // ... <-- RSP points here + // + // Case 1: In the simple case of no stack realignment and no dynamic + // allocas, both "fixed" stack objects (arguments and CSRs) are addressable + // with fixed offsets from RSP. + // + // Case 2: In the case of stack realignment with no dynamic allocas, fixed + // stack objects are addressed with RBP and regular stack objects with RSP. + // + // Case 3: In the case of dynamic allocas and stack realignment, RSP is used + // to address stack arguments for outgoing calls and nothing else. The "base + // pointer" points to local variables, and RBP points to fixed objects. + // + // In cases 2 and 3, we can only answer for non-fixed stack objects, and the + // answer we give is relative to the SP after the prologue, and not the + // SP in the middle of the function. + + if (MFI->isFixedObjectIndex(FI) && TRI->needsStackRealignment(MF) && + !STI.isTargetWin64()) + return getFrameIndexReference(MF, FI, FrameReg); + + // If !hasReservedCallFrame the function might have SP adjustement in the + // body. So, even though the offset is statically known, it depends on where + // we are in the function. + const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering(); + if (!IgnoreSPUpdates && !TFI->hasReservedCallFrame(MF)) + return getFrameIndexReference(MF, FI, FrameReg); + + // We don't handle tail calls, and shouldn't be seeing them either. + assert(MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta() >= 0 && + "we don't handle this case!"); // Fill in FrameReg output argument. FrameReg = TRI->getStackRegister(); @@ -1851,16 +1896,37 @@ bool X86FrameLowering::spillCalleeSavedRegisters( return true; // Push GPRs. It increases frame size. + const MachineFunction &MF = *MBB.getParent(); unsigned Opc = STI.is64Bit() ? X86::PUSH64r : X86::PUSH32r; for (unsigned i = CSI.size(); i != 0; --i) { unsigned Reg = CSI[i - 1].getReg(); if (!X86::GR64RegClass.contains(Reg) && !X86::GR32RegClass.contains(Reg)) continue; - // Add the callee-saved register as live-in. It's killed at the spill. - MBB.addLiveIn(Reg); - BuildMI(MBB, MI, DL, TII.get(Opc)).addReg(Reg, RegState::Kill) + const MachineRegisterInfo &MRI = MF.getRegInfo(); + bool isLiveIn = MRI.isLiveIn(Reg); + if (!isLiveIn) + MBB.addLiveIn(Reg); + + // Decide whether we can add a kill flag to the use. + bool CanKill = !isLiveIn; + // Check if any subregister is live-in + if (CanKill) { + for (MCRegAliasIterator AReg(Reg, TRI, false); AReg.isValid(); ++AReg) { + if (MRI.isLiveIn(*AReg)) { + CanKill = false; + break; + } + } + } + + // Do not set a kill flag on values that are also marked as live-in. This + // happens with the @llvm-returnaddress intrinsic and with arguments + // passed in callee saved registers. + // Omitting the kill flags is conservatively correct even if the live-in + // is not used after all. + BuildMI(MBB, MI, DL, TII.get(Opc)).addReg(Reg, getKillRegState(CanKill)) .setMIFlag(MachineInstr::FrameSetup); } @@ -1891,7 +1957,7 @@ bool X86FrameLowering::restoreCalleeSavedRegisters(MachineBasicBlock &MBB, if (CSI.empty()) return false; - if (isFuncletReturnInstr(MI) && STI.isOSWindows()) { + if (isFuncletReturnInstr(*MI) && STI.isOSWindows()) { // Don't restore CSRs in 32-bit EH funclets. Matches // spillCalleeSavedRegisters. if (STI.is32Bit()) @@ -2250,11 +2316,33 @@ void X86FrameLowering::adjustForSegmentedStacks( checkMBB->addSuccessor(allocMBB); checkMBB->addSuccessor(&PrologueMBB); -#ifdef XDEBUG +#ifdef EXPENSIVE_CHECKS MF.verify(); #endif } +/// Lookup an ERTS parameter in the !hipe.literals named metadata node. +/// HiPE provides Erlang Runtime System-internal parameters, such as PCB offsets +/// to fields it needs, through a named metadata node "hipe.literals" containing +/// name-value pairs. +static unsigned getHiPELiteral( + NamedMDNode *HiPELiteralsMD, const StringRef LiteralName) { + for (int i = 0, e = HiPELiteralsMD->getNumOperands(); i != e; ++i) { + MDNode *Node = HiPELiteralsMD->getOperand(i); + if (Node->getNumOperands() != 2) continue; + MDString *NodeName = dyn_cast<MDString>(Node->getOperand(0)); + ValueAsMetadata *NodeVal = dyn_cast<ValueAsMetadata>(Node->getOperand(1)); + if (!NodeName || !NodeVal) continue; + ConstantInt *ValConst = dyn_cast_or_null<ConstantInt>(NodeVal->getValue()); + if (ValConst && NodeName->getString() == LiteralName) { + return ValConst->getZExtValue(); + } + } + + report_fatal_error("HiPE literal " + LiteralName + + " required but not provided"); +} + /// Erlang programs may need a special prologue to handle the stack size they /// might need at runtime. That is because Erlang/OTP does not implement a C /// stack but uses a custom implementation of hybrid stack/heap architecture. @@ -2280,7 +2368,14 @@ void X86FrameLowering::adjustForHiPEPrologue( assert(&(*MF.begin()) == &PrologueMBB && "Shrink-wrapping not supported yet"); // HiPE-specific values - const unsigned HipeLeafWords = 24; + NamedMDNode *HiPELiteralsMD = MF.getMMI().getModule() + ->getNamedMetadata("hipe.literals"); + if (!HiPELiteralsMD) + report_fatal_error( + "Can't generate HiPE prologue without runtime parameters"); + const unsigned HipeLeafWords + = getHiPELiteral(HiPELiteralsMD, + Is64Bit ? "AMD64_LEAF_WORDS" : "X86_LEAF_WORDS"); const unsigned CCRegisteredArgs = Is64Bit ? 6 : 5; const unsigned Guaranteed = HipeLeafWords * SlotSize; unsigned CallerStkArity = MF.getFunction()->arg_size() > CCRegisteredArgs ? @@ -2300,15 +2395,13 @@ void X86FrameLowering::adjustForHiPEPrologue( if (MFI->hasCalls()) { unsigned MoreStackForCalls = 0; - for (MachineFunction::iterator MBBI = MF.begin(), MBBE = MF.end(); - MBBI != MBBE; ++MBBI) - for (MachineBasicBlock::iterator MI = MBBI->begin(), ME = MBBI->end(); - MI != ME; ++MI) { - if (!MI->isCall()) + for (auto &MBB : MF) { + for (auto &MI : MBB) { + if (!MI.isCall()) continue; // Get callee operand. - const MachineOperand &MO = MI->getOperand(0); + const MachineOperand &MO = MI.getOperand(0); // Only take account of global function calls (no closures etc.). if (!MO.isGlobal()) @@ -2334,6 +2427,7 @@ void X86FrameLowering::adjustForHiPEPrologue( MoreStackForCalls = std::max(MoreStackForCalls, (HipeLeafWords - 1 - CalleeStkArity) * SlotSize); } + } MaxStack += MoreStackForCalls; } @@ -2353,20 +2447,19 @@ void X86FrameLowering::adjustForHiPEPrologue( unsigned ScratchReg, SPReg, PReg, SPLimitOffset; unsigned LEAop, CMPop, CALLop; + SPLimitOffset = getHiPELiteral(HiPELiteralsMD, "P_NSP_LIMIT"); if (Is64Bit) { SPReg = X86::RSP; PReg = X86::RBP; LEAop = X86::LEA64r; CMPop = X86::CMP64rm; CALLop = X86::CALL64pcrel32; - SPLimitOffset = 0x90; } else { SPReg = X86::ESP; PReg = X86::EBP; LEAop = X86::LEA32r; CMPop = X86::CMP32rm; CALLop = X86::CALLpcrel32; - SPLimitOffset = 0x4c; } ScratchReg = GetScratchRegister(Is64Bit, IsLP64, MF, true); @@ -2395,13 +2488,15 @@ void X86FrameLowering::adjustForHiPEPrologue( incStackMBB->addSuccessor(&PrologueMBB, {99, 100}); incStackMBB->addSuccessor(incStackMBB, {1, 100}); } -#ifdef XDEBUG +#ifdef EXPENSIVE_CHECKS MF.verify(); #endif } bool X86FrameLowering::adjustStackWithPops(MachineBasicBlock &MBB, - MachineBasicBlock::iterator MBBI, DebugLoc DL, int Offset) const { + MachineBasicBlock::iterator MBBI, + const DebugLoc &DL, + int Offset) const { if (Offset <= 0) return false; @@ -2440,7 +2535,8 @@ bool X86FrameLowering::adjustStackWithPops(MachineBasicBlock &MBB, bool IsDef = false; for (const MachineOperand &MO : Prev->implicit_operands()) { - if (MO.isReg() && MO.isDef() && MO.getReg() == Candidate) { + if (MO.isReg() && MO.isDef() && + TRI->isSuperOrSubRegisterEq(MO.getReg(), Candidate)) { IsDef = true; break; } @@ -2468,7 +2564,7 @@ bool X86FrameLowering::adjustStackWithPops(MachineBasicBlock &MBB, return true; } -void X86FrameLowering:: +MachineBasicBlock::iterator X86FrameLowering:: eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, MachineBasicBlock::iterator I) const { bool reserveCallFrame = hasReservedCallFrame(MF); @@ -2488,7 +2584,7 @@ eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, // amount of space needed for the outgoing arguments up to the next // alignment boundary. unsigned StackAlign = getStackAlignment(); - Amount = RoundUpToAlignment(Amount, StackAlign); + Amount = alignTo(Amount, StackAlign); MachineModuleInfo &MMI = MF.getMMI(); const Function *Fn = MF.getFunction(); @@ -2512,7 +2608,7 @@ eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, MCCFIInstruction::createGnuArgsSize(nullptr, Amount)); if (Amount == 0) - return; + return I; // Factor out the amount that gets handled inside the sequence // (Pushes of argument for frame setup, callee pops for frame destroy) @@ -2525,13 +2621,23 @@ eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, BuildCFI(MBB, I, DL, MCCFIInstruction::createAdjustCfaOffset(nullptr, -InternalAmt)); - if (Amount) { - // Add Amount to SP to destroy a frame, and subtract to setup. - int Offset = isDestroy ? Amount : -Amount; - - if (!(Fn->optForMinSize() && - adjustStackWithPops(MBB, I, DL, Offset))) - BuildStackAdjustment(MBB, I, DL, Offset, /*InEpilogue=*/false); + // Add Amount to SP to destroy a frame, or subtract to setup. + int64_t StackAdjustment = isDestroy ? Amount : -Amount; + int64_t CfaAdjustment = -StackAdjustment; + + if (StackAdjustment) { + // Merge with any previous or following adjustment instruction. Note: the + // instructions merged with here do not have CFI, so their stack + // adjustments do not feed into CfaAdjustment. + StackAdjustment += mergeSPUpdates(MBB, I, true); + StackAdjustment += mergeSPUpdates(MBB, I, false); + + if (StackAdjustment) { + if (!(Fn->optForMinSize() && + adjustStackWithPops(MBB, I, DL, StackAdjustment))) + BuildStackAdjustment(MBB, I, DL, StackAdjustment, + /*InEpilogue=*/false); + } } if (DwarfCFI && !hasFP(MF)) { @@ -2541,18 +2647,16 @@ eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, // CFI only for EH purposes or for debugging. EH only requires the CFA // offset to be correct at each call site, while for debugging we want // it to be more precise. - int CFAOffset = Amount; + // TODO: When not using precise CFA, we also need to adjust for the // InternalAmt here. - - if (CFAOffset) { - CFAOffset = isDestroy ? -CFAOffset : CFAOffset; - BuildCFI(MBB, I, DL, - MCCFIInstruction::createAdjustCfaOffset(nullptr, CFAOffset)); + if (CfaAdjustment) { + BuildCFI(MBB, I, DL, MCCFIInstruction::createAdjustCfaOffset( + nullptr, CfaAdjustment)); } } - return; + return I; } if (isDestroy && InternalAmt) { @@ -2562,11 +2666,20 @@ eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, // We are not tracking the stack pointer adjustment by the callee, so make // sure we restore the stack pointer immediately after the call, there may // be spill code inserted between the CALL and ADJCALLSTACKUP instructions. + MachineBasicBlock::iterator CI = I; MachineBasicBlock::iterator B = MBB.begin(); - while (I != B && !std::prev(I)->isCall()) - --I; - BuildStackAdjustment(MBB, I, DL, -InternalAmt, /*InEpilogue=*/false); + while (CI != B && !std::prev(CI)->isCall()) + --CI; + BuildStackAdjustment(MBB, CI, DL, -InternalAmt, /*InEpilogue=*/false); } + + return I; +} + +bool X86FrameLowering::canUseAsPrologue(const MachineBasicBlock &MBB) const { + assert(MBB.getParent() && "Block is not attached to a function!"); + const MachineFunction &MF = *MBB.getParent(); + return !TRI->needsStackRealignment(MF) || !MBB.isLiveIn(X86::EFLAGS); } bool X86FrameLowering::canUseAsEpilogue(const MachineBasicBlock &MBB) const { @@ -2604,7 +2717,7 @@ bool X86FrameLowering::enableShrinkWrapping(const MachineFunction &MF) const { MachineBasicBlock::iterator X86FrameLowering::restoreWin32EHStackPointers( MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, - DebugLoc DL, bool RestoreSP) const { + const DebugLoc &DL, bool RestoreSP) const { assert(STI.isTargetWindowsMSVC() && "funclets only supported in MSVC env"); assert(STI.isTargetWin32() && "EBP/ESI restoration only required on win32"); assert(STI.is32Bit() && !Uses64BitFramePtr && @@ -2664,6 +2777,150 @@ MachineBasicBlock::iterator X86FrameLowering::restoreWin32EHStackPointers( return MBBI; } +namespace { +// Struct used by orderFrameObjects to help sort the stack objects. +struct X86FrameSortingObject { + bool IsValid = false; // true if we care about this Object. + unsigned ObjectIndex = 0; // Index of Object into MFI list. + unsigned ObjectSize = 0; // Size of Object in bytes. + unsigned ObjectAlignment = 1; // Alignment of Object in bytes. + unsigned ObjectNumUses = 0; // Object static number of uses. +}; + +// The comparison function we use for std::sort to order our local +// stack symbols. The current algorithm is to use an estimated +// "density". This takes into consideration the size and number of +// uses each object has in order to roughly minimize code size. +// So, for example, an object of size 16B that is referenced 5 times +// will get higher priority than 4 4B objects referenced 1 time each. +// It's not perfect and we may be able to squeeze a few more bytes out of +// it (for example : 0(esp) requires fewer bytes, symbols allocated at the +// fringe end can have special consideration, given their size is less +// important, etc.), but the algorithmic complexity grows too much to be +// worth the extra gains we get. This gets us pretty close. +// The final order leaves us with objects with highest priority going +// at the end of our list. +struct X86FrameSortingComparator { + inline bool operator()(const X86FrameSortingObject &A, + const X86FrameSortingObject &B) { + uint64_t DensityAScaled, DensityBScaled; + + // For consistency in our comparison, all invalid objects are placed + // at the end. This also allows us to stop walking when we hit the + // first invalid item after it's all sorted. + if (!A.IsValid) + return false; + if (!B.IsValid) + return true; + + // The density is calculated by doing : + // (double)DensityA = A.ObjectNumUses / A.ObjectSize + // (double)DensityB = B.ObjectNumUses / B.ObjectSize + // Since this approach may cause inconsistencies in + // the floating point <, >, == comparisons, depending on the floating + // point model with which the compiler was built, we're going + // to scale both sides by multiplying with + // A.ObjectSize * B.ObjectSize. This ends up factoring away + // the division and, with it, the need for any floating point + // arithmetic. + DensityAScaled = static_cast<uint64_t>(A.ObjectNumUses) * + static_cast<uint64_t>(B.ObjectSize); + DensityBScaled = static_cast<uint64_t>(B.ObjectNumUses) * + static_cast<uint64_t>(A.ObjectSize); + + // If the two densities are equal, prioritize highest alignment + // objects. This allows for similar alignment objects + // to be packed together (given the same density). + // There's room for improvement here, also, since we can pack + // similar alignment (different density) objects next to each + // other to save padding. This will also require further + // complexity/iterations, and the overall gain isn't worth it, + // in general. Something to keep in mind, though. + if (DensityAScaled == DensityBScaled) + return A.ObjectAlignment < B.ObjectAlignment; + + return DensityAScaled < DensityBScaled; + } +}; +} // namespace + +// Order the symbols in the local stack. +// We want to place the local stack objects in some sort of sensible order. +// The heuristic we use is to try and pack them according to static number +// of uses and size of object in order to minimize code size. +void X86FrameLowering::orderFrameObjects( + const MachineFunction &MF, SmallVectorImpl<int> &ObjectsToAllocate) const { + const MachineFrameInfo *MFI = MF.getFrameInfo(); + + // Don't waste time if there's nothing to do. + if (ObjectsToAllocate.empty()) + return; + + // Create an array of all MFI objects. We won't need all of these + // objects, but we're going to create a full array of them to make + // it easier to index into when we're counting "uses" down below. + // We want to be able to easily/cheaply access an object by simply + // indexing into it, instead of having to search for it every time. + std::vector<X86FrameSortingObject> SortingObjects(MFI->getObjectIndexEnd()); + + // Walk the objects we care about and mark them as such in our working + // struct. + for (auto &Obj : ObjectsToAllocate) { + SortingObjects[Obj].IsValid = true; + SortingObjects[Obj].ObjectIndex = Obj; + SortingObjects[Obj].ObjectAlignment = MFI->getObjectAlignment(Obj); + // Set the size. + int ObjectSize = MFI->getObjectSize(Obj); + if (ObjectSize == 0) + // Variable size. Just use 4. + SortingObjects[Obj].ObjectSize = 4; + else + SortingObjects[Obj].ObjectSize = ObjectSize; + } + + // Count the number of uses for each object. + for (auto &MBB : MF) { + for (auto &MI : MBB) { + if (MI.isDebugValue()) + continue; + for (const MachineOperand &MO : MI.operands()) { + // Check to see if it's a local stack symbol. + if (!MO.isFI()) + continue; + int Index = MO.getIndex(); + // Check to see if it falls within our range, and is tagged + // to require ordering. + if (Index >= 0 && Index < MFI->getObjectIndexEnd() && + SortingObjects[Index].IsValid) + SortingObjects[Index].ObjectNumUses++; + } + } + } + + // Sort the objects using X86FrameSortingAlgorithm (see its comment for + // info). + std::stable_sort(SortingObjects.begin(), SortingObjects.end(), + X86FrameSortingComparator()); + + // Now modify the original list to represent the final order that + // we want. The order will depend on whether we're going to access them + // from the stack pointer or the frame pointer. For SP, the list should + // end up with the END containing objects that we want with smaller offsets. + // For FP, it should be flipped. + int i = 0; + for (auto &Obj : SortingObjects) { + // All invalid items are sorted at the end, so it's safe to stop. + if (!Obj.IsValid) + break; + ObjectsToAllocate[i++] = Obj.ObjectIndex; + } + + // Flip it if we're accessing off of the FP. + if (!TRI->needsStackRealignment(MF) && hasFP(MF)) + std::reverse(ObjectsToAllocate.begin(), ObjectsToAllocate.end()); +} + + unsigned X86FrameLowering::getWinEHParentFrameOffset(const MachineFunction &MF) const { // RDX, the parent frame pointer, is homed into 16(%rsp) in the prologue. unsigned Offset = 16; @@ -2691,14 +2948,30 @@ void X86FrameLowering::processFunctionBeforeFrameFinalized( // were no fixed objects, use offset -SlotSize, which is immediately after the // return address. Fixed objects have negative frame indices. MachineFrameInfo *MFI = MF.getFrameInfo(); + WinEHFuncInfo &EHInfo = *MF.getWinEHFuncInfo(); int64_t MinFixedObjOffset = -SlotSize; for (int I = MFI->getObjectIndexBegin(); I < 0; ++I) MinFixedObjOffset = std::min(MinFixedObjOffset, MFI->getObjectOffset(I)); + for (WinEHTryBlockMapEntry &TBME : EHInfo.TryBlockMap) { + for (WinEHHandlerType &H : TBME.HandlerArray) { + int FrameIndex = H.CatchObj.FrameIndex; + if (FrameIndex != INT_MAX) { + // Ensure alignment. + unsigned Align = MFI->getObjectAlignment(FrameIndex); + MinFixedObjOffset -= std::abs(MinFixedObjOffset) % Align; + MinFixedObjOffset -= MFI->getObjectSize(FrameIndex); + MFI->setObjectOffset(FrameIndex, MinFixedObjOffset); + } + } + } + + // Ensure alignment. + MinFixedObjOffset -= std::abs(MinFixedObjOffset) % 8; int64_t UnwindHelpOffset = MinFixedObjOffset - SlotSize; int UnwindHelpFI = MFI->CreateFixedObject(SlotSize, UnwindHelpOffset, /*Immutable=*/false); - MF.getWinEHFuncInfo()->UnwindHelpFrameIdx = UnwindHelpFI; + EHInfo.UnwindHelpFrameIdx = UnwindHelpFI; // Store -2 into UnwindHelp on function entry. We have to scan forwards past // other frame setup instructions. |