X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FMC%2FMCAssembler.cpp;h=35f39564cd5565f7ecbffb4d584767ebe7a6902b;hb=a8bf473fb15bbedc78144ec57438f884616283b0;hp=432e44a24aa327d0fc2e129cc614c67c2fcbbeed;hpb=2a6e3f5be21fdcb1fa81533ec81ba879df815607;p=oota-llvm.git diff --git a/lib/MC/MCAssembler.cpp b/lib/MC/MCAssembler.cpp index 432e44a24aa..35f39564cd5 100644 --- a/lib/MC/MCAssembler.cpp +++ b/lib/MC/MCAssembler.cpp @@ -19,19 +19,26 @@ #include "llvm/ADT/Statistic.h" #include "llvm/ADT/StringExtras.h" #include "llvm/ADT/Twine.h" +#include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/Support/Debug.h" #include "llvm/Target/TargetRegistry.h" #include "llvm/Target/TargetAsmBackend.h" -// FIXME: Gross. -#include "../Target/X86/X86FixupKinds.h" - #include using namespace llvm; +namespace { +namespace stats { STATISTIC(EmittedFragments, "Number of emitted assembler fragments"); +STATISTIC(EvaluateFixup, "Number of evaluated fixups"); +STATISTIC(FragmentLayouts, "Number of fragment layouts"); +STATISTIC(ObjectBytes, "Number of emitted object file bytes"); +STATISTIC(RelaxationSteps, "Number of assembler layout and relaxation steps"); +STATISTIC(RelaxedInstructions, "Number of relaxed instructions"); +STATISTIC(SectionLayouts, "Number of section layouts"); +} +} // FIXME FIXME FIXME: There are number of places in this file where we convert // what is a 64-bit assembler value used for computation into a value in the @@ -40,24 +47,146 @@ STATISTIC(EmittedFragments, "Number of emitted assembler fragments"); /* *** */ -MCFragment::MCFragment() : Kind(FragmentType(~0)) { +MCAsmLayout::MCAsmLayout(MCAssembler &Asm) + : Assembler(Asm), LastValidFragment(0) + { + // Compute the section layout order. Virtual sections must go last. + for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it) + if (!Asm.getBackend().isVirtualSection(it->getSection())) + SectionOrder.push_back(&*it); + for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it) + if (Asm.getBackend().isVirtualSection(it->getSection())) + SectionOrder.push_back(&*it); } -MCFragment::MCFragment(FragmentType _Kind, MCSectionData *_Parent) - : Kind(_Kind), - Parent(_Parent), - FileSize(~UINT64_C(0)) -{ - if (Parent) - Parent->getFragmentList().push_back(this); +bool MCAsmLayout::isSectionUpToDate(const MCSectionData *SD) const { + // The first section is always up-to-date. + unsigned Index = SD->getLayoutOrder(); + if (!Index) + return true; + + // Otherwise, sections are always implicitly computed when the preceeding + // fragment is layed out. + const MCSectionData *Prev = getSectionOrder()[Index - 1]; + return isFragmentUpToDate(&(Prev->getFragmentList().back())); +} + +bool MCAsmLayout::isFragmentUpToDate(const MCFragment *F) const { + return (LastValidFragment && + F->getLayoutOrder() <= LastValidFragment->getLayoutOrder()); +} + +void MCAsmLayout::UpdateForSlide(MCFragment *F, int SlideAmount) { + // If this fragment wasn't already up-to-date, we don't need to do anything. + if (!isFragmentUpToDate(F)) + return; + + // Otherwise, reset the last valid fragment to the predecessor of the + // invalidated fragment. + LastValidFragment = F->getPrevNode(); + if (!LastValidFragment) { + unsigned Index = F->getParent()->getLayoutOrder(); + if (Index != 0) { + MCSectionData *Prev = getSectionOrder()[Index - 1]; + LastValidFragment = &(Prev->getFragmentList().back()); + } + } +} + +void MCAsmLayout::EnsureValid(const MCFragment *F) const { + // Advance the layout position until the fragment is up-to-date. + while (!isFragmentUpToDate(F)) { + // Advance to the next fragment. + MCFragment *Cur = LastValidFragment; + if (Cur) + Cur = Cur->getNextNode(); + if (!Cur) { + unsigned NextIndex = 0; + if (LastValidFragment) + NextIndex = LastValidFragment->getParent()->getLayoutOrder() + 1; + Cur = SectionOrder[NextIndex]->begin(); + } + + const_cast(this)->LayoutFragment(Cur); + } +} + +void MCAsmLayout::FragmentReplaced(MCFragment *Src, MCFragment *Dst) { + if (LastValidFragment == Src) + LastValidFragment = Dst; + + Dst->Offset = Src->Offset; + Dst->EffectiveSize = Src->EffectiveSize; +} + +uint64_t MCAsmLayout::getFragmentAddress(const MCFragment *F) const { + assert(F->getParent() && "Missing section()!"); + return getSectionAddress(F->getParent()) + getFragmentOffset(F); +} + +uint64_t MCAsmLayout::getFragmentEffectiveSize(const MCFragment *F) const { + EnsureValid(F); + assert(F->EffectiveSize != ~UINT64_C(0) && "Address not set!"); + return F->EffectiveSize; +} + +uint64_t MCAsmLayout::getFragmentOffset(const MCFragment *F) const { + EnsureValid(F); + assert(F->Offset != ~UINT64_C(0) && "Address not set!"); + return F->Offset; +} + +uint64_t MCAsmLayout::getSymbolAddress(const MCSymbolData *SD) const { + assert(SD->getFragment() && "Invalid getAddress() on undefined symbol!"); + return getFragmentAddress(SD->getFragment()) + SD->getOffset(); +} + +uint64_t MCAsmLayout::getSectionAddress(const MCSectionData *SD) const { + EnsureValid(SD->begin()); + assert(SD->Address != ~UINT64_C(0) && "Address not set!"); + return SD->Address; +} + +uint64_t MCAsmLayout::getSectionAddressSize(const MCSectionData *SD) const { + // The size is the last fragment's end offset. + const MCFragment &F = SD->getFragmentList().back(); + return getFragmentOffset(&F) + getFragmentEffectiveSize(&F); +} + +uint64_t MCAsmLayout::getSectionFileSize(const MCSectionData *SD) const { + // Virtual sections have no file size. + if (getAssembler().getBackend().isVirtualSection(SD->getSection())) + return 0; + + // Otherwise, the file size is the same as the address space size. + return getSectionAddressSize(SD); +} + +uint64_t MCAsmLayout::getSectionSize(const MCSectionData *SD) const { + // The logical size is the address space size minus any tail padding. + uint64_t Size = getSectionAddressSize(SD); + const MCAlignFragment *AF = + dyn_cast(&(SD->getFragmentList().back())); + if (AF && AF->hasOnlyAlignAddress()) + Size -= getFragmentEffectiveSize(AF); + + return Size; +} + +/* *** */ + +MCFragment::MCFragment() : Kind(FragmentType(~0)) { } MCFragment::~MCFragment() { } -uint64_t MCFragment::getAddress() const { - assert(getParent() && "Missing Section!"); - return getParent()->getAddress() + Offset; +MCFragment::MCFragment(FragmentType _Kind, MCSectionData *_Parent) + : Kind(_Kind), Parent(_Parent), Atom(0), Offset(~UINT64_C(0)), + EffectiveSize(~UINT64_C(0)) +{ + if (Parent) + Parent->getFragmentList().push_back(this); } /* *** */ @@ -68,8 +197,6 @@ MCSectionData::MCSectionData(const MCSection &_Section, MCAssembler *A) : Section(&_Section), Alignment(1), Address(~UINT64_C(0)), - Size(~UINT64_C(0)), - FileSize(~UINT64_C(0)), HasInstructions(false) { if (A) @@ -84,7 +211,8 @@ MCSymbolData::MCSymbolData(const MCSymbol &_Symbol, MCFragment *_Fragment, uint64_t _Offset, MCAssembler *A) : Symbol(&_Symbol), Fragment(_Fragment), Offset(_Offset), IsExternal(false), IsPrivateExtern(false), - CommonSize(0), CommonAlign(0), Flags(0), Index(0) + CommonSize(0), SymbolSize(0), CommonAlign(0), + Flags(0), Index(0) { if (A) A->getSymbolList().push_back(this); @@ -95,7 +223,7 @@ MCSymbolData::MCSymbolData(const MCSymbol &_Symbol, MCFragment *_Fragment, MCAssembler::MCAssembler(MCContext &_Context, TargetAsmBackend &_Backend, MCCodeEmitter &_Emitter, raw_ostream &_OS) : Context(_Context), Backend(_Backend), Emitter(_Emitter), - OS(_OS), SubsectionsViaSymbols(false) + OS(_OS), RelaxAll(false), SubsectionsViaSymbols(false) { } @@ -103,8 +231,7 @@ MCAssembler::~MCAssembler() { } static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm, - const MCAsmFixup &Fixup, - const MCDataFragment *DF, + const MCFixup &Fixup, const MCValue Target, const MCSection *BaseSection) { // The effective fixup address is @@ -121,7 +248,7 @@ static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm, // resolved. This also works in conjunction with absolutized .set, which // requires the compiler to use .set to absolutize the differences between // symbols which the compiler knows to be assembly time constants, so we don't - // need to worry about consider symbol differences fully resolved. + // need to worry about considering symbol differences fully resolved. // Non-relative fixups are only resolved if constant. if (!BaseSection) @@ -141,8 +268,8 @@ static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm, } static bool isScatteredFixupFullyResolved(const MCAssembler &Asm, - const MCAsmFixup &Fixup, - const MCDataFragment *DF, + const MCAsmLayout &Layout, + const MCFixup &Fixup, const MCValue Target, const MCSymbolData *BaseSymbol) { // The effective fixup address is @@ -162,7 +289,7 @@ static bool isScatteredFixupFullyResolved(const MCAssembler &Asm, if (A->getKind() != MCSymbolRefExpr::VK_None) return false; - A_Base = Asm.getAtom(&Asm.getSymbolData(A->getSymbol())); + A_Base = Asm.getAtom(Layout, &Asm.getSymbolData(A->getSymbol())); if (!A_Base) return false; } @@ -172,7 +299,7 @@ static bool isScatteredFixupFullyResolved(const MCAssembler &Asm, if (B->getKind() != MCSymbolRefExpr::VK_None) return false; - B_Base = Asm.getAtom(&Asm.getSymbolData(B->getSymbol())); + B_Base = Asm.getAtom(Layout, &Asm.getSymbolData(B->getSymbol())); if (!B_Base) return false; } @@ -186,61 +313,46 @@ static bool isScatteredFixupFullyResolved(const MCAssembler &Asm, return !B_Base && BaseSymbol == A_Base; } -bool MCAssembler::isSymbolLinkerVisible(const MCSymbolData *SD) const { +bool MCAssembler::isSymbolLinkerVisible(const MCSymbol &Symbol) const { // Non-temporary labels should always be visible to the linker. - if (!SD->getSymbol().isTemporary()) + if (!Symbol.isTemporary()) return true; // Absolute temporary labels are never visible. - if (!SD->getFragment()) + if (!Symbol.isInSection()) return false; // Otherwise, check if the section requires symbols even for temporary labels. - return getBackend().doesSectionRequireSymbols( - SD->getFragment()->getParent()->getSection()); -} - -const MCSymbolData *MCAssembler::getAtomForAddress(const MCSectionData *Section, - uint64_t Address) const { - const MCSymbolData *Best = 0; - for (MCAssembler::const_symbol_iterator it = symbol_begin(), - ie = symbol_end(); it != ie; ++it) { - // Ignore non-linker visible symbols. - if (!isSymbolLinkerVisible(it)) - continue; - - // Ignore symbols not in the same section. - if (!it->getFragment() || it->getFragment()->getParent() != Section) - continue; - - // Otherwise, find the closest symbol preceding this address (ties are - // resolved in favor of the last defined symbol). - if (it->getAddress() <= Address && - (!Best || it->getAddress() >= Best->getAddress())) - Best = it; - } - - return Best; + return getBackend().doesSectionRequireSymbols(Symbol.getSection()); } -const MCSymbolData *MCAssembler::getAtom(const MCSymbolData *SD) const { +const MCSymbolData *MCAssembler::getAtom(const MCAsmLayout &Layout, + const MCSymbolData *SD) const { // Linker visible symbols define atoms. - if (isSymbolLinkerVisible(SD)) + if (isSymbolLinkerVisible(SD->getSymbol())) return SD; // Absolute and undefined symbols have no defining atom. if (!SD->getFragment()) return 0; - // Otherwise, search by address. - return getAtomForAddress(SD->getFragment()->getParent(), SD->getAddress()); + // Non-linker visible symbols in sections which can't be atomized have no + // defining atom. + if (!getBackend().isSectionAtomizable( + SD->getFragment()->getParent()->getSection())) + return 0; + + // Otherwise, return the atom for the containing fragment. + return SD->getFragment()->getAtom(); } -bool MCAssembler::EvaluateFixup(const MCAsmLayout &Layout, MCAsmFixup &Fixup, - MCDataFragment *DF, +bool MCAssembler::EvaluateFixup(const MCAsmLayout &Layout, + const MCFixup &Fixup, const MCFragment *DF, MCValue &Target, uint64_t &Value) const { - if (!Fixup.Value->EvaluateAsRelocatable(Target, &Layout)) - llvm_report_error("expected relocatable expression"); + ++stats::EvaluateFixup; + + if (!Fixup.getValue()->EvaluateAsRelocatable(Target, &Layout)) + report_fatal_error("expected relocatable expression"); // FIXME: How do non-scattered symbols work in ELF? I presume the linker // doesn't support small relocations, but then under what criteria does the @@ -248,18 +360,18 @@ bool MCAssembler::EvaluateFixup(const MCAsmLayout &Layout, MCAsmFixup &Fixup, Value = Target.getConstant(); - bool IsPCRel = - Emitter.getFixupKindInfo(Fixup.Kind).Flags & MCFixupKindInfo::FKF_IsPCRel; + bool IsPCRel = Emitter.getFixupKindInfo( + Fixup.getKind()).Flags & MCFixupKindInfo::FKF_IsPCRel; bool IsResolved = true; if (const MCSymbolRefExpr *A = Target.getSymA()) { if (A->getSymbol().isDefined()) - Value += getSymbolData(A->getSymbol()).getAddress(); + Value += Layout.getSymbolAddress(&getSymbolData(A->getSymbol())); else IsResolved = false; } if (const MCSymbolRefExpr *B = Target.getSymB()) { if (B->getSymbol().isDefined()) - Value -= getSymbolData(B->getSymbol()).getAddress(); + Value -= Layout.getSymbolAddress(&getSymbolData(B->getSymbol())); else IsResolved = false; } @@ -272,193 +384,182 @@ bool MCAssembler::EvaluateFixup(const MCAsmLayout &Layout, MCAsmFixup &Fixup, // symbol) that the fixup value is relative to. const MCSymbolData *BaseSymbol = 0; if (IsPCRel) { - BaseSymbol = getAtomForAddress( - DF->getParent(), DF->getAddress() + Fixup.Offset); + BaseSymbol = DF->getAtom(); if (!BaseSymbol) IsResolved = false; } if (IsResolved) - IsResolved = isScatteredFixupFullyResolved(*this, Fixup, DF, Target, + IsResolved = isScatteredFixupFullyResolved(*this, Layout, Fixup, Target, BaseSymbol); } else { const MCSection *BaseSection = 0; if (IsPCRel) BaseSection = &DF->getParent()->getSection(); - IsResolved = isScatteredFixupFullyResolvedSimple(*this, Fixup, DF, Target, + IsResolved = isScatteredFixupFullyResolvedSimple(*this, Fixup, Target, BaseSection); } } if (IsPCRel) - Value -= DF->getAddress() + Fixup.Offset; + Value -= Layout.getFragmentAddress(DF) + Fixup.getOffset(); return IsResolved; } -void MCAssembler::LayoutSection(MCSectionData &SD, - MCAsmLayout &Layout) { - uint64_t Address = SD.getAddress(); +uint64_t MCAssembler::ComputeFragmentSize(MCAsmLayout &Layout, + const MCFragment &F, + uint64_t SectionAddress, + uint64_t FragmentOffset) const { + switch (F.getKind()) { + case MCFragment::FT_Data: + return cast(F).getContents().size(); + case MCFragment::FT_Fill: + return cast(F).getSize(); + case MCFragment::FT_Inst: + return cast(F).getInstSize(); - for (MCSectionData::iterator it = SD.begin(), ie = SD.end(); it != ie; ++it) { - MCFragment &F = *it; + case MCFragment::FT_Align: { + const MCAlignFragment &AF = cast(F); - F.setOffset(Address - SD.getAddress()); + assert((!AF.hasOnlyAlignAddress() || !AF.getNextNode()) && + "Invalid OnlyAlignAddress bit, not the last fragment!"); - // Evaluate fragment size. - switch (F.getKind()) { - case MCFragment::FT_Align: { - MCAlignFragment &AF = cast(F); + uint64_t Size = OffsetToAlignment(SectionAddress + FragmentOffset, + AF.getAlignment()); - uint64_t Size = OffsetToAlignment(Address, AF.getAlignment()); - if (Size > AF.getMaxBytesToEmit()) - AF.setFileSize(0); - else - AF.setFileSize(Size); - break; - } + // Honor MaxBytesToEmit. + if (Size > AF.getMaxBytesToEmit()) + return 0; - case MCFragment::FT_Data: - F.setFileSize(cast(F).getContents().size()); - break; + return Size; + } - case MCFragment::FT_Fill: { - MCFillFragment &FF = cast(F); - F.setFileSize(FF.getValueSize() * FF.getCount()); - break; - } + case MCFragment::FT_Org: { + const MCOrgFragment &OF = cast(F); - case MCFragment::FT_Org: { - MCOrgFragment &OF = cast(F); + // FIXME: We should compute this sooner, we don't want to recurse here, and + // we would like to be more functional. + int64_t TargetLocation; + if (!OF.getOffset().EvaluateAsAbsolute(TargetLocation, &Layout)) + report_fatal_error("expected assembly-time absolute expression"); - int64_t TargetLocation; - if (!OF.getOffset().EvaluateAsAbsolute(TargetLocation, &Layout)) - llvm_report_error("expected assembly-time absolute expression"); + // FIXME: We need a way to communicate this error. + int64_t Offset = TargetLocation - FragmentOffset; + if (Offset < 0) + report_fatal_error("invalid .org offset '" + Twine(TargetLocation) + + "' (at offset '" + Twine(FragmentOffset) + "'"); - // FIXME: We need a way to communicate this error. - int64_t Offset = TargetLocation - F.getOffset(); - if (Offset < 0) - llvm_report_error("invalid .org offset '" + Twine(TargetLocation) + - "' (at offset '" + Twine(F.getOffset()) + "'"); + return Offset; + } + } - F.setFileSize(Offset); - break; - } + assert(0 && "invalid fragment kind"); + return 0; +} - case MCFragment::FT_ZeroFill: { - MCZeroFillFragment &ZFF = cast(F); +void MCAsmLayout::LayoutFile() { + // Initialize the first section and set the valid fragment layout point. All + // actual layout computations are done lazily. + LastValidFragment = 0; + if (!getSectionOrder().empty()) + getSectionOrder().front()->Address = 0; +} - // Align the fragment offset; it is safe to adjust the offset freely since - // this is only in virtual sections. - Address = RoundUpToAlignment(Address, ZFF.getAlignment()); - F.setOffset(Address - SD.getAddress()); +void MCAsmLayout::LayoutFragment(MCFragment *F) { + MCFragment *Prev = F->getPrevNode(); + + // We should never try to recompute something which is up-to-date. + assert(!isFragmentUpToDate(F) && "Attempt to recompute up-to-date fragment!"); + // We should never try to compute the fragment layout if the section isn't + // up-to-date. + assert(isSectionUpToDate(F->getParent()) && + "Attempt to compute fragment before it's section!"); + // We should never try to compute the fragment layout if it's predecessor + // isn't up-to-date. + assert((!Prev || isFragmentUpToDate(Prev)) && + "Attempt to compute fragment before it's predecessor!"); + + ++stats::FragmentLayouts; + + // Compute the fragment start address. + uint64_t StartAddress = F->getParent()->Address; + uint64_t Address = StartAddress; + if (Prev) + Address += Prev->Offset + Prev->EffectiveSize; + + // Compute fragment offset and size. + F->Offset = Address - StartAddress; + F->EffectiveSize = getAssembler().ComputeFragmentSize(*this, *F, StartAddress, + F->Offset); + LastValidFragment = F; + + // If this is the last fragment in a section, update the next section address. + if (!F->getNextNode()) { + unsigned NextIndex = F->getParent()->getLayoutOrder() + 1; + if (NextIndex != getSectionOrder().size()) + LayoutSection(getSectionOrder()[NextIndex]); + } +} - // FIXME: This is misnamed. - F.setFileSize(ZFF.getSize()); - break; - } - } +void MCAsmLayout::LayoutSection(MCSectionData *SD) { + unsigned SectionOrderIndex = SD->getLayoutOrder(); - Address += F.getFileSize(); - } - - // Set the section sizes. - SD.setSize(Address - SD.getAddress()); - if (getBackend().isVirtualSection(SD.getSection())) - SD.setFileSize(0); - else - SD.setFileSize(Address - SD.getAddress()); -} - -/// WriteNopData - Write optimal nops to the output file for the \arg Count -/// bytes. This returns the number of bytes written. It may return 0 if -/// the \arg Count is more than the maximum optimal nops. -/// -/// FIXME this is X86 32-bit specific and should move to a better place. -static uint64_t WriteNopData(uint64_t Count, MCObjectWriter *OW) { - static const uint8_t Nops[16][16] = { - // nop - {0x90}, - // xchg %ax,%ax - {0x66, 0x90}, - // nopl (%[re]ax) - {0x0f, 0x1f, 0x00}, - // nopl 0(%[re]ax) - {0x0f, 0x1f, 0x40, 0x00}, - // nopl 0(%[re]ax,%[re]ax,1) - {0x0f, 0x1f, 0x44, 0x00, 0x00}, - // nopw 0(%[re]ax,%[re]ax,1) - {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00}, - // nopl 0L(%[re]ax) - {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00}, - // nopl 0L(%[re]ax,%[re]ax,1) - {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}, - // nopw 0L(%[re]ax,%[re]ax,1) - {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}, - // nopw %cs:0L(%[re]ax,%[re]ax,1) - {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}, - // nopl 0(%[re]ax,%[re]ax,1) - // nopw 0(%[re]ax,%[re]ax,1) - {0x0f, 0x1f, 0x44, 0x00, 0x00, - 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00}, - // nopw 0(%[re]ax,%[re]ax,1) - // nopw 0(%[re]ax,%[re]ax,1) - {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00, - 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00}, - // nopw 0(%[re]ax,%[re]ax,1) - // nopl 0L(%[re]ax) */ - {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00, - 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00}, - // nopl 0L(%[re]ax) - // nopl 0L(%[re]ax) - {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00, - 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00}, - // nopl 0L(%[re]ax) - // nopl 0L(%[re]ax,%[re]ax,1) - {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00, - 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00} - }; - - if (Count > 15) - return 0; + ++stats::SectionLayouts; + + // Compute the section start address. + uint64_t StartAddress = 0; + if (SectionOrderIndex) { + MCSectionData *Prev = getSectionOrder()[SectionOrderIndex - 1]; + StartAddress = getSectionAddress(Prev) + getSectionAddressSize(Prev); + } - for (uint64_t i = 0; i < Count; i++) - OW->Write8(uint8_t(Nops[Count - 1][i])); + // Honor the section alignment requirements. + StartAddress = RoundUpToAlignment(StartAddress, SD->getAlignment()); - return Count; + // Set the section address. + SD->Address = StartAddress; } /// WriteFragmentData - Write the \arg F data to the output file. -static void WriteFragmentData(const MCFragment &F, MCObjectWriter *OW) { +static void WriteFragmentData(const MCAssembler &Asm, const MCAsmLayout &Layout, + const MCFragment &F, MCObjectWriter *OW) { uint64_t Start = OW->getStream().tell(); (void) Start; - ++EmittedFragments; + ++stats::EmittedFragments; // FIXME: Embed in fragments instead? + uint64_t FragmentSize = Layout.getFragmentEffectiveSize(&F); switch (F.getKind()) { case MCFragment::FT_Align: { MCAlignFragment &AF = cast(F); - uint64_t Count = AF.getFileSize() / AF.getValueSize(); + uint64_t Count = FragmentSize / AF.getValueSize(); + + assert(AF.getValueSize() && "Invalid virtual align in concrete fragment!"); // FIXME: This error shouldn't actually occur (the front end should emit // multiple .align directives to enforce the semantics it wants), but is // severe enough that we want to report it. How to handle this? - if (Count * AF.getValueSize() != AF.getFileSize()) - llvm_report_error("undefined .align directive, value size '" + + if (Count * AF.getValueSize() != FragmentSize) + report_fatal_error("undefined .align directive, value size '" + Twine(AF.getValueSize()) + "' is not a divisor of padding size '" + - Twine(AF.getFileSize()) + "'"); + Twine(FragmentSize) + "'"); // See if we are aligning with nops, and if so do that first to try to fill // the Count bytes. Then if that did not fill any bytes or there are any // bytes left to fill use the the Value and ValueSize to fill the rest. - if (AF.getEmitNops()) { - uint64_t NopByteCount = WriteNopData(Count, OW); - Count -= NopByteCount; + // If we are aligning with nops, ask that target to emit the right data. + if (AF.hasEmitNops()) { + if (!Asm.getBackend().WriteNopData(Count, OW)) + report_fatal_error("unable to write nop sequence of " + + Twine(Count) + " bytes"); + break; } + // Otherwise, write out in multiples of the value size. for (uint64_t i = 0; i != Count; ++i) { switch (AF.getValueSize()) { default: @@ -473,13 +574,18 @@ static void WriteFragmentData(const MCFragment &F, MCObjectWriter *OW) { } case MCFragment::FT_Data: { - OW->WriteBytes(cast(F).getContents().str()); + MCDataFragment &DF = cast(F); + assert(FragmentSize == DF.getContents().size() && "Invalid size!"); + OW->WriteBytes(DF.getContents().str()); break; } case MCFragment::FT_Fill: { MCFillFragment &FF = cast(F); - for (uint64_t i = 0, e = FF.getCount(); i != e; ++i) { + + assert(FF.getValueSize() && "Invalid virtual align in concrete fragment!"); + + for (uint64_t i = 0, e = FF.getSize() / FF.getValueSize(); i != e; ++i) { switch (FF.getValueSize()) { default: assert(0 && "Invalid size!"); @@ -492,29 +598,47 @@ static void WriteFragmentData(const MCFragment &F, MCObjectWriter *OW) { break; } + case MCFragment::FT_Inst: + llvm_unreachable("unexpected inst fragment after lowering"); + break; + case MCFragment::FT_Org: { MCOrgFragment &OF = cast(F); - for (uint64_t i = 0, e = OF.getFileSize(); i != e; ++i) + for (uint64_t i = 0, e = FragmentSize; i != e; ++i) OW->Write8(uint8_t(OF.getValue())); break; } - - case MCFragment::FT_ZeroFill: { - assert(0 && "Invalid zero fill fragment in concrete section!"); - break; - } } - assert(OW->getStream().tell() - Start == F.getFileSize()); + assert(OW->getStream().tell() - Start == FragmentSize); } void MCAssembler::WriteSectionData(const MCSectionData *SD, + const MCAsmLayout &Layout, MCObjectWriter *OW) const { // Ignore virtual sections. if (getBackend().isVirtualSection(SD->getSection())) { - assert(SD->getFileSize() == 0); + assert(Layout.getSectionFileSize(SD) == 0 && "Invalid size for section!"); + + // Check that contents are only things legal inside a virtual section. + for (MCSectionData::const_iterator it = SD->begin(), + ie = SD->end(); it != ie; ++it) { + switch (it->getKind()) { + default: + assert(0 && "Invalid fragment in virtual section!"); + case MCFragment::FT_Align: + assert(!cast(it)->getValueSize() && + "Invalid align in virtual section!"); + break; + case MCFragment::FT_Fill: + assert(!cast(it)->getValueSize() && + "Invalid fill in virtual section!"); + break; + } + } + return; } @@ -523,32 +647,124 @@ void MCAssembler::WriteSectionData(const MCSectionData *SD, for (MCSectionData::const_iterator it = SD->begin(), ie = SD->end(); it != ie; ++it) - WriteFragmentData(*it, OW); + WriteFragmentData(*this, Layout, *it, OW); + + assert(OW->getStream().tell() - Start == Layout.getSectionFileSize(SD)); +} + +void MCAssembler::AddSectionToTheEnd(MCSectionData &SD, MCAsmLayout &Layout) { + // Create dummy fragments and assign section ordinals. + unsigned SectionIndex = 0; + for (MCAssembler::iterator it = begin(), ie = end(); it != ie; ++it) + SectionIndex++; + + SD.setOrdinal(SectionIndex); + + // Assign layout order indices to sections and fragments. + unsigned FragmentIndex = 0; + unsigned i = 0; + for (unsigned e = Layout.getSectionOrder().size(); i != e; ++i) { + MCSectionData *SD = Layout.getSectionOrder()[i]; + + for (MCSectionData::iterator it2 = SD->begin(), + ie2 = SD->end(); it2 != ie2; ++it2) + FragmentIndex++; + } + + SD.setLayoutOrder(i); + for (MCSectionData::iterator it2 = SD.begin(), + ie2 = SD.end(); it2 != ie2; ++it2) { + it2->setLayoutOrder(FragmentIndex++); + } + Layout.getSectionOrder().push_back(&SD); - // Add section padding. - assert(SD->getFileSize() >= SD->getSize() && "Invalid section sizes!"); - OW->WriteZeros(SD->getFileSize() - SD->getSize()); + Layout.LayoutSection(&SD); + + // Layout until everything fits. + while (LayoutOnce(Layout)) + continue; - assert(OW->getStream().tell() - Start == SD->getFileSize()); } -void MCAssembler::Finish() { +void MCAssembler::Finish(MCObjectWriter *Writer) { DEBUG_WITH_TYPE("mc-dump", { llvm::errs() << "assembler backend - pre-layout\n--\n"; dump(); }); - // Layout until everything fits. + // Create the layout object. MCAsmLayout Layout(*this); + + // Insert additional align fragments for concrete sections to explicitly pad + // the previous section to match their alignment requirements. This is for + // 'gas' compatibility, it shouldn't strictly be necessary. + // + // FIXME: This may be Mach-O specific. + for (unsigned i = 1, e = Layout.getSectionOrder().size(); i < e; ++i) { + MCSectionData *SD = Layout.getSectionOrder()[i]; + + // Ignore sections without alignment requirements. + unsigned Align = SD->getAlignment(); + if (Align <= 1) + continue; + + // Ignore virtual sections, they don't cause file size modifications. + if (getBackend().isVirtualSection(SD->getSection())) + continue; + + // Otherwise, create a new align fragment at the end of the previous + // section. + MCAlignFragment *AF = new MCAlignFragment(Align, 0, 1, Align, + Layout.getSectionOrder()[i - 1]); + AF->setOnlyAlignAddress(true); + } + + // Create dummy fragments and assign section ordinals. + unsigned SectionIndex = 0; + for (MCAssembler::iterator it = begin(), ie = end(); it != ie; ++it) { + // Create dummy fragments to eliminate any empty sections, this simplifies + // layout. + if (it->getFragmentList().empty()) + new MCFillFragment(0, 1, 0, it); + + it->setOrdinal(SectionIndex++); + } + + // Assign layout order indices to sections and fragments. + unsigned FragmentIndex = 0; + for (unsigned i = 0, e = Layout.getSectionOrder().size(); i != e; ++i) { + MCSectionData *SD = Layout.getSectionOrder()[i]; + SD->setLayoutOrder(i); + + for (MCSectionData::iterator it2 = SD->begin(), + ie2 = SD->end(); it2 != ie2; ++it2) + it2->setLayoutOrder(FragmentIndex++); + } + + // Layout until everything fits. while (LayoutOnce(Layout)) continue; DEBUG_WITH_TYPE("mc-dump", { - llvm::errs() << "assembler backend - post-layout\n--\n"; + llvm::errs() << "assembler backend - post-relaxation\n--\n"; dump(); }); - llvm::OwningPtr Writer(getBackend().createObjectWriter(OS)); - if (!Writer) - llvm_report_error("unable to create object writer!"); + // Finalize the layout, including fragment lowering. + FinishLayout(Layout); + + DEBUG_WITH_TYPE("mc-dump", { + llvm::errs() << "assembler backend - final-layout\n--\n"; + dump(); }); + + uint64_t StartOffset = OS.tell(); + + llvm::OwningPtr OwnWriter(0); + if (Writer == 0) { + //no custom Writer_ : create the default one life-managed by OwningPtr + OwnWriter.reset(getBackend().createObjectWriter(OS)); + Writer = OwnWriter.get(); + if (!Writer) + report_fatal_error("unable to create object writer!"); + } // Allow the object writer a chance to perform post-layout binding (for // example, to set the index fields in the symbol data). @@ -564,7 +780,7 @@ void MCAssembler::Finish() { for (MCDataFragment::fixup_iterator it3 = DF->fixup_begin(), ie3 = DF->fixup_end(); it3 != ie3; ++it3) { - MCAsmFixup &Fixup = *it3; + MCFixup &Fixup = *it3; // Evaluate the fixup. MCValue Target; @@ -573,7 +789,7 @@ void MCAssembler::Finish() { // The fixup was unresolved, we need a relocation. Inform the object // writer of the relocation, and give it an opportunity to adjust the // fixup value if need be. - Writer->RecordRelocation(*this, *DF, Fixup, Target, FixedValue); + Writer->RecordRelocation(*this, Layout, DF, Fixup, Target,FixedValue); } getBackend().ApplyFixup(Fixup, *DF, FixedValue); @@ -582,15 +798,16 @@ void MCAssembler::Finish() { } // Write the object file. - Writer->WriteObject(*this); - OS.flush(); + Writer->WriteObject(*this, Layout); + + stats::ObjectBytes += OS.tell() - StartOffset; } -bool MCAssembler::FixupNeedsRelaxation(MCAsmFixup &Fixup, MCDataFragment *DF, +bool MCAssembler::FixupNeedsRelaxation(const MCFixup &Fixup, + const MCFragment *DF, const MCAsmLayout &Layout) const { - // Currently we only need to relax X86::reloc_pcrel_1byte. - if (unsigned(Fixup.Kind) != X86::reloc_pcrel_1byte) - return false; + if (getRelaxAll()) + return true; // If we cannot resolve the fixup value, it requires relaxation. MCValue Target; @@ -599,147 +816,132 @@ bool MCAssembler::FixupNeedsRelaxation(MCAsmFixup &Fixup, MCDataFragment *DF, return true; // Otherwise, relax if the value is too big for a (signed) i8. + // + // FIXME: This is target dependent! return int64_t(Value) != int64_t(int8_t(Value)); } -bool MCAssembler::LayoutOnce(MCAsmLayout &Layout) { - // Layout the concrete sections and fragments. - uint64_t Address = 0; - MCSectionData *Prev = 0; - for (iterator it = begin(), ie = end(); it != ie; ++it) { - MCSectionData &SD = *it; +bool MCAssembler::FragmentNeedsRelaxation(const MCInstFragment *IF, + const MCAsmLayout &Layout) const { + // If this inst doesn't ever need relaxation, ignore it. This occurs when we + // are intentionally pushing out inst fragments, or because we relaxed a + // previous instruction to one that doesn't need relaxation. + if (!getBackend().MayNeedRelaxation(IF->getInst())) + return false; - // Skip virtual sections. - if (getBackend().isVirtualSection(SD.getSection())) - continue; + for (MCInstFragment::const_fixup_iterator it = IF->fixup_begin(), + ie = IF->fixup_end(); it != ie; ++it) + if (FixupNeedsRelaxation(*it, IF, Layout)) + return true; - // Align this section if necessary by adding padding bytes to the previous - // section. - if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment())) { - assert(Prev && "Missing prev section!"); - Prev->setFileSize(Prev->getFileSize() + Pad); - Address += Pad; - } + return false; +} - // Layout the section fragments and its size. - SD.setAddress(Address); - LayoutSection(SD, Layout); - Address += SD.getFileSize(); +bool MCAssembler::LayoutOnce(MCAsmLayout &Layout) { + ++stats::RelaxationSteps; - Prev = &SD; - } + // Layout the sections in order. + Layout.LayoutFile(); - // Layout the virtual sections. + // Scan for fragments that need relaxation. + bool WasRelaxed = false; for (iterator it = begin(), ie = end(); it != ie; ++it) { MCSectionData &SD = *it; - if (!getBackend().isVirtualSection(SD.getSection())) - continue; - - // Align this section if necessary by adding padding bytes to the previous - // section. - if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment())) - Address += Pad; + for (MCSectionData::iterator it2 = SD.begin(), + ie2 = SD.end(); it2 != ie2; ++it2) { + // Check if this is an instruction fragment that needs relaxation. + MCInstFragment *IF = dyn_cast(it2); + if (!IF || !FragmentNeedsRelaxation(IF, Layout)) + continue; - SD.setAddress(Address); - LayoutSection(SD, Layout); - Address += SD.getSize(); + ++stats::RelaxedInstructions; + + // FIXME-PERF: We could immediately lower out instructions if we can tell + // they are fully resolved, to avoid retesting on later passes. + + // Relax the fragment. + + MCInst Relaxed; + getBackend().RelaxInstruction(IF->getInst(), Relaxed); + + // Encode the new instruction. + // + // FIXME-PERF: If it matters, we could let the target do this. It can + // probably do so more efficiently in many cases. + SmallVector Fixups; + SmallString<256> Code; + raw_svector_ostream VecOS(Code); + getEmitter().EncodeInstruction(Relaxed, VecOS, Fixups); + VecOS.flush(); + + // Update the instruction fragment. + int SlideAmount = Code.size() - IF->getInstSize(); + IF->setInst(Relaxed); + IF->getCode() = Code; + IF->getFixups().clear(); + // FIXME: Eliminate copy. + for (unsigned i = 0, e = Fixups.size(); i != e; ++i) + IF->getFixups().push_back(Fixups[i]); + + // Update the layout, and remember that we relaxed. + Layout.UpdateForSlide(IF, SlideAmount); + WasRelaxed = true; + } } - // Scan the fixups in order and relax any that don't fit. + return WasRelaxed; +} + +void MCAssembler::FinishLayout(MCAsmLayout &Layout) { + // Lower out any instruction fragments, to simplify the fixup application and + // output. + // + // FIXME-PERF: We don't have to do this, but the assumption is that it is + // cheap (we will mostly end up eliminating fragments and appending on to data + // fragments), so the extra complexity downstream isn't worth it. Evaluate + // this assumption. for (iterator it = begin(), ie = end(); it != ie; ++it) { MCSectionData &SD = *it; for (MCSectionData::iterator it2 = SD.begin(), ie2 = SD.end(); it2 != ie2; ++it2) { - MCDataFragment *DF = dyn_cast(it2); - if (!DF) + MCInstFragment *IF = dyn_cast(it2); + if (!IF) continue; - for (MCDataFragment::fixup_iterator it3 = DF->fixup_begin(), - ie3 = DF->fixup_end(); it3 != ie3; ++it3) { - MCAsmFixup &Fixup = *it3; - - // Check whether we need to relax this fixup. - if (!FixupNeedsRelaxation(Fixup, DF, Layout)) - continue; - - // Relax the instruction. - // - // FIXME: This is a huge temporary hack which just looks for x86 - // branches; the only thing we need to relax on x86 is - // 'X86::reloc_pcrel_1byte'. Once we have MCInst fragments, this will be - // replaced by a TargetAsmBackend hook (most likely tblgen'd) to relax - // an individual MCInst. - SmallVectorImpl &C = DF->getContents(); - uint64_t PrevOffset = Fixup.Offset; - unsigned Amt = 0; - - // jcc instructions - if (unsigned(C[Fixup.Offset-1]) >= 0x70 && - unsigned(C[Fixup.Offset-1]) <= 0x7f) { - C[Fixup.Offset] = C[Fixup.Offset-1] + 0x10; - C[Fixup.Offset-1] = char(0x0f); - ++Fixup.Offset; - Amt = 4; - - // jmp rel8 - } else if (C[Fixup.Offset-1] == char(0xeb)) { - C[Fixup.Offset-1] = char(0xe9); - Amt = 3; - - } else - llvm_unreachable("unknown 1 byte pcrel instruction!"); - - Fixup.Value = MCBinaryExpr::Create( - MCBinaryExpr::Sub, Fixup.Value, - MCConstantExpr::Create(3, getContext()), - getContext()); - C.insert(C.begin() + Fixup.Offset, Amt, char(0)); - Fixup.Kind = MCFixupKind(X86::reloc_pcrel_4byte); - - // Update the remaining fixups, which have slid. - // - // FIXME: This is bad for performance, but will be eliminated by the - // move to MCInst specific fragments. - ++it3; - for (; it3 != ie3; ++it3) - it3->Offset += Amt; - - // Update all the symbols for this fragment, which may have slid. - // - // FIXME: This is really really bad for performance, but will be - // eliminated by the move to MCInst specific fragments. - for (MCAssembler::symbol_iterator it = symbol_begin(), - ie = symbol_end(); it != ie; ++it) { - MCSymbolData &SD = *it; - - if (it->getFragment() != DF) - continue; - - if (SD.getOffset() > PrevOffset) - SD.setOffset(SD.getOffset() + Amt); - } - - // Restart layout. - // - // FIXME: This is O(N^2), but will be eliminated once we have a smart - // MCAsmLayout object. - return true; - } + // Create a new data fragment for the instruction. + // + // FIXME-PERF: Reuse previous data fragment if possible. + MCDataFragment *DF = new MCDataFragment(); + SD.getFragmentList().insert(it2, DF); + + // Update the data fragments layout data. + DF->setParent(IF->getParent()); + DF->setAtom(IF->getAtom()); + DF->setLayoutOrder(IF->getLayoutOrder()); + Layout.FragmentReplaced(IF, DF); + + // Copy in the data and the fixups. + DF->getContents().append(IF->getCode().begin(), IF->getCode().end()); + for (unsigned i = 0, e = IF->getFixups().size(); i != e; ++i) + DF->getFixups().push_back(IF->getFixups()[i]); + + // Delete the instruction fragment and update the iterator. + SD.getFragmentList().erase(IF); + it2 = DF; } } - - return false; } // Debugging methods namespace llvm { -raw_ostream &operator<<(raw_ostream &OS, const MCAsmFixup &AF) { - OS << ""; +raw_ostream &operator<<(raw_ostream &OS, const MCFixup &AF) { + OS << ""; return OS; } @@ -748,83 +950,82 @@ raw_ostream &operator<<(raw_ostream &OS, const MCAsmFixup &AF) { void MCFragment::dump() { raw_ostream &OS = llvm::errs(); - OS << ""; -} - -void MCAlignFragment::dump() { - raw_ostream &OS = llvm::errs(); - - OS << "MCFragment::dump(); - OS << "\n "; - OS << " Alignment:" << getAlignment() - << " Value:" << getValue() << " ValueSize:" << getValueSize() - << " MaxBytesToEmit:" << getMaxBytesToEmit() << ">"; -} + OS << "<"; + switch (getKind()) { + case MCFragment::FT_Align: OS << "MCAlignFragment"; break; + case MCFragment::FT_Data: OS << "MCDataFragment"; break; + case MCFragment::FT_Fill: OS << "MCFillFragment"; break; + case MCFragment::FT_Inst: OS << "MCInstFragment"; break; + case MCFragment::FT_Org: OS << "MCOrgFragment"; break; + } -void MCDataFragment::dump() { - raw_ostream &OS = llvm::errs(); + OS << ""; - OS << "MCFragment::dump(); - OS << "\n "; - OS << " Contents:["; - for (unsigned i = 0, e = getContents().size(); i != e; ++i) { - if (i) OS << ","; - OS << hexdigit((Contents[i] >> 4) & 0xF) << hexdigit(Contents[i] & 0xF); - } - OS << "] (" << getContents().size() << " bytes)"; - - if (!getFixups().empty()) { - OS << ",\n "; - OS << " Fixups:["; - for (fixup_iterator it = fixup_begin(), ie = fixup_end(); it != ie; ++it) { - if (it != fixup_begin()) OS << ",\n "; - OS << *it; + switch (getKind()) { + case MCFragment::FT_Align: { + const MCAlignFragment *AF = cast(this); + if (AF->hasEmitNops()) + OS << " (emit nops)"; + if (AF->hasOnlyAlignAddress()) + OS << " (only align section)"; + OS << "\n "; + OS << " Alignment:" << AF->getAlignment() + << " Value:" << AF->getValue() << " ValueSize:" << AF->getValueSize() + << " MaxBytesToEmit:" << AF->getMaxBytesToEmit() << ">"; + break; + } + case MCFragment::FT_Data: { + const MCDataFragment *DF = cast(this); + OS << "\n "; + OS << " Contents:["; + const SmallVectorImpl &Contents = DF->getContents(); + for (unsigned i = 0, e = Contents.size(); i != e; ++i) { + if (i) OS << ","; + OS << hexdigit((Contents[i] >> 4) & 0xF) << hexdigit(Contents[i] & 0xF); } - OS << "]"; + OS << "] (" << Contents.size() << " bytes)"; + + if (!DF->getFixups().empty()) { + OS << ",\n "; + OS << " Fixups:["; + for (MCDataFragment::const_fixup_iterator it = DF->fixup_begin(), + ie = DF->fixup_end(); it != ie; ++it) { + if (it != DF->fixup_begin()) OS << ",\n "; + OS << *it; + } + OS << "]"; + } + break; + } + case MCFragment::FT_Fill: { + const MCFillFragment *FF = cast(this); + OS << " Value:" << FF->getValue() << " ValueSize:" << FF->getValueSize() + << " Size:" << FF->getSize(); + break; + } + case MCFragment::FT_Inst: { + const MCInstFragment *IF = cast(this); + OS << "\n "; + OS << " Inst:"; + IF->getInst().dump_pretty(OS); + break; + } + case MCFragment::FT_Org: { + const MCOrgFragment *OF = cast(this); + OS << "\n "; + OS << " Offset:" << OF->getOffset() << " Value:" << OF->getValue(); + break; + } } - OS << ">"; } -void MCFillFragment::dump() { - raw_ostream &OS = llvm::errs(); - - OS << "MCFragment::dump(); - OS << "\n "; - OS << " Value:" << getValue() << " ValueSize:" << getValueSize() - << " Count:" << getCount() << ">"; -} - -void MCOrgFragment::dump() { - raw_ostream &OS = llvm::errs(); - - OS << "MCFragment::dump(); - OS << "\n "; - OS << " Offset:" << getOffset() << " Value:" << getValue() << ">"; -} - -void MCZeroFillFragment::dump() { - raw_ostream &OS = llvm::errs(); - - OS << "MCFragment::dump(); - OS << "\n "; - OS << " Size:" << getSize() << " Alignment:" << getAlignment() << ">"; -} - void MCSectionData::dump() { raw_ostream &OS = llvm::errs(); OS << "