MachO: support N_INDR aliases in the assembler

[t.p.northover]

Summary

Hi,

This patch should allow LLVM to emit N_INDR symbols (essentially telling the linker that some symbol should be resolved to another named one when it becomes available).

I'm not all that good at the actual MachO object format though, so I may have made some mistakes. Does it look OK?

Cheers.

Tim.

[kledzik]

What is the visibility of alias symbols? An earlier patch had them global by default, which is unlike other symbols in assembler which are non-extern by default.

The test case relies on the order of the symbols (or relies on llvm-nm sorting them a certain way). Is there a defined order to where the N_INDR symbols are in the symbol table? The range of global symbols are sorted alphabetically.

Since no public linker supports this yet, we need to make sure that the C side is not enabled and hope no assembly files rely on the old behavior.

[t.p.northover]

Hi Nick,

Thanks for taking a look.

include/llvm/MC/MCMachObjectWriter.h

class MachObjectWriter : public MCObjectWriter {
 
   SmallString<256> StringTable;
   std::vector<MachSymbolData> LocalSymbolData;
   std::vector<MachSymbolData> ExternalSymbolData;
   std::vector<MachSymbolData> UndefinedSymbolData;
 
   /// @}
 
+  MachSymbolData *findSymbolData(const MCSymbol &Sym);
+
 public:
   MachObjectWriter(MCMachObjectTargetWriter *MOTW, raw_ostream &_OS,
                    bool _IsLittleEndian)
     : MCObjectWriter(_OS, _IsLittleEndian), TargetObjectWriter(MOTW) {
   }
 
   /// @name Lifetime management Methods
   /// @{

include/llvm/Object/SymbolicFile.h

public:
   // FIXME: should we add a SF_Text?
   enum Flags : unsigned {
     SF_None = 0,
     SF_Undefined = 1U << 0,      // Symbol is defined in another object file
     SF_Global = 1U << 1,         // Global symbol
     SF_Weak = 1U << 2,           // Weak symbol
     SF_Absolute = 1U << 3,       // Absolute symbol
     SF_Common = 1U << 4,         // Symbol has common linkage
+    SF_Indirect = 1U << 5,
     SF_FormatSpecific = 1U << 5  // Specific to the object file format
                                  // (e.g. section symbols)
   };
 
   BasicSymbolRef() : OwningObject(nullptr) { }
   BasicSymbolRef(DataRefImpl SymbolP, const SymbolicFile *Owner);
 
   bool operator==(const BasicSymbolRef &Other) const;

lib/MC/MachObjectWriter.cpp

void MachObjectWriter::WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
   Write32(0); // extreloff
   Write32(0); // nextrel
   Write32(0); // locreloff
   Write32(0); // nlocrel
 
   assert(OS.tell() - Start == sizeof(MachO::dysymtab_command));
 }
 
+MachObjectWriter::MachSymbolData *
+MachObjectWriter::findSymbolData(const MCSymbol &Sym) {
+  for (auto &Entry : LocalSymbolData)
+    if (&Entry.SymbolData->getSymbol() == &Sym)
+      return &Entry;
+
+  for (auto &Entry : ExternalSymbolData)
+    if (&Entry.SymbolData->getSymbol() == &Sym)
+      return &Entry;
+
+  for (auto &Entry : UndefinedSymbolData)
+    if (&Entry.SymbolData->getSymbol() == &Sym)
+      return &Entry;
+
+  return nullptr;
+}
+
 void MachObjectWriter::WriteNlist(MachSymbolData &MSD,
                                   const MCAsmLayout &Layout) {
   MCSymbolData &Data = *MSD.SymbolData;
-  const MCSymbol &Symbol = Data.getSymbol();
+  const MCSymbol *Symbol = &Data.getSymbol();
+  const MCSymbol *AliasedSymbol = &Symbol->AliasedSymbol();
+  uint8_t SectionIndex = MSD.SectionIndex;
   uint8_t Type = 0;
   uint16_t Flags = Data.getFlags();
   uint64_t Address = 0;
+  bool IsAlias = Symbol != AliasedSymbol;
+
+  MachSymbolData *AliaseeInfo;
+  if (IsAlias) {
+    AliaseeInfo = findSymbolData(*AliasedSymbol);
+    if (AliaseeInfo)
+      SectionIndex = AliaseeInfo->SectionIndex;
+    Symbol = AliasedSymbol;
+  }
 
   // Set the N_TYPE bits. See <mach-o/nlist.h>.
   //
   // FIXME: Are the prebound or indirect fields possible here?
-  if (Symbol.isUndefined())
+  if (IsAlias && Symbol->isUndefined())
+    Type = MachO::N_INDR;
+  else if (Symbol->isUndefined())
     Type = MachO::N_UNDF;
-  else if (Symbol.isAbsolute())
+  else if (Symbol->isAbsolute())
     Type = MachO::N_ABS;
   else
     Type = MachO::N_SECT;
 
   // FIXME: Set STAB bits.
 
   if (Data.isPrivateExtern())
     Type |= MachO::N_PEXT;
 
   // Set external bit.
-  if (Data.isExternal() || Symbol.isUndefined())
+  if (Data.isExternal() || (!IsAlias && Symbol->isUndefined()))
     Type |= MachO::N_EXT;
 
   // Compute the symbol address.
-  if (Symbol.isDefined()) {
+  if (IsAlias && Symbol->isUndefined())
+    Address = AliaseeInfo->StringIndex;
+  else if (Symbol->isDefined())
     Address = getSymbolAddress(&Data, Layout);
-  } else if (Data.isCommon()) {
+  else if (Data.isCommon()) {
     // Common symbols are encoded with the size in the address
     // field, and their alignment in the flags.
     Address = Data.getCommonSize();
 
     // Common alignment is packed into the 'desc' bits.
     if (unsigned Align = Data.getCommonAlignment()) {
       unsigned Log2Size = Log2_32(Align);
       assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
       if (Log2Size > 15)
         report_fatal_error("invalid 'common' alignment '" +
-                           Twine(Align) + "' for '" + Symbol.getName() + "'",
+                           Twine(Align) + "' for '" + Symbol->getName() + "'",
                            false);
       // FIXME: Keep this mask with the SymbolFlags enumeration.
       Flags = (Flags & 0xF0FF) | (Log2Size << 8);
     }
   }
 
-  if (Layout.getAssembler().isThumbFunc(&Symbol))
+  if (Layout.getAssembler().isThumbFunc(Symbol))
     Flags |= SF_ThumbFunc;
 
   // struct nlist (12 bytes)
 
   Write32(MSD.StringIndex);
   Write8(Type);
-  Write8(MSD.SectionIndex);
+  Write8(SectionIndex);
 
   // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
   // value.
   Write16(Flags);
   if (is64Bit())
     Write64(Address);
   else
     Write32(Address);

lib/Object/MachOObjectFile.cpp

uint32_t MachOObjectFile::getSymbolFlags(DataRefImpl DRI) const {
   uint8_t MachOType = Entry.n_type;
   uint16_t MachOFlags = Entry.n_desc;
 
   uint32_t Result = SymbolRef::SF_None;
 
   if ((MachOType & MachO::N_TYPE) == MachO::N_UNDF)
     Result |= SymbolRef::SF_Undefined;
 
+  if ((MachOType & MachO::N_TYPE) == MachO::N_INDR)
+    Result |= SymbolRef::SF_Indirect;
+
   if (MachOType & MachO::N_STAB)
     Result |= SymbolRef::SF_FormatSpecific;
 
   if (MachOType & MachO::N_EXT) {
     Result |= SymbolRef::SF_Global;
     if ((MachOType & MachO::N_TYPE) == MachO::N_UNDF) {
       uint64_t Value;
       getSymbolAddress(DRI, Value);

test/MC/MachO/aliased-symbols.s

+// RUN: llvm-mc -triple thumbv7m-apple-darwin-eabi %s -filetype=obj -o %t
+// RUN:     llvm-readobj -symbols %t | FileCheck %s
+
+        .data
+        var1 = var2
+        .long var1
+        .long var2
+        .long var2 + 4
+defined_early:
+        .long 0
+
+        alias_to_early = defined_early
+        alias_to_late = defined_late
+
+defined_late:
+        .long 0
+
+        .global extern_test
+        extern_test = var2
+
+        alias_to_local = Ltmp0
+Ltmp0:
+
+// CHECK: Symbols [
+
+        // defined_early was defined. Actually has value 0xc.
+// CHECK: Symbol {
+// CHECK-NEXT:   Name: defined_early
+// CHECK-NEXT:   Type: Section (0xE)
+// CHECK-NEXT:   Section: __data (0x2)
+// CHECK-NEXT:   RefType: UndefinedNonLazy (0x0)
+// CHECK-NEXT:   Flags [ (0x0)
+// CHECK-NEXT:   ]
+// CHECK-NEXT:   Value: 0x[[DEFINED_EARLY:[0-9A-F]+]]
+// CHECK-NEXT: }
+
+        // alias_to_early was an alias to defined_early. But we can resolve it.
+// CHECK: Symbol {
+// CHECK-NEXT:   Name: alias_to_early
+// CHECK-NEXT:   Type: Section (0xE)
+// CHECK-NEXT:   Section: __data (0x2)
+// CHECK-NEXT:   RefType: UndefinedNonLazy (0x0)
+// CHECK-NEXT:   Flags [ (0x0)
+// CHECK-NEXT:   ]
+// CHECK-NEXT:   Value: 0x[[DEFINED_EARLY]]
+// CHECK-NEXT: }
+
+        // defined_late was defined. Just after defined_early.
+// CHECK: Symbol {
+// CHECK-NEXT:   Name: defined_late
+// CHECK-NEXT:   Type: Section (0xE)
+// CHECK-NEXT:   Section: __data (0x2)
+// CHECK-NEXT:   RefType: UndefinedNonLazy (0x0)
+// CHECK-NEXT:   Flags [ (0x0)
+// CHECK-NEXT:   ]
+// CHECK-NEXT:   Value: 0x[[DEFINED_LATE:[0-9A-F]+]]
+// CHECK-NEXT: }
+
+        // alias_to_late was an alias to defined_late. But we can resolve it.
+// CHECK: Symbol {
+// CHECK-NEXT:   Name: alias_to_late
+// CHECK-NEXT:   Type: Section (0xE)
+// CHECK-NEXT:   Section: __data (0x2)
+// CHECK-NEXT:   RefType: UndefinedNonLazy (0x0)
+// CHECK-NEXT:   Flags [ (0x0)
+// CHECK-NEXT:   ]
+// CHECK-NEXT:   Value: 0x[[DEFINED_LATE]]
+// CHECK-NEXT: }
+
+        // alias_to_local is an alias, but what it points to has no
+        // MachO representation. We must resolve it.
+// CHECK: Symbol {
+// CHECK-NEXT:   Name: alias_to_local (37)
+// CHECK-NEXT:   Type: Section (0xE)
+// CHECK-NEXT:   Section:  (0x0)
+// CHECK-NEXT:   RefType: UndefinedNonLazy (0x0)
+// CHECK-NEXT:   Flags [ (0x0)
+// CHECK-NEXT:   ]
+// CHECK-NEXT:   Value: 0x14
+// CHECK-NEXT: }
+
+        // extern_test was a pure alias to the unknown "var2".
+        // N_INDR and Extern.
+// CHECK:   Name: extern_test
+// CHECK-NEXT:   Extern
+// CHECK-NEXT:   Type: Indirect (0xA)
+// CHECK-NEXT:   Section:  (0x0)
+// CHECK-NEXT:   RefType: UndefinedNonLazy (0x0)
+// CHECK-NEXT:   Flags [ (0x0)
+// CHECK-NEXT:   ]
+// CHECK-NEXT:   Value: 0x[[VAR2_STRINGINDEX:[0-9a-f]+]]
+// CHECK-NEXT: }
+
+        // var1 was another alias to an unknown variable. Not extern this time.
+// CHECK: Symbol {
+// CHECK-NEXT:   Name: var1 (1)
+// CHECK-NEXT:   Type: Indirect (0xA)
+// CHECK-NEXT:   Section:  (0x0)
+// CHECK-NEXT:   RefType: UndefinedNonLazy (0x0)
+// CHECK-NEXT:   Flags [ (0x0)
+// CHECK-NEXT:   ]
+// CHECK-NEXT:   Value: 0x[[VAR2_STRINGINDEX]]
+// CHECK-NEXT: }
+
+        // var2 was a normal undefined (extern) symbol.
+// CHECK: Symbol {
+// CHECK-NEXT:   Name: var2
+// CHECK-NEXT:   Extern
+// CHECK-NEXT:   Type: Undef (0x0)
+// CHECK-NEXT:   Section:  (0x0)
+// CHECK-NEXT:   RefType: UndefinedNonLazy (0x0)
+// CHECK-NEXT:   Flags [ (0x0)
+// CHECK-NEXT:   ]
+// CHECK-NEXT:   Value: 0x0
+// CHECK-NEXT: }

test/MC/MachO/variable-exprs.s

 // CHECK-I386:     ('n_type', 0x1)
 // CHECK-I386:     ('n_sect', 0)
 // CHECK-I386:     ('n_desc', 0)
 // CHECK-I386:     ('n_value', 0)
 // CHECK-I386:     ('_string', 'd')
 // CHECK-I386:    ),
 // CHECK-I386:     # Symbol 8
 // CHECK-I386:    (('n_strx', 1)
-// CHECK-I386:     ('n_type', 0x1)
+// CHECK-I386:     ('n_type', 0xb)
 // CHECK-I386:     ('n_sect', 0)
 // CHECK-I386:     ('n_desc', 0)
-// CHECK-I386:     ('n_value', 0)
+// CHECK-I386:     ('n_value', 4)
 // CHECK-I386:     ('_string', 'd2')
 // CHECK-I386:    ),
 // CHECK-I386:     # Symbol 9
 // CHECK-I386:    (('n_strx', 6)
 // CHECK-I386:     ('n_type', 0x1)
 // CHECK-I386:     ('n_sect', 0)
 // CHECK-I386:     ('n_desc', 0)
 // CHECK-I386:     ('n_value', 0)

 // CHECK-X86_64:     ('n_type', 0x1)
 // CHECK-X86_64:     ('n_sect', 0)
 // CHECK-X86_64:     ('n_desc', 0)
 // CHECK-X86_64:     ('n_value', 0)
 // CHECK-X86_64:     ('_string', 'd')
 // CHECK-X86_64:    ),
 // CHECK-X86_64:     # Symbol 8
 // CHECK-X86_64:    (('n_strx', 1)
-// CHECK-X86_64:     ('n_type', 0x1)
+// CHECK-X86_64:     ('n_type', 0xb)
 // CHECK-X86_64:     ('n_sect', 0)
 // CHECK-X86_64:     ('n_desc', 0)
-// CHECK-X86_64:     ('n_value', 0)
+// CHECK-X86_64:     ('n_value', 4)
 // CHECK-X86_64:     ('_string', 'd2')
 // CHECK-X86_64:    ),
 // CHECK-X86_64:     # Symbol 9
 // CHECK-X86_64:    (('n_strx', 6)
 // CHECK-X86_64:     ('n_type', 0x1)
 // CHECK-X86_64:     ('n_sect', 0)
 // CHECK-X86_64:     ('n_desc', 0)
 // CHECK-X86_64:     ('n_value', 0)

tools/llvm-nm/llvm-nm.cpp

 
 static char getSymbolNMTypeChar(MachOObjectFile &Obj, basic_symbol_iterator I) {
   DataRefImpl Symb = I->getRawDataRefImpl();
   uint8_t NType = getNType(Obj, Symb);
 
   switch (NType & MachO::N_TYPE) {
   case MachO::N_ABS:
     return 's';
+  case MachO::N_INDR:
+    return 'i';
   case MachO::N_SECT: {
     section_iterator Sec = Obj.section_end();
     Obj.getSymbolSection(Symb, Sec);
     DataRefImpl Ref = Sec->getRawDataRefImpl();
     StringRef SectionName;
     Obj.getSectionName(Ref, SectionName);
     StringRef SegmentName = Obj.getSectionFinalSegmentName(Ref);
     if (SegmentName == "__TEXT" && SectionName == "__text")