MC: Allocate initial fragment and define section symbol in changeSection

13a79bb (2017) introduced fragment
creation in MCContext for createELFSectionImpl, which was inappropriate.
Fragments should only be created when using MCSteramer, not during
`MCContext::get*Section` calls.
`initMachOMCObjectFileInfo` defines multiple sections, some of which may
not be used by the code generator. This caused symbol names matching
these sections to be incorrectly marked as undefined (see
https://reviews.llvm.org/D55173).

The fragment code was later replicated in other file formats, such as
WebAssembly (see https://reviews.llvm.org/D46561), XCOFF, and GOFF.

This patch fixes the problem by moving initial fragment allocation from
MCContext::createSection to MCStreamer::changeSection.
While MCContext still creates a section symbol, the symbol is not
attached to the initial fragment.
In addition, move `emitLabel`/`setFragment` from `switchSection*` and
overridden changeSection to `MCObjectStreamer::changeSection` for
consistency.

* test/CodeGen/XCore/section-name.ll now passes. XCore doesn't support
  MCObjectStreamer. I don't think the MCAsmStreamer output behavior
  change matters.

Pull Request: llvm#150574

Author: MaskRay

llvm/include/llvm/MC/MCContext.h

@@ -333,8 +333,6 @@ class MCContext {
   void reportCommon(SMLoc Loc,
                     std::function<void(SMDiagnostic &, const SourceMgr *)>);
 
-  MCFragment *allocInitialFragment(MCSection &Sec);
-
   MCSymbolTableEntry &getSymbolTableEntry(StringRef Name);
 
   MCSymbol *createSymbolImpl(const MCSymbolTableEntry *Name, bool IsTemporary);

llvm/include/llvm/MC/MCSection.h

@@ -589,6 +589,8 @@ class LLVM_ABI MCSection {
   /// offset between two locations may not be fully resolved.
   bool LinkerRelaxable : 1;
 
+  MCFragment DummyFragment;
+
   // Mapping from subsection number to fragment list. At layout time, the
   // subsection 0 list is replaced with concatenated fragments from all
   // subsections.
@@ -650,7 +652,7 @@ class LLVM_ABI MCSection {
   bool isLinkerRelaxable() const { return LinkerRelaxable; }
   void setLinkerRelaxable() { LinkerRelaxable = true; }
 
-  MCFragment &getDummyFragment() { return *Subsections[0].second.Head; }
+  MCFragment &getDummyFragment() { return DummyFragment; }
 
   FragList *curFragList() const { return CurFragList; }
   iterator begin() const { return iterator(CurFragList->Head); }

llvm/include/llvm/MC/MCXCOFFStreamer.h

@@ -22,6 +22,7 @@ class MCXCOFFStreamer : public MCObjectStreamer {
 
   XCOFFObjectWriter &getWriter();
 
+  void changeSection(MCSection *Section, uint32_t Subsection = 0) override;
   bool emitSymbolAttribute(MCSymbol *Symbol, MCSymbolAttr Attribute) override;
   void emitCommonSymbol(MCSymbol *Symbol, uint64_t Size,
                         Align ByteAlignment) override;

llvm/lib/MC/ELFObjectWriter.cpp

@@ -559,20 +559,7 @@ void ELFWriter::computeSymbolTable(const RevGroupMapTy &RevGroupMap) {
     } else {
       const MCSectionELF &Section =
           static_cast<const MCSectionELF &>(Symbol.getSection());
-
-      // We may end up with a situation when section symbol is technically
-      // defined, but should not be. That happens because we explicitly
-      // pre-create few .debug_* sections to have accessors.
-      // And if these sections were not really defined in the code, but were
-      // referenced, we simply error out.
-      if (!Section.isRegistered()) {
-        assert(static_cast<const MCSymbolELF &>(Symbol).getType() ==
-               ELF::STT_SECTION);
-        Ctx.reportError(SMLoc(),
-                        "Undefined section reference: " + Symbol.getName());
-        continue;
-      }
-
+      assert(Section.isRegistered());
       if (Mode == NonDwoOnly && isDwoSection(Section))
         continue;
       MSD.SectionIndex = Section.getOrdinal();

llvm/lib/MC/MCAssembler.cpp

@@ -106,7 +106,6 @@ void MCAssembler::reset() {
 bool MCAssembler::registerSection(MCSection &Section) {
   if (Section.isRegistered())
     return false;
-  assert(Section.curFragList()->Head && "allocInitialFragment not called");
   Sections.push_back(&Section);
   Section.setIsRegistered(true);
   return true;

llvm/lib/MC/MCContext.cpp

@@ -200,16 +200,6 @@ MCInst *MCContext::createMCInst() {
   return new (MCInstAllocator.Allocate()) MCInst;
 }
 
-// Allocate the initial MCFragment for the begin symbol.
-MCFragment *MCContext::allocInitialFragment(MCSection &Sec) {
-  assert(!Sec.curFragList()->Head);
-  auto *F = allocFragment<MCFragment>();
-  F->setParent(&Sec);
-  Sec.curFragList()->Head = F;
-  Sec.curFragList()->Tail = F;
-  return F;
-}
-
 //===----------------------------------------------------------------------===//
 // Symbol Manipulation
 //===----------------------------------------------------------------------===//
@@ -443,24 +433,28 @@ MCSymbol *MCContext::getDirectionalLocalSymbol(unsigned LocalLabelVal,
   return getOrCreateDirectionalLocalSymbol(LocalLabelVal, Instance);
 }
 
+// Create a section symbol, with a distinct one for each section of the same.
+// The first symbol is used for assembly code references.
 template <typename Symbol>
 Symbol *MCContext::getOrCreateSectionSymbol(StringRef Section) {
   Symbol *R;
   auto &SymEntry = getSymbolTableEntry(Section);
   MCSymbol *Sym = SymEntry.second.Symbol;
-  // A section symbol can not redefine regular symbols. There may be multiple
-  // sections with the same name, in which case the first such section wins.
   if (Sym && Sym->isDefined() &&
       (!Sym->isInSection() || Sym->getSection().getBeginSymbol() != Sym))
     reportError(SMLoc(), "invalid symbol redefinition");
-  if (Sym && Sym->isUndefined()) {
+  // Use the symbol's index to track if it has been used as a section symbol.
+  // Set to -1 to catch potential bugs if misused as a symbol index.
+  if (Sym && Sym->getIndex() != -1u) {
     R = cast<Symbol>(Sym);
   } else {
     SymEntry.second.Used = true;
     R = new (&SymEntry, *this) Symbol(&SymEntry, /*isTemporary=*/false);
     if (!Sym)
       SymEntry.second.Symbol = R;
   }
+  // Mark as section symbol.
+  R->setIndex(-1u);
   return R;
 }
 
@@ -568,7 +562,6 @@ MCSectionMachO *MCContext::getMachOSection(StringRef Segment, StringRef Section,
       MCSectionMachO(Segment, Name.substr(Name.size() - Section.size()),
                      TypeAndAttributes, Reserved2, Kind, Begin);
   R.first->second = Ret;
-  allocInitialFragment(*Ret);
   return Ret;
 }
 
@@ -579,15 +572,8 @@ MCSectionELF *MCContext::createELFSectionImpl(StringRef Section, unsigned Type,
                                               bool Comdat, unsigned UniqueID,
                                               const MCSymbolELF *LinkedToSym) {
   auto *R = getOrCreateSectionSymbol<MCSymbolELF>(Section);
-  R->setBinding(ELF::STB_LOCAL);
-  R->setType(ELF::STT_SECTION);
-
-  auto *Ret = new (ELFAllocator.Allocate()) MCSectionELF(
+  return new (ELFAllocator.Allocate()) MCSectionELF(
       Section, Type, Flags, EntrySize, Group, Comdat, UniqueID, R, LinkedToSym);
-
-  auto *F = allocInitialFragment(*Ret);
-  R->setFragment(F);
-  return Ret;
 }
 
 MCSectionELF *
@@ -743,7 +729,6 @@ MCSectionGOFF *MCContext::getGOFFSection(SectionKind Kind, StringRef Name,
       MCSectionGOFF(CachedName, Kind, IsVirtual, Attributes,
                     static_cast<MCSectionGOFF *>(Parent));
   Iter->second = GOFFSection;
-  allocInitialFragment(*GOFFSection);
   return GOFFSection;
 }
 
@@ -798,8 +783,7 @@ MCSectionCOFF *MCContext::getCOFFSection(StringRef Section,
   MCSectionCOFF *Result = new (COFFAllocator.Allocate()) MCSectionCOFF(
       CachedName, Characteristics, COMDATSymbol, Selection, UniqueID, Begin);
   Iter->second = Result;
-  auto *F = allocInitialFragment(*Result);
-  Begin->setFragment(F);
+  Begin->setFragment(&Result->getDummyFragment());
   return Result;
 }
 
@@ -870,8 +854,6 @@ MCSectionWasm *MCContext::getWasmSection(const Twine &Section, SectionKind Kind,
       MCSectionWasm(CachedName, Kind, Flags, GroupSym, UniqueID, Begin);
   Entry.second = Result;
 
-  auto *F = allocInitialFragment(*Result);
-  Begin->setFragment(F);
   return Result;
 }
 
@@ -927,24 +909,11 @@ MCSectionXCOFF *MCContext::getXCOFFSection(
                        MultiSymbolsAllowed);
 
   Entry.second = Result;
-
-  auto *F = allocInitialFragment(*Result);
-
-  // We might miss calculating the symbols difference as absolute value before
-  // adding fixups when symbol_A without the fragment set is the csect itself
-  // and symbol_B is in it.
-  // TODO: Currently we only set the fragment for XMC_PR csects and DWARF
-  // sections because we don't have other cases that hit this problem yet.
-  if (IsDwarfSec || CsectProp->MappingClass == XCOFF::XMC_PR)
-    QualName->setFragment(F);
-
   return Result;
 }
 
 MCSectionSPIRV *MCContext::getSPIRVSection() {
   MCSectionSPIRV *Result = new (SPIRVAllocator.Allocate()) MCSectionSPIRV();
-
-  allocInitialFragment(*Result);
   return Result;
 }
 
@@ -964,7 +933,6 @@ MCSectionDXContainer *MCContext::getDXContainerSection(StringRef Section,
       new (DXCAllocator.Allocate()) MCSectionDXContainer(Name, K, nullptr);
 
   // The first fragment will store the header
-  allocInitialFragment(*MapIt->second);
   return MapIt->second;
 }
 

llvm/lib/MC/MCELFStreamer.cpp

@@ -89,7 +89,9 @@ void MCELFStreamer::changeSection(MCSection *Section, uint32_t Subsection) {
     getWriter().markGnuAbi();
 
   MCObjectStreamer::changeSection(Section, Subsection);
-  Asm.registerSymbol(*Section->getBeginSymbol());
+  auto *Sym = static_cast<MCSymbolELF *>(Section->getBeginSymbol());
+  Sym->setBinding(ELF::STB_LOCAL);
+  Sym->setType(ELF::STT_SECTION);
 }
 
 void MCELFStreamer::emitWeakReference(MCSymbol *Alias, const MCSymbol *Target) {

llvm/lib/MC/MCMachOStreamer.cpp

@@ -140,6 +140,8 @@ void MCMachOStreamer::changeSection(MCSection *Section, uint32_t Subsection) {
     MCSymbol *Label = getContext().createLinkerPrivateTempSymbol();
     Section->setBeginSymbol(Label);
     HasSectionLabel[Section] = true;
+    if (!Label->isInSection())
+      emitLabel(Label);
   }
 }
 

llvm/lib/MC/MCObjectStreamer.cpp

@@ -185,10 +185,10 @@ void MCObjectStreamer::emitLabel(MCSymbol *Symbol, SMLoc Loc) {
 
   getAssembler().registerSymbol(*Symbol);
 
-  // If there is a current fragment, mark the symbol as pointing into it.
-  // Otherwise queue the label and set its fragment pointer when we emit the
-  // next fragment.
-  MCFragment *F = getCurrentFragment();
+  // Set the fragment and offset. This function might be called by
+  // changeSection, when the section stack top hasn't been changed to the new
+  // section.
+  MCFragment *F = CurFrag;
   Symbol->setFragment(F);
   Symbol->setOffset(F->getContents().size());
 
@@ -247,6 +247,15 @@ void MCObjectStreamer::changeSection(MCSection *Section, uint32_t Subsection) {
   assert(Section && "Cannot switch to a null section!");
   getContext().clearDwarfLocSeen();
 
+  // Register the section and create an initial fragment for subsection 0
+  // if `Subsection` is non-zero.
+  bool NewSec = getAssembler().registerSection(*Section);
+  MCFragment *F0 = nullptr;
+  if (NewSec && Subsection) {
+    changeSection(Section, 0);
+    F0 = CurFrag;
+  }
+
   auto &Subsections = Section->Subsections;
   size_t I = 0, E = Subsections.size();
   while (I != E && Subsections[I].first < Subsection)
@@ -262,7 +271,13 @@ void MCObjectStreamer::changeSection(MCSection *Section, uint32_t Subsection) {
   Section->CurFragList = &Subsections[I].second;
   CurFrag = Section->CurFragList->Tail;
 
-  getAssembler().registerSection(*Section);
+  // Define the section symbol at subsection 0's initial fragment if required.
+  if (!NewSec)
+    return;
+  if (auto *Sym = Section->getBeginSymbol()) {
+    Sym->setFragment(Subsection ? F0 : CurFrag);
+    getAssembler().registerSymbol(*Sym);
+  }
 }
 
 void MCObjectStreamer::switchSectionNoPrint(MCSection *Section) {

llvm/lib/MC/MCSection.cpp

@@ -22,8 +22,7 @@ MCSection::MCSection(SectionVariant V, StringRef Name, bool IsText, bool IsBss,
                      MCSymbol *Begin)
     : Begin(Begin), HasInstructions(false), IsRegistered(false), IsText(IsText),
       IsBss(IsBss), LinkerRelaxable(false), Name(Name), Variant(V) {
-  // The initial subsection number is 0. Create a fragment list.
-  CurFragList = &Subsections.emplace_back(0u, FragList{}).second;
+  DummyFragment.setParent(this);
 }
 
 MCSymbol *MCSection::getEndSymbol(MCContext &Ctx) {