MCFragment: Use trailing data for fixed-size part

Reapply after llvm#151724 switched to `char *Data`, fixing a
-fsanitize=pointer-overflow issue in MCAssembler::layout.

---

The fixed-size content of the MCFragment object is now stored as
trailing data, replacing ContentStart/ContentEnd with ContentSize. The
available space for trailing data is tracked using `FragSpace`. If the
available space is insufficient, a new block is allocated within the
bump allocator `MCObjectStreamer::FragStorage`.

FragList::Tail cannot be reused when switching sections or subsections,
as it is not associated with the fragment space tracked by `FragSpace`.
Instead, allocate a new fragment, which becomes less expensive after llvm#150574.

Data can only be appended to the tail fragment of a subsection, not to
fragments in the middle. Post-assembler-layout adjustments (such as
.llvm_addrsig and .llvm.call-graph-profile) have been updated to use the
variable-size part instead.

Pull Request: llvm#150846

Author: MaskRay

llvm/include/llvm/MC/MCObjectStreamer.h

@@ -52,6 +52,10 @@ class MCObjectStreamer : public MCStreamer {
   DenseMap<const MCSymbol *, SmallVector<PendingAssignment, 1>>
       pendingAssignments;
 
+  SmallVector<std::unique_ptr<uint8_t[]>, 0> FragStorage;
+  // Available bytes in the current block for trailing data or new fragments.
+  size_t FragSpace = 0;
+
   void emitInstToData(const MCInst &Inst, const MCSubtargetInfo &);
   void emitCFIStartProcImpl(MCDwarfFrameInfo &Frame) override;
   void emitCFIEndProcImpl(MCDwarfFrameInfo &Frame) override;
@@ -84,11 +88,18 @@ class MCObjectStreamer : public MCStreamer {
   // Add a fragment with a variable-size tail and start a new empty fragment.
   void insert(MCFragment *F);
 
+  uint8_t *getCurFragEnd() const {
+    return reinterpret_cast<uint8_t *>(CurFrag + 1) + CurFrag->getFixedSize();
+  }
+  MCFragment *allocFragSpace(size_t Headroom);
   // Add a new fragment to the current section without a variable-size tail.
   void newFragment();
 
+  void ensureHeadroom(size_t Headroom);
   void appendContents(ArrayRef<char> Contents);
-  void appendContents(size_t Num, char Elt);
+  void appendContents(size_t Num, uint8_t Elt);
+  // Add a fixup to the current fragment. Call ensureHeadroom beforehand to
+  // ensure the fixup and appended content apply to the same fragment.
   void addFixup(const MCExpr *Value, MCFixupKind Kind);
 
   void emitLabel(MCSymbol *Symbol, SMLoc Loc = SMLoc()) override;

llvm/include/llvm/MC/MCSection.h

@@ -91,10 +91,10 @@ class MCFragment {
   bool AllowAutoPadding : 1;
 
   // Track content and fixups for the fixed-size part as fragments are
-  // appended to the section. The content remains immutable, except when
-  // modified by applyFixup.
-  uint32_t ContentStart = 0;
-  uint32_t ContentEnd = 0;
+  // appended to the section. The content is stored as trailing data of the
+  // MCFragment. The content remains immutable, except when modified by
+  // applyFixup.
+  uint32_t FixedSize = 0;
   uint32_t FixupStart = 0;
   uint32_t FixupEnd = 0;
 
@@ -188,18 +188,6 @@ class MCFragment {
   //== Content-related functions manage parent's storage using ContentStart and
   // ContentSize.
 
-  // Get a SmallVector reference. The caller should call doneAppending to update
-  // `ContentEnd`.
-  SmallVectorImpl<char> &getContentsForAppending();
-  void doneAppending();
-  void appendContents(ArrayRef<char> Contents) {
-    getContentsForAppending().append(Contents.begin(), Contents.end());
-    doneAppending();
-  }
-  void appendContents(size_t Num, char Elt) {
-    getContentsForAppending().append(Num, Elt);
-    doneAppending();
-  }
   MutableArrayRef<char> getContents();
   ArrayRef<char> getContents() const;
 
@@ -208,10 +196,10 @@ class MCFragment {
   MutableArrayRef<char> getVarContents();
   ArrayRef<char> getVarContents() const;
 
-  size_t getFixedSize() const { return ContentEnd - ContentStart; }
+  size_t getFixedSize() const { return FixedSize; }
   size_t getVarSize() const { return VarContentEnd - VarContentStart; }
   size_t getSize() const {
-    return ContentEnd - ContentStart + (VarContentEnd - VarContentStart);
+    return FixedSize + (VarContentEnd - VarContentStart);
   }
 
   //== Fixup-related functions manage parent's storage using FixupStart and
@@ -626,28 +614,11 @@ class LLVM_ABI MCSection {
   bool isBssSection() const { return IsBss; }
 };
 
-inline SmallVectorImpl<char> &MCFragment::getContentsForAppending() {
-  SmallVectorImpl<char> &S = getParent()->ContentStorage;
-  if (LLVM_UNLIKELY(ContentEnd != S.size())) {
-    // Move the elements to the end. Reserve space to avoid invalidating
-    // S.begin()+I for `append`.
-    auto Size = ContentEnd - ContentStart;
-    auto I = std::exchange(ContentStart, S.size());
-    S.reserve(S.size() + Size);
-    S.append(S.begin() + I, S.begin() + I + Size);
-  }
-  return S;
-}
-inline void MCFragment::doneAppending() {
-  ContentEnd = getParent()->ContentStorage.size();
-}
 inline MutableArrayRef<char> MCFragment::getContents() {
-  return MutableArrayRef(getParent()->ContentStorage)
-      .slice(ContentStart, ContentEnd - ContentStart);
+  return {reinterpret_cast<char *>(this + 1), FixedSize};
 }
 inline ArrayRef<char> MCFragment::getContents() const {
-  return ArrayRef(getParent()->ContentStorage)
-      .slice(ContentStart, ContentEnd - ContentStart);
+  return {reinterpret_cast<const char *>(this + 1), FixedSize};
 }
 
 inline MutableArrayRef<char> MCFragment::getVarContents() {

llvm/lib/MC/MCObjectStreamer.cpp

@@ -46,27 +46,83 @@ MCAssembler *MCObjectStreamer::getAssemblerPtr() {
   return nullptr;
 }
 
+constexpr size_t FragBlockSize = 16384;
+// Ensure the new fragment can at least store a few bytes.
+constexpr size_t NewFragHeadroom = 8;
+
+static_assert(NewFragHeadroom >= alignof(MCFragment));
+static_assert(FragBlockSize >= sizeof(MCFragment) + NewFragHeadroom);
+
+MCFragment *MCObjectStreamer::allocFragSpace(size_t Headroom) {
+  auto Size = std::max(FragBlockSize, sizeof(MCFragment) + Headroom);
+  FragSpace = Size - sizeof(MCFragment);
+  auto Block = std::unique_ptr<uint8_t[]>(new uint8_t[Size]);
+  auto *F = reinterpret_cast<MCFragment *>(Block.get());
+  FragStorage.push_back(std::move(Block));
+  return F;
+}
+
 void MCObjectStreamer::newFragment() {
-  addFragment(getContext().allocFragment<MCFragment>());
+  MCFragment *F;
+  if (LLVM_LIKELY(sizeof(MCFragment) + NewFragHeadroom <= FragSpace)) {
+    auto End = reinterpret_cast<size_t>(getCurFragEnd());
+    F = reinterpret_cast<MCFragment *>(
+        alignToPowerOf2(End, alignof(MCFragment)));
+    FragSpace -= size_t(F) - End + sizeof(MCFragment);
+  } else {
+    F = allocFragSpace(0);
+  }
+  new (F) MCFragment();
+  addFragment(F);
+}
+
+void MCObjectStreamer::ensureHeadroom(size_t Headroom) {
+  if (Headroom <= FragSpace)
+    return;
+  auto *F = allocFragSpace(Headroom);
+  new (F) MCFragment();
+  addFragment(F);
 }
 
-void MCObjectStreamer::insert(MCFragment *F) {
-  assert(F->getKind() != MCFragment::FT_Data &&
+void MCObjectStreamer::insert(MCFragment *Frag) {
+  assert(Frag->getKind() != MCFragment::FT_Data &&
          "F should have a variable-size tail");
+  // Frag is not connected to FragSpace. Before modifying CurFrag with
+  // addFragment(Frag), allocate an empty fragment to maintain FragSpace
+  // connectivity, potentially reusing CurFrag's associated space.
+  MCFragment *F;
+  if (LLVM_LIKELY(sizeof(MCFragment) + NewFragHeadroom <= FragSpace)) {
+    auto End = reinterpret_cast<size_t>(getCurFragEnd());
+    F = reinterpret_cast<MCFragment *>(
+        alignToPowerOf2(End, alignof(MCFragment)));
+    FragSpace -= size_t(F) - End + sizeof(MCFragment);
+  } else {
+    F = allocFragSpace(0);
+  }
+  new (F) MCFragment();
+
+  addFragment(Frag);
   addFragment(F);
-  newFragment();
 }
 
 void MCObjectStreamer::appendContents(ArrayRef<char> Contents) {
-  CurFrag->appendContents(Contents);
+  ensureHeadroom(Contents.size());
+  assert(FragSpace >= Contents.size());
+  llvm::copy(Contents, getCurFragEnd());
+  CurFrag->FixedSize += Contents.size();
+  FragSpace -= Contents.size();
 }
 
-void MCObjectStreamer::appendContents(size_t Num, char Elt) {
-  CurFrag->appendContents(Num, Elt);
+void MCObjectStreamer::appendContents(size_t Num, uint8_t Elt) {
+  ensureHeadroom(Num);
+  MutableArrayRef<uint8_t> Data(getCurFragEnd(), Num);
+  llvm::fill(Data, Elt);
+  CurFrag->FixedSize += Num;
+  FragSpace -= Num;
 }
 
 void MCObjectStreamer::addFixup(const MCExpr *Value, MCFixupKind Kind) {
-  CurFrag->addFixup(MCFixup::create(CurFrag->getFixedSize(), Value, Kind));
+  CurFrag->addFixup(MCFixup::create(getCurFragSize(), Value, Kind));
 }
 
 // As a compile-time optimization, avoid allocating and evaluating an MCExpr
@@ -115,6 +171,8 @@ void MCObjectStreamer::reset() {
   }
   EmitEHFrame = true;
   EmitDebugFrame = false;
+  FragStorage.clear();
+  FragSpace = 0;
   MCStreamer::reset();
 }
 
@@ -143,7 +201,6 @@ void MCObjectStreamer::emitCFISections(bool EH, bool Debug, bool SFrame) {
 void MCObjectStreamer::emitValueImpl(const MCExpr *Value, unsigned Size,
                                      SMLoc Loc) {
   MCStreamer::emitValueImpl(Value, Size, Loc);
-  MCFragment *DF = getCurrentFragment();
 
   MCDwarfLineEntry::make(this, getCurrentSectionOnly());
 
@@ -158,9 +215,9 @@ void MCObjectStreamer::emitValueImpl(const MCExpr *Value, unsigned Size,
     emitIntValue(AbsValue, Size);
     return;
   }
-  DF->addFixup(MCFixup::create(DF->getContents().size(), Value,
-                               MCFixup::getDataKindForSize(Size)));
-  DF->appendContents(Size, 0);
+  ensureHeadroom(Size);
+  addFixup(Value, MCFixup::getDataKindForSize(Size));
+  appendContents(Size, 0);
 }
 
 MCSymbol *MCObjectStreamer::emitCFILabel() {
@@ -194,7 +251,7 @@ void MCObjectStreamer::emitLabel(MCSymbol *Symbol, SMLoc Loc) {
   // section.
   MCFragment *F = CurFrag;
   Symbol->setFragment(F);
-  Symbol->setOffset(F->getContents().size());
+  Symbol->setOffset(F->getFixedSize());
 
   emitPendingAssignments(Symbol);
 }
@@ -260,20 +317,38 @@ void MCObjectStreamer::changeSection(MCSection *Section, uint32_t Subsection) {
     F0 = CurFrag;
   }
 
+  // To maintain connectivity between CurFrag and FragSpace when CurFrag is
+  // modified, allocate an empty fragment and append it to the fragment list.
+  // (Subsections[I].second.Tail is not connected to FragSpace.)
+  MCFragment *F;
+  if (LLVM_LIKELY(sizeof(MCFragment) + NewFragHeadroom <= FragSpace)) {
+    auto End = reinterpret_cast<size_t>(getCurFragEnd());
+    F = reinterpret_cast<MCFragment *>(
+        alignToPowerOf2(End, alignof(MCFragment)));
+    FragSpace -= size_t(F) - End + sizeof(MCFragment);
+  } else {
+    F = allocFragSpace(0);
+  }
+  new (F) MCFragment();
+  F->setParent(Section);
+
   auto &Subsections = Section->Subsections;
   size_t I = 0, E = Subsections.size();
   while (I != E && Subsections[I].first < Subsection)
     ++I;
   // If the subsection number is not in the sorted Subsections list, create a
   // new fragment list.
   if (I == E || Subsections[I].first != Subsection) {
-    auto *F = getContext().allocFragment<MCFragment>();
-    F->setParent(Section);
     Subsections.insert(Subsections.begin() + I,
                        {Subsection, MCSection::FragList{F, F}});
+    Section->CurFragList = &Subsections[I].second;
+    CurFrag = F;
+  } else {
+    Section->CurFragList = &Subsections[I].second;
+    CurFrag = Subsections[I].second.Tail;
+    // Ensure CurFrag is associated with FragSpace.
+    addFragment(F);
   }
-  Section->CurFragList = &Subsections[I].second;
-  CurFrag = Section->CurFragList->Tail;
 
   // Define the section symbol at subsection 0's initial fragment if required.
   if (!NewSec)
@@ -344,11 +419,15 @@ void MCObjectStreamer::emitInstToData(const MCInst &Inst,
   MCFragment *F = getCurrentFragment();
 
   // Append the instruction to the data fragment.
-  size_t CodeOffset = F->getContents().size();
+  size_t CodeOffset = getCurFragSize();
+  SmallString<16> Content;
   SmallVector<MCFixup, 1> Fixups;
-  getAssembler().getEmitter().encodeInstruction(
-      Inst, F->getContentsForAppending(), Fixups, STI);
-  F->doneAppending();
+  getAssembler().getEmitter().encodeInstruction(Inst, Content, Fixups, STI);
+  appendContents(Content);
+  if (CurFrag != F) {
+    F = CurFrag;
+    CodeOffset = 0;
+  }
   F->setHasInstructions(STI);
 
   if (Fixups.empty())

llvm/lib/MC/MCWin64EH.cpp

@@ -318,6 +318,9 @@ static void EmitUnwindInfo(MCStreamer &streamer, WinEH::FrameInfo *info) {
 
   // Emit the epilog instructions.
   if (EnableUnwindV2) {
+    // Ensure the fixups and appended content apply to the same fragment.
+    OS->ensureHeadroom(info->EpilogMap.size() * 2);
+
     bool IsLast = true;
     for (const auto &Epilog : llvm::reverse(info->EpilogMap)) {
       if (IsLast) {

llvm/lib/MC/MCWinCOFFStreamer.cpp

@@ -280,6 +280,7 @@ void MCWinCOFFStreamer::emitCOFFSymbolIndex(MCSymbol const *Symbol) {
 void MCWinCOFFStreamer::emitCOFFSectionIndex(const MCSymbol *Symbol) {
   visitUsedSymbol(*Symbol);
   const MCSymbolRefExpr *SRE = MCSymbolRefExpr::create(Symbol, getContext());
+  ensureHeadroom(2);
   addFixup(SRE, FK_SecRel_2);
   appendContents(2, 0);
 }
@@ -293,6 +294,7 @@ void MCWinCOFFStreamer::emitCOFFSecRel32(const MCSymbol *Symbol,
   if (Offset)
     MCE = MCBinaryExpr::createAdd(
         MCE, MCConstantExpr::create(Offset, getContext()), getContext());
+  ensureHeadroom(4);
   addFixup(MCE, FK_SecRel_4);
   // Emit 4 bytes (zeros) to the object file.
   appendContents(4, 0);
@@ -308,6 +310,7 @@ void MCWinCOFFStreamer::emitCOFFImgRel32(const MCSymbol *Symbol,
   if (Offset)
     MCE = MCBinaryExpr::createAdd(
         MCE, MCConstantExpr::create(Offset, getContext()), getContext());
+  ensureHeadroom(4);
   addFixup(MCE, FK_Data_4);
   // Emit 4 bytes (zeros) to the object file.
   appendContents(4, 0);
@@ -318,6 +321,7 @@ void MCWinCOFFStreamer::emitCOFFSecNumber(MCSymbol const *Symbol) {
   // Create Symbol for section number.
   const MCExpr *MCE = MCCOFFSectionNumberTargetExpr::create(
       *Symbol, this->getWriter(), getContext());
+  ensureHeadroom(4);
   addFixup(MCE, FK_Data_4);
   // Emit 4 bytes (zeros) to the object file.
   appendContents(4, 0);
@@ -328,6 +332,7 @@ void MCWinCOFFStreamer::emitCOFFSecOffset(MCSymbol const *Symbol) {
   // Create Symbol for section offset.
   const MCExpr *MCE =
       MCCOFFSectionOffsetTargetExpr::create(*Symbol, getContext());
+  ensureHeadroom(4);
   addFixup(MCE, FK_Data_4);
   // Emit 4 bytes (zeros) to the object file.
   appendContents(4, 0);

llvm/lib/Target/Mips/MCTargetDesc/MipsTargetStreamer.cpp

@@ -1034,37 +1034,43 @@ MCELFStreamer &MipsTargetELFStreamer::getStreamer() {
 
 void MipsTargetELFStreamer::emitGPRel32Value(const MCExpr *Value) {
   auto &S = getStreamer();
+  S.ensureHeadroom(4);
   S.addFixup(Value, Mips::fixup_Mips_GPREL32);
   S.appendContents(4, 0);
 }
 
 void MipsTargetELFStreamer::emitGPRel64Value(const MCExpr *Value) {
   auto &S = getStreamer();
+  S.ensureHeadroom(8);
   // fixup_Mips_GPREL32 desginates R_MIPS_GPREL32+R_MIPS_64 on MIPS64.
   S.addFixup(Value, Mips::fixup_Mips_GPREL32);
   S.appendContents(8, 0);
 }
 
 void MipsTargetELFStreamer::emitDTPRel32Value(const MCExpr *Value) {
   auto &S = getStreamer();
+  S.ensureHeadroom(4);
   S.addFixup(Value, Mips::fixup_Mips_DTPREL32);
   S.appendContents(4, 0);
 }
 
 void MipsTargetELFStreamer::emitDTPRel64Value(const MCExpr *Value) {
   auto &S = getStreamer();
+  S.ensureHeadroom(8);
   S.addFixup(Value, Mips::fixup_Mips_DTPREL64);
   S.appendContents(8, 0);
 }
 
 void MipsTargetELFStreamer::emitTPRel32Value(const MCExpr *Value) {
   auto &S = getStreamer();
+  S.ensureHeadroom(4);
   S.addFixup(Value, Mips::fixup_Mips_TPREL32);
   S.appendContents(4, 0);
 }
 
 void MipsTargetELFStreamer::emitTPRel64Value(const MCExpr *Value) {
   auto &S = getStreamer();
+  S.ensureHeadroom(8);
   S.addFixup(Value, Mips::fixup_Mips_TPREL64);
   S.appendContents(8, 0);
 }