[lld-macho] Emit STABS symbols for debugging, and drop debug sections

[int3]

Summary

Debug sections contain a large amount of data. In order not to bloat the size
of the final binary, we remove them and instead emit STABS symbols for
dsymutil and the debugger to locate their contents in the object files.

With this diff, dsymutil is able to locate the debug info. However, we need
a few more features before lldb is able to work well with our binaries --
e.g. having LC_DYSYMTAB accurately reflect the number of local symbols,
emitting LC_UUID, and more. Those will be handled in follow-up diffs.

Note also that the STABS we emit differ slightly from what ld64 does. First, we
emit the path to the source file as one N_SO symbol instead of two. (ld64
emits one N_SO for the dirname and one of the basename.) Second, we do not
emit N_BNSYM and N_ENSYM STABS to mark the start and end of functions,
because the N_FUN STABS already serve that purpose. @clayborg recommended
these changes based on his knowledge of what the debugging tools look for.

Additionally, this current implementation doesn't accurately reflect the size
of function symbols. It uses the size of their containing sectioins as a proxy,
but that is only accurate if .subsections_with_symbols is set, and if there
isn't an N_ALT_ENTRY in that particular subsection. I think we have two
options to solve this:

1. We can split up subsections by symbol even if .subsections_with_symbols is not set, but include constraints to ensure those subsections retain their order in the final output. This is ld64's approach.
2. We could just add a size field to our Symbol class. This seems simpler, and I'm more inclined toward it, but I'm not sure if there are use cases that it doesn't handle well. As such I'm punting on the decision for now.

[int3]

don't run test on windows; fixing the paths there is more work than I'd like to do for now

[smeenai]

We could just add a size field to our Symbol class. This seems simpler, and I'm more inclined toward it, but I'm not sure if there are use cases that it doesn't handle well. As such I'm punting on the decision for now.

Yeah, I like this more too, but punting on it until later is fine.

As far as I can see, we only need a very small amount of information from the DWARF in the object file (the source file information). Do you know if the DWARF parsing is done lazily (so we only pay the cost for the parts of the DWARF we care about), or if it'll attempt to parse the entirety of the object file's DWARF (which would be wasteful)?

lld/MachO/SyntheticSections.cpp

618

I guess we can optimize this later to not duplicate strings in the string table? (Since the symbol table entry will also add the symbol name.)

649–650

I think this is valid. How does ld64 make this determination though?

[clayborg]

lld/MachO/SyntheticSections.cpp

610

stab.value should be set to the modification date of the .o file here. If this is a path to a .o file, then the mod time of the .o file, and if this is a "foo.a(bar.o)", it needs to be the modification time of the .o file as mentioned in the BSD archive file data structures. This is important because sometimes .a file have multiple .o files with the same name and the only way to tell them apart is the mod time.

617

"stab.sect" needs to be set to be the 1 based mach-o section index that contains the "stab.value" address within the mach-o file. Any symbol that you emit that has an address in "stab.sect" has its "stab.sect" set to the 1 based section index. So you take all of the sections from all LC_SEGMENT or LC_SEGMENT_64 load commands and make a list of them and the first one will have index 1. From my python tool I have something that dumps the sections and shows the index:

$ mach_o.py a.out --sections
FILE OFF    INDEX ADDRESS            SIZE               OFFSET     ALIGN      RELOFF     NRELOC     FLAGS      RESERVED1  RESERVED2  RESERVED3  NAME
=========== ===== ------------------ ------------------ ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ----------------------
0x000000b0: [  1] 0x0000000100000ec0 0x00000000000000c1 0x00000ec0 0x00000004 0x00000000 0x00000000 0x80000400 0x00000000 0x00000000 0x00000000 __TEXT.__text
0x00000100: [  2] 0x0000000100000f82 0x0000000000000006 0x00000f82 0x00000001 0x00000000 0x00000000 0x80000408 0x00000000 0x00000006 0x00000000 __TEXT.__stubs
0x00000150: [  3] 0x0000000100000f88 0x000000000000001a 0x00000f88 0x00000002 0x00000000 0x00000000 0x80000400 0x00000000 0x00000000 0x00000000 __TEXT.__stub_helper
0x000001a0: [  4] 0x0000000100000fa2 0x000000000000000d 0x00000fa2 0x00000000 0x00000000 0x00000000 0x00000002 0x00000000 0x00000000 0x00000000 __TEXT.__cstring
0x000001f0: [  5] 0x0000000100000fb0 0x0000000000000048 0x00000fb0 0x00000002 0x00000000 0x00000000 0x00000000 0x00000000 0x00000000 0x00000000 __TEXT.__unwind_info
0x00000288: [  6] 0x0000000100001000 0x0000000000000008 0x00001000 0x00000003 0x00000000 0x00000000 0x00000006 0x00000001 0x00000000 0x00000000 __DATA_CONST.__got
0x00000320: [  7] 0x0000000100002000 0x0000000000000008 0x00002000 0x00000003 0x00000000 0x00000000 0x00000007 0x00000002 0x00000000 0x00000000 __DATA.__la_symbol_ptr
0x00000370: [  8] 0x0000000100002008 0x0000000000000008 0x00002008 0x00000003 0x00000000 0x00000000 0x00000000 0x00000000 0x00000000 0x00000000 __DATA.__data

LLDB uses this section index to grab the right section when it parses the symbol.

618

Yes, this should be uniqued if it isn't happening already. Large C++ compiles will really trigger this and cause huge symbol table bloat.

635

Yes, we need to emit N_GSYM symbols for globals, and N_STSYM symbols for static variables. Those need to be linked in the debug info.

N_STSYM always has the address set to a valid address. The stab entry should have the "value" set to the address, and "sect" set to the right section index just like the N_FUN entry.

N_GSYM may or may not have the address filed in. If the address of the global can be calculated, the "value" should be set to the address, and "sect" set to the right section index just like the N_FUN entry. If the address can't be calculated, N_GSYM can be zero and we will need to find the global by matching up with the non STAB symbol whose name must match exactly (and that symbol will have "value" and "sect" set correctly).

649–650

Not necessarily. Read only globals can be in __text IIRC on ARM and possibly ARM64 or even other architectures.

653

Is this truly unreachable code?

657–659

Is this list sorted by file? We don't want to be emitting multiple N_SO + N_OSO entries for the same source file. dsymutil will be a lot less efficient if we have a few entries for "/tmp/foo.cpp" and "/tmp/foo.o" and then a few function, then some other source + object file, and then "/tmp/foo.cpp" and "/tmp/foo.o" again.

697–698

So in mach-o files, and preparing for lld emitting a LC_DYSYMTAB load command, all local symbols must be in a contiguous block of symbols. In fact all locals, external, and undefined symbols must be in a contiguous block. That is required because if "ilocalsym" represents the first local symbol table index of all of the local symbols and "nlocalsym" is the out. Same for external with "iextdefsym" and "nextdefsym", and for undefined symbols with "iundefsym" and "nundefsym".

$ otool -lv a.out
...
Load command 7
            cmd LC_DYSYMTAB
        cmdsize 80
      ilocalsym 0
      nlocalsym 17
     iextdefsym 17
     nextdefsym 4
      iundefsym 21
      nundefsym 2

[int3]

@smeenai Dwarf parsing appears to be on-demand, though I don't know if it's parsing just the subset of things we end up using. Specifically, compile_units() calls parseNormalUnits, which parses all units of type DW_SECT_INFO and DW_SECT_EXT_TYPES. I'll add a TODO to investigate more later.

@clayborg thanks for all the insightful comments! I'll work the modtime and n_sect follow-up diffs tomorrow so that they can be reviewed as a coherent whole.

lld/MachO/SyntheticSections.cpp

610

yeah I was going to do it in a follow-up. Will add a TODO here. Didn't think about the archive case though, thanks for the tip!

617

Ah, I was following ld64's code here which also hard-codes a sect value of 1 for N_FUN stabs, but I didn't understand the significance of the value. Thanks for the explanation! I think it works out in ld64's case because ld64 puts all functions in __text in the final binary. Functions that are in non-text sections in the input object files all get coalesced into __text. LLD has yet to implement that behavior, but we plan on doing so.

I think I'll add support for this in the same diff where I'll add support for N_GSYM. That way we can write a test for values of n_sect other than 1.

618

ld64 doesn't appear to do it, but it does seem like a worthwhile optimization. I'll add a TODO inside StringTableSection::addString.

649–650

ld64 checks if their atoms have a type of typeCode. That type is determined by their sectionType() method, which uses a combination of input section flags and section names to determine it. I *think* any __text input section will be of typeCode, but I'll have to investigate further. I think we can punt on this.

653

The only file types are object files, bitcode files, archives, and dylibs. We extract ObjFiles from ArchiveFiles & create them from BitcodeFiles before processing their InputSections, so we shouldn't see either of those here. And we shouldn't be emitting debug info for functions that our output binary references from dylibs (I think... correct me if I'm wrong).

657–659

yeah sorting did occur to me but I thought I'd implement it later. It's probably quite straightforward though, so I can do it in this diff

697–698

yup, that is done in D89285: [lld-macho] Emit local symbols in symtab; record metadata in LC_DYSYMTAB

[int3]

lld/MachO/SyntheticSections.cpp

610

Done in D91318: [lld-macho] Add archive name and file modtime to STABS output

[int3]

update

[int3]

Sorry this took a while... got a bit sidetracked

lld/MachO/SyntheticSections.cpp

635

Addressed in D92366: [lld-macho] Flesh out STABS implementation. Though I'm not sure how to test the case where we have a defined global whose address can't be calculated. How does that situation arise?

653

After thinking about it a bit more, we should really refactor things such that this cast isn't necessary. But it's a bit hairy, so not in this diff

657–659

Done in D92366: [lld-macho] Flesh out STABS implementation

[clayborg]

So marking as needing changes because the size of the N_FUN entries must be correct. dsymutil creates an address map when remapping symbols with ranges and if the size of the N_FUN stabs entry is too large, it will cause major problems when linking the DWARF. We should also find a rock solid way to identify N_FUN entries if possible instead of relying on the section name being "__text".

lld/MachO/SyntheticSections.cpp

610

Sounds good!

617

Sounds good.

627

This can't be wrong and must be correct. If it is wrong, dsymutil will link the wrong address ranges for each function and really bad DWARF will result.

635

This can be done by just declaring a global variable and not assigning it a value:

$ cat main.c
int global;

int main(int argc, const char **argv) {
  return 0;
}
$ clang -c -g main.c
$ clang -g main.o
$ dsymutil -s a.out 
----------------------------------------------------------------------
Symbol table for: 'a.out' (x86_64)
----------------------------------------------------------------------
Index    n_strx   n_type             n_sect n_desc n_value
======== -------- ------------------ ------ ------ ----------------
[     0] 00000035 64 (N_SO         ) 00     0000   0000000000000000 '/tmp/'
[     1] 0000003b 64 (N_SO         ) 00     0000   0000000000000000 'main.c'
[     2] 00000042 66 (N_OSO        ) 03     0001   000000005fc5ec13 '/private/tmp/main.o'
[     3] 00000001 2e (N_BNSYM      ) 01     0000   0000000100003fa0
[     4] 00000056 24 (N_FUN        ) 01     0000   0000000100003fa0 '_main'
[     5] 00000001 24 (N_FUN        ) 00     0000   0000000000000016
[     6] 00000001 4e (N_ENSYM      ) 01     0000   0000000000000016
[     7] 0000005c 20 (N_GSYM       ) 00     0000   0000000000000000 '_global'
[     8] 00000001 64 (N_SO         ) 01     0000   0000000000000000
[     9] 00000002 0f (     SECT EXT) 01     0010   0000000100000000 '__mh_execute_header'
[    10] 00000016 0f (     SECT EXT) 03     0000   0000000100004000 '_global'
[    11] 0000001e 0f (     SECT EXT) 01     0000   0000000100003fa0 '_main'
[    12] 00000024 01 (     UNDF EXT) 00     0100   0000000000000000 'dyld_stub_binder'

Note how symbol 7 has no valid address even though the actual non debug symbol does symbol 10. If lld can always emit the valid address for the STAB entry, that is quite ok. Just noting that ld64 doesn't always know it for some reason (or what ever linker the clang driver is evoking these days...).

650

I think we really need to get this right. Any section can contain functions. Most of the time, functions are in __text, but not always.

659

nice

[int3]

lld/MachO/SyntheticSections.cpp

627

I'd like to punt on this for now. The goal for now is not to get the STABS implementation to 100% correctness, but to create sort-of-working binaries for local testing. As mentioned in the commit message, I'm on the fence about the best way to store symbol size in LLD's current architecture, and I think it will be clearer once more of the linker has been implemented.

635

ah... yeah LLD is able to emit the address for BSS symbols. I've added a test for that particular case in the other diff

650

addressed in D92430: [lld-macho] Add isCodeSection()

[clayborg]

lld/MachO/SyntheticSections.cpp

627

understood, just wanted you to be aware that debugging these binaries or using a dSYM file generated from these binaries will have issues until this is fixed.

[thakis]

lld/MachO/InputFiles.cpp

401

Since the arg here is called err, was this supposed to call error() instead of warn()? If not, this maybe deserves a comment.

[int3]

lld/MachO/InputFiles.cpp

401

This is copypasta from *ObjFile<ELFT>::getDwarf(). I'm not actually sure what kinds of malformed input triggers this...

lld/MachO/CMakeLists.txt

 set(LLVM_TARGET_DEFINITIONS Options.td)
 tablegen(LLVM Options.inc -gen-opt-parser-defs)
 add_public_tablegen_target(MachOOptionsTableGen)
 
 include_directories(${LLVM_MAIN_SRC_DIR}/../libunwind/include)
 
 add_lld_library(lldMachO2
   Arch/X86_64.cpp
   UnwindInfoSection.cpp
   Driver.cpp
   DriverUtils.cpp
+  Dwarf.cpp
   ExportTrie.cpp
   InputFiles.cpp
   InputSection.cpp
   MergedOutputSection.cpp
   ObjC.cpp
   OutputSection.cpp
   OutputSegment.cpp
   SymbolTable.cpp

lld/MachO/Dwarf.h

+//===- DWARF.h -----------------------------------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===-------------------------------------------------------------------===//
+
+#ifndef LLD_MACHO_DWARF_H
+#define LLD_MACHO_DWARF_H
+
+#include "llvm/ADT/StringRef.h"
+#include "llvm/DebugInfo/DWARF/DWARFObject.h"
+
+namespace lld {
+namespace macho {
+
+class ObjFile;
+
+// Implements the interface between LLVM's DWARF-parsing utilities and LLD's
+// InputSection structures.
+class DwarfObject final : public llvm::DWARFObject {
+public:
+  bool isLittleEndian() const override { return true; }
+
+  llvm::Optional<llvm::RelocAddrEntry> find(const llvm::DWARFSection &sec,
+                                            uint64_t pos) const override {
+    // TODO: implement this
+    return llvm::None;
+  }
+
+  void forEachInfoSections(
+      llvm::function_ref<void(const llvm::DWARFSection &)> f) const override {
+    f(infoSection);
+  }
+
+  llvm::StringRef getAbbrevSection() const override { return abbrevSection; }
+  llvm::StringRef getStrSection() const override { return strSection; }
+
+  // Returns an instance of DwarfObject if the given object file has the
+  // relevant DWARF debug sections.
+  static std::unique_ptr<DwarfObject> create(ObjFile *);
+
+private:
+  llvm::DWARFSection infoSection;
+  llvm::StringRef abbrevSection;
+  llvm::StringRef strSection;
+};
+
+} // namespace macho
+} // namespace lld
+
+#endif

lld/MachO/Dwarf.cpp

+//===- DWARF.cpp ----------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "Dwarf.h"
+#include "InputFiles.h"
+#include "InputSection.h"
+#include "OutputSegment.h"
+
+#include <memory>
+
+using namespace lld;
+using namespace lld::macho;
+using namespace llvm;
+
+std::unique_ptr<DwarfObject> DwarfObject::create(ObjFile *obj) {
+  auto dObj = std::make_unique<DwarfObject>();
+  bool hasDwarfInfo = false;
+  for (SubsectionMap subsecMap : obj->subsections) {
+    for (auto it : subsecMap) {
+      InputSection *isec = it.second;
+      if (!(isDebugSection(isec->flags) &&
+            isec->segname == segment_names::dwarf))
+        continue;
+
+      if (isec->name == "__debug_info") {
+        dObj->infoSection.Data = toStringRef(isec->data);
+        hasDwarfInfo = true;
+        continue;
+      }
+
+      if (StringRef *s = StringSwitch<StringRef *>(isec->name)
+                             .Case("__debug_abbrev", &dObj->abbrevSection)
+                             .Case("__debug_str", &dObj->strSection)
+                             .Default(nullptr)) {
+        *s = toStringRef(isec->data);
+        hasDwarfInfo = true;
+      }
+    }
+  }
+
+  if (hasDwarfInfo)
+    return dObj;
+  return nullptr;
+}

lld/MachO/InputFiles.h

 #define LLD_MACHO_INPUT_FILES_H
 
 #include "MachOStructs.h"
 
 #include "lld/Common/LLVM.h"
 #include "lld/Common/Memory.h"
 #include "llvm/ADT/DenseSet.h"
 #include "llvm/BinaryFormat/MachO.h"
+#include "llvm/DebugInfo/DWARF/DWARFUnit.h"
 #include "llvm/Object/Archive.h"
 #include "llvm/Support/MemoryBuffer.h"
 #include "llvm/TextAPI/MachO/InterfaceFile.h"
 #include "llvm/TextAPI/MachO/TextAPIReader.h"
 
 #include <map>
 #include <vector>
 
private:
   const StringRef name;
 };
 
 // .o file
 class ObjFile : public InputFile {
 public:
   explicit ObjFile(MemoryBufferRef mb);
   static bool classof(const InputFile *f) { return f->kind() == ObjKind; }
+
+  llvm::DWARFUnit *compileUnit = nullptr;
+
+private:
+  void parseDebugInfo();
 };
 
 // command-line -sectcreate file
 class OpaqueFile : public InputFile {
 public:
   explicit OpaqueFile(MemoryBufferRef mb, StringRef segName,
                       StringRef sectName);
   static bool classof(const InputFile *f) { return f->kind() == OpaqueKind; }

lld/MachO/InputFiles.cpp

 // Without the above differences, I think you can use your knowledge about ELF
 // and COFF for Mach-O.
 //
 //===----------------------------------------------------------------------===//
 
 #include "InputFiles.h"
 #include "Config.h"
 #include "DriverUtils.h"
+#include "Dwarf.h"
 #include "ExportTrie.h"
 #include "InputSection.h"
 #include "MachOStructs.h"
 #include "ObjC.h"
 #include "OutputSection.h"
 #include "OutputSegment.h"
 #include "SymbolTable.h"
 #include "Symbols.h"
 #include "Target.h"
 
+#include "lld/Common/DWARF.h"
 #include "lld/Common/ErrorHandler.h"
 #include "lld/Common/Memory.h"
 #include "llvm/ADT/iterator.h"
 #include "llvm/BinaryFormat/MachO.h"
 #include "llvm/Support/Endian.h"
 #include "llvm/Support/MemoryBuffer.h"
 #include "llvm/Support/Path.h"
 
if (const load_command *cmd = findCommand(hdr, LC_SYMTAB)) {
     bool subsectionsViaSymbols = hdr->flags & MH_SUBSECTIONS_VIA_SYMBOLS;
     parseSymbols(nList, strtab, subsectionsViaSymbols);
   }
 
   // The relocations may refer to the symbols, so we parse them after we have
   // parsed all the symbols.
   for (size_t i = 0, n = subsections.size(); i < n; ++i)
     parseRelocations(sectionHeaders[i], subsections[i]);
+
+  parseDebugInfo();
+}
+
+void ObjFile::parseDebugInfo() {
+  std::unique_ptr<DwarfObject> dObj = DwarfObject::create(this);
+  if (!dObj)
+    return;
+
+  auto *ctx = make<DWARFContext>(
+      std::move(dObj), "",
+      [&](Error err) { warn(getName() + ": " + toString(std::move(err))); },
+      [&](Error warning) {
+        warn(getName() + ": " + toString(std::move(warning)));
+      });
+
+  const DWARFContext::compile_unit_range &units = ctx->compile_units();
+  auto it = units.begin();
+  compileUnit = it->get();
+  assert(std::next(it) == units.end());
 }
 
 // The path can point to either a dylib or a .tbd file.
 static Optional<DylibFile *> loadDylib(StringRef path, DylibFile *umbrella) {
   Optional<MemoryBufferRef> mbref = readFile(path);
   if (!mbref) {
     error("could not read dylib file at " + path);
     return {};

lld/MachO/InputSection.h

struct Reloc {
   // to.
   uint32_t offset;
   // Adding this offset to the address of the referent symbol or subsection
   // gives the destination that this relocation refers to.
   uint64_t addend;
   llvm::PointerUnion<Symbol *, InputSection *> referent;
 };
 
-inline bool isZeroFill(uint8_t flags) {
+inline bool isZeroFill(uint32_t flags) {
   return llvm::MachO::isVirtualSection(flags & llvm::MachO::SECTION_TYPE);
 }
 
-inline bool isThreadLocalVariables(uint8_t flags) {
+inline bool isThreadLocalVariables(uint32_t flags) {
   return (flags & llvm::MachO::SECTION_TYPE) ==
          llvm::MachO::S_THREAD_LOCAL_VARIABLES;
 }
 
+inline bool isDebugSection(uint32_t flags) {
+  return (flags & llvm::MachO::SECTION_ATTRIBUTES_USR) ==
+         llvm::MachO::S_ATTR_DEBUG;
+}
+
 class InputSection {
 public:
   virtual ~InputSection() = default;
   virtual uint64_t getSize() const { return data.size(); }
   virtual uint64_t getFileSize() const {
     return isZeroFill(flags) ? 0 : getSize();
   }
   uint64_t getFileOffset() const;

lld/MachO/OutputSegment.h

 namespace segment_names {
 
 constexpr const char pageZero[] = "__PAGEZERO";
 constexpr const char text[] = "__TEXT";
 constexpr const char data[] = "__DATA";
 constexpr const char linkEdit[] = "__LINKEDIT";
 constexpr const char dataConst[] = "__DATA_CONST";
 constexpr const char ld[] = "__LD"; // output only with -r
+constexpr const char dwarf[] = "__DWARF";
 
 } // namespace segment_names
 
 class OutputSection;
 class InputSection;
 
 class OutputSegment {
 public:

lld/MachO/SyntheticSections.h

 #include "OutputSection.h"
 #include "OutputSegment.h"
 #include "Target.h"
 
 #include "llvm/ADT/PointerUnion.h"
 #include "llvm/ADT/SetVector.h"
 #include "llvm/Support/raw_ostream.h"
 
+namespace llvm {
+class DWARFUnit;
+} // namespace llvm
+
 namespace lld {
 namespace macho {
 
 namespace section_names {
 
 constexpr const char pageZero[] = "__pagezero";
 constexpr const char common[] = "__common";
 constexpr const char header[] = "__mach_header";
private:
   size_t size = 1;
 };
 
 struct SymtabEntry {
   Symbol *sym;
   size_t strx;
 };
 
+struct StabsEntry {
+  uint8_t type;
+  uint32_t strx = 0;
+  uint8_t sect = 0;
+  uint16_t desc = 0;
+  uint64_t value = 0;
+
+  explicit StabsEntry(uint8_t type): type(type) {}
+};
+
 class SymtabSection : public LinkEditSection {
 public:
   SymtabSection(StringTableSection &);
   void finalizeContents();
-  size_t getNumSymbols() const { return symbols.size(); }
+  size_t getNumSymbols() const { return stabs.size() + symbols.size(); }
   uint64_t getRawSize() const override;
   void writeTo(uint8_t *buf) const override;
 
 private:
+  void emitBeginSourceStab(llvm::DWARFUnit *compileUnit);
+  void emitEndSourceStab();
+  void emitObjectFileStab(InputFile *);
+  void emitFunStabs(Defined *);
+
   StringTableSection &stringTableSection;
+  std::vector<StabsEntry> stabs;
   std::vector<SymtabEntry> symbols;
 };
 
 // The indirect symbol table is a list of 32-bit integers that serve as indices
 // into the (actual) symbol table. The indirect symbol table is a
 // concatentation of several sub-arrays of indices, each sub-array belonging to
 // a separate section. The starting offset of each sub-array is stored in the
 // reserved1 header field of the respective section.

lld/MachO/SyntheticSections.cpp

 #include "OutputSegment.h"
 #include "SymbolTable.h"
 #include "Symbols.h"
 #include "Writer.h"
 
 #include "lld/Common/ErrorHandler.h"
 #include "lld/Common/Memory.h"
 #include "llvm/Support/EndianStream.h"
+#include "llvm/Support/FileSystem.h"
 #include "llvm/Support/LEB128.h"
+#include "llvm/Support/Path.h"
 
 using namespace llvm;
 using namespace llvm::support;
 using namespace llvm::support::endian;
 using namespace lld;
 using namespace lld::macho;
 
 InStruct macho::in;

 
 void ExportSection::writeTo(uint8_t *buf) const { trieBuilder.writeTo(buf); }
 
 SymtabSection::SymtabSection(StringTableSection &stringTableSection)
     : LinkEditSection(segment_names::linkEdit, section_names::symbolTable),
       stringTableSection(stringTableSection) {}
 
 uint64_t SymtabSection::getRawSize() const {
-  return symbols.size() * sizeof(structs::nlist_64);
+  return getNumSymbols() * sizeof(structs::nlist_64);
+}
+
+void SymtabSection::emitBeginSourceStab(DWARFUnit *compileUnit) {
+  StabsEntry stab(MachO::N_SO);
+  SmallString<261> dir(compileUnit->getCompilationDir());
+  StringRef sep = sys::path::get_separator();
+  // We don't use `path::append` here because we want an empty `dir` to result
+  // in an absolute path. `append` would give us a relative path for that case.
+  if (!dir.endswith(sep))
+    dir += sep;
+  stab.strx = stringTableSection.addString(
+      saver.save(dir + compileUnit->getUnitDIE().getShortName()));
+  stabs.emplace_back(std::move(stab));
+}
+
+void SymtabSection::emitEndSourceStab() {
+  StabsEntry stab(MachO::N_SO);
+  stab.sect = 1;
+  stabs.emplace_back(std::move(stab));
+}
+
+void SymtabSection::emitObjectFileStab(InputFile *file) {
+  StabsEntry stab(MachO::N_OSO);
+  stab.sect = target->cpuSubtype;
+  SmallString<261> path(file->getName());
+  std::error_code ec = sys::fs::make_absolute(path);
+  if (ec)
+    fatal("failed to get absolute path for " + file->getName());
+
+  stab.strx = stringTableSection.addString(saver.save(path.str()));
+  stab.desc = 1;
+  stabs.emplace_back(std::move(stab));
+}
+
+void SymtabSection::emitFunStabs(Defined *defined) {
+  {
+    StabsEntry stab(MachO::N_FUN);
+    stab.sect = 1;
+    stab.strx = stringTableSection.addString(defined->getName());
+    stab.value = defined->getVA();
+    stabs.emplace_back(std::move(stab));
+  }
+
+  {
+    StabsEntry stab(MachO::N_FUN);
+    // FIXME this should be the size of the symbol. Using the section size in
+    // lieu is only correct if .subsections_via_symbols is set.
+    stab.value = defined->isec->getSize();
+    stabs.emplace_back(std::move(stab));
+  }
 }
 
 void SymtabSection::finalizeContents() {
-  // TODO support other symbol types
+  InputFile *lastFile = nullptr;
   for (Symbol *sym : symtab->getSymbols()) {
+    // TODO support other symbol types
     if (isa<Defined>(sym) || sym->isInGot() || sym->isInStubs()) {
       sym->symtabIndex = symbols.size();
       symbols.push_back({sym, stringTableSection.addString(sym->getName())});
     }
+
+    // Emit STABS symbols so that dsymutil and/or the debugger can map address
+    // regions in the final binary to the source and object files from which
+    // they originated.
+    if (auto *defined = dyn_cast<Defined>(sym)) {
+      if (defined->isAbsolute())
+        continue;
+
+      InputSection *isec = defined->isec;
+      // XXX is it right to assume that all symbols in __text are function
+      // symbols?
+      if (isec->name == "__text") {
+        ObjFile *file = dyn_cast<ObjFile>(isec->file);
+        assert(file);
+        if (!file->compileUnit)
+          continue;
+
+        if (lastFile == nullptr || lastFile != file) {
+          if (lastFile != nullptr)
+            emitEndSourceStab();
+          lastFile = file;
+
+          emitBeginSourceStab(file->compileUnit);
+          emitObjectFileStab(file);
+        }
+        emitFunStabs(defined);
       }
+      // TODO emit stabs for non-function symbols too
+    }
+  }
+
+  if (!stabs.empty())
+    emitEndSourceStab();
 }
 
 void SymtabSection::writeTo(uint8_t *buf) const {
   auto *nList = reinterpret_cast<structs::nlist_64 *>(buf);
   for (const SymtabEntry &entry : symbols) {
     nList->n_strx = entry.strx;
     // TODO support other symbol types
     // TODO populate n_desc with more flags
     if (auto *defined = dyn_cast<Defined>(entry.sym)) {
       if (defined->isAbsolute()) {
         nList->n_type = MachO::N_EXT | MachO::N_ABS;
         nList->n_sect = MachO::NO_SECT;
         nList->n_value = defined->value;
       } else {
         nList->n_type = MachO::N_EXT | MachO::N_SECT;
         nList->n_sect = defined->isec->parent->index;
         // For the N_SECT symbol type, n_value is the address of the symbol
-        nList->n_value = defined->value + defined->isec->getVA();
+        nList->n_value = defined->getVA();
       }
       nList->n_desc |= defined->isWeakDef() ? MachO::N_WEAK_DEF : 0;
     }
     ++nList;
   }
+
+  // Emit the stabs entries after the "real" symbols. We cannot emit them
+  // before as that would render Symbol::symtabIndex inaccurate.
+  for (const StabsEntry &entry : stabs) {
+    nList->n_strx = entry.strx;
+    nList->n_type = entry.type;
+    nList->n_sect = entry.sect;
+    nList->n_desc = entry.desc;
+    nList->n_value = entry.value;
+    ++nList;
+  }
 }
 
 IndirectSymtabSection::IndirectSymtabSection()
     : LinkEditSection(segment_names::linkEdit,
                       section_names::indirectSymbolTable) {}
 
 uint32_t IndirectSymtabSection::getNumSymbols() const {
   return in.got->getEntries().size() + in.tlvPointers->getEntries().size() +

lld/MachO/Writer.cpp

void Writer::createOutputSections() {
   default:
     llvm_unreachable("unhandled output file type");
   }
 
   // Then merge input sections into output sections.
   MapVector<std::pair<StringRef, StringRef>, MergedOutputSection *>
       mergedOutputSections;
   for (InputSection *isec : inputSections) {
+    // Instead of emitting DWARF sections, we emit STABS symbols to the object
+    // files that contain them.
+    if (isDebugSection(isec->flags) && isec->segname == segment_names::dwarf)
+      continue;
     MergedOutputSection *&osec =
         mergedOutputSections[{isec->segname, isec->name}];
     if (osec == nullptr)
       osec = make<MergedOutputSection>(isec->name);
     osec->mergeInput(isec);
   }
 
   for (const auto &it : mergedOutputSections) {
     StringRef segname = it.first.first;
     MergedOutputSection *osec = it.second;
     if (unwindInfoSection && segname == segment_names::ld) {
       assert(osec->name == section_names::compactUnwind);
       unwindInfoSection->setCompactUnwindSection(osec);
-    } else
+    } else {
       getOrCreateOutputSegment(segname)->addOutputSection(osec);
     }
+  }
 
   for (SyntheticSection *ssec : syntheticSections) {
     auto it = mergedOutputSections.find({ssec->segname, ssec->name});
     if (it == mergedOutputSections.end()) {
       if (ssec->isNeeded())
         getOrCreateOutputSegment(ssec->segname)->addOutputSection(ssec);
     } else {
       error("section from " + it->second->firstSection()->file->getName() +

lld/test/MachO/stabs.s

+# REQUIRES: x86
+# RUN: split-file %s %t
+# RUN: llvm-mc -filetype=obj -triple=x86_64-apple-darwin %t/test.s -o %t/test.o
+# RUN: llvm-mc -filetype=obj -triple=x86_64-apple-darwin %t/foo.s -o %t/foo.o
+
+# RUN: %lld -lSystem %t/test.o %t/foo.o -o %t/test
+# RUN: llvm-nm -pa %t/test | FileCheck %s -DDIR=%t
+
+## Check that we emit absolute paths to the object files in our OSO entries
+## even if our inputs are relative paths.
+# RUN: cd %t && %lld -lSystem test.o foo.o -o test
+# RUN: llvm-nm -pa %t/test | FileCheck %s -DDIR=%t
+
+# CHECK-DAG:  [[#%x, MAIN:]]   T _main
+# CHECK-DAG:  [[#%x, FOO: ]]   T _foo
+# CHECK:      0000000000000000 - 00 0000    SO /tmp/test.cpp
+# CHECK-NEXT: 0000000000000000 - 03 0001   OSO [[DIR]]/test.o
+# CHECK-NEXT: [[#MAIN]]        - 01 0000   FUN _main
+# CHECK-NEXT: 0000000000000001 - 00 0000   FUN
+# CHECK-NEXT: 0000000000000000 - 01 0000    SO
+# CHECK-NEXT: 0000000000000000 - 00 0000    SO /foo.cpp
+# CHECK-NEXT: 0000000000000000 - 03 0001   OSO [[DIR]]/foo.o
+# CHECK-NEXT: [[#FOO]]         - 01 0000   FUN _foo
+# CHECK-NEXT: 0000000000000001 - 00 0000   FUN
+# CHECK-NEXT: 0000000000000000 - 01 0000    SO
+
+#--- test.s
+.text
+.globl  _main
+_main:
+Lfunc_begin0:
+  retq
+Lfunc_end0:
+
+.section  __DWARF,__debug_str,regular,debug
+  .asciz  "test.cpp"             ## string offset=0
+  .asciz  "/tmp"                 ## string offset=9
+.section  __DWARF,__debug_abbrev,regular,debug
+Lsection_abbrev:
+  .byte  1                       ## Abbreviation Code
+  .byte  17                      ## DW_TAG_compile_unit
+  .byte  1                       ## DW_CHILDREN_yes
+  .byte  3                       ## DW_AT_name
+  .byte  14                      ## DW_FORM_strp
+  .byte  27                      ## DW_AT_comp_dir
+  .byte  14                      ## DW_FORM_strp
+  .byte  17                      ## DW_AT_low_pc
+  .byte  1                       ## DW_FORM_addr
+  .byte  18                      ## DW_AT_high_pc
+  .byte  6                       ## DW_FORM_data4
+  .byte  0                       ## EOM(1)
+.section  __DWARF,__debug_info,regular,debug
+.set Lset0, Ldebug_info_end0-Ldebug_info_start0 ## Length of Unit
+  .long  Lset0
+Ldebug_info_start0:
+  .short  4                       ## DWARF version number
+.set Lset1, Lsection_abbrev-Lsection_abbrev ## Offset Into Abbrev. Section
+  .long  Lset1
+  .byte  8                       ## Address Size (in bytes)
+  .byte  1                       ## Abbrev [1] 0xb:0x48 DW_TAG_compile_unit
+  .long  0                       ## DW_AT_name
+  .long  9                       ## DW_AT_comp_dir
+  .quad  Lfunc_begin0            ## DW_AT_low_pc
+.set Lset3, Lfunc_end0-Lfunc_begin0     ## DW_AT_high_pc
+  .long  Lset3
+  .byte  0                       ## End Of Children Mark
+Ldebug_info_end0:
+.subsections_via_symbols
+.section  __DWARF,__debug_line,regular,debug
+
+#--- foo.s
+.text
+.globl  _foo
+_foo:
+Lfunc_begin0:
+  retq
+Lfunc_end0:
+
+.section  __DWARF,__debug_str,regular,debug
+  .asciz  "foo.cpp"              ## string offset=0
+  .asciz  ""                     ## string offset=8
+.section  __DWARF,__debug_abbrev,regular,debug
+Lsection_abbrev:
+  .byte  1                       ## Abbreviation Code
+  .byte  17                      ## DW_TAG_compile_unit
+  .byte  1                       ## DW_CHILDREN_yes
+  .byte  3                       ## DW_AT_name
+  .byte  14                      ## DW_FORM_strp
+  .byte  27                      ## DW_AT_comp_dir
+  .byte  14                      ## DW_FORM_strp
+  .byte  17                      ## DW_AT_low_pc
+  .byte  1                       ## DW_FORM_addr
+  .byte  18                      ## DW_AT_high_pc
+  .byte  6                       ## DW_FORM_data4
+  .byte  0                       ## EOM(1)
+.section  __DWARF,__debug_info,regular,debug
+.set Lset0, Ldebug_info_end0-Ldebug_info_start0 ## Length of Unit
+  .long  Lset0
+Ldebug_info_start0:
+  .short  4                       ## DWARF version number
+.set Lset1, Lsection_abbrev-Lsection_abbrev ## Offset Into Abbrev. Section
+  .long  Lset1
+  .byte  8                       ## Address Size (in bytes)
+  .byte  1                       ## Abbrev [1] 0xb:0x48 DW_TAG_compile_unit
+  .long  0                       ## DW_AT_name
+  .long  8                       ## DW_AT_comp_dir
+  .quad  Lfunc_begin0            ## DW_AT_low_pc
+.set Lset3, Lfunc_end0-Lfunc_begin0     ## DW_AT_high_pc
+  .long  Lset3
+  .byte  0                       ## End Of Children Mark
+Ldebug_info_end0:
+.subsections_via_symbols
+.section  __DWARF,__debug_line,regular,debug