[lld-macho] Stricter Bitcode Symbol Resolution

[kyulee]

Summary

LLD resolves symbols before performing LTO compilation, assuming that the symbols in question are resolved by the resulting object files from LTO. However, there is a scenario where the prevailing symbols might be resolved incorrectly due to specific assembly symbols not appearing in the symbol table of the bitcode. This patch deals with such a scenario by generating an error instead of silently allowing a mis-linkage.
If a prevailing symbol is resolved through post-loaded archives via LC linker options, a warning will now be issued.

[kyulee]

Update format

[kyulee]

Update tests.

[thevinster]

Overall, lgtm. Minor comments

lld/MachO/SymbolTable.cpp

155

Can we also add a condition to cast this file as a bitcode file just make sure that it is in fact a bitcode file?

156–166

So if I understand the comment here correctly, A has a module asm symbol that is defined in B, but incorrectly resolves it to the bitcode file instead of the object file? This dependency issue kinda seems like a bug rather than a limitation and I wonder if this scenario should be supported. Though using module asm in source is already sketchy, in itself, so it might be worth error-ing to as a forcing function to make users fix their errors at the right level.

[kyulee]

Incorporate the comment.

[kyulee]

lld/MachO/SymbolTable.cpp

156–166

This is the case where we error out where we can't resolve the prevailing symbol correctly.
I think blocking module asm itself is an overkill given there exists user code that already has assembly.
I think this is a peculiar case where module asm is used to hold a reference that is even not defined in this compilation unit.

[kyulee]

Fix the rebase.

lld/MachO/InputFiles.h

struct FDE {
   Symbol *personality;
   InputSection *lsda;
 };
 
 // .o file
 class ObjFile final : public InputFile {
 public:
   ObjFile(MemoryBufferRef mb, uint32_t modTime, StringRef archiveName,
-          bool lazy = false, bool forceHidden = false, bool compatArch = true);
+          bool lazy = false, bool forceHidden = false, bool compatArch = true,
+          bool builtFromBitcode = false);
   ArrayRef<llvm::MachO::data_in_code_entry> getDataInCode() const;
   ArrayRef<uint8_t> getOptimizationHints() const;
   template <class LP> void parse();
   template <class LP>
   void parseLinkerOptions(llvm::SmallVectorImpl<StringRef> &LinkerOptions);
 
   static bool classof(const InputFile *f) { return f->kind() == ObjKind; }
 
   std::string sourceFile() const;
   // Parses line table information for diagnostics. compileUnit should be used
   // for other purposes.
   lld::DWARFCache *getDwarf();
 
   llvm::DWARFUnit *compileUnit = nullptr;
   std::unique_ptr<lld::DWARFCache> dwarfCache;
   Section *addrSigSection = nullptr;
   const uint32_t modTime;
   bool forceHidden;
+  bool builtFromBitcode;
   std::vector<ConcatInputSection *> debugSections;
   std::vector<CallGraphEntry> callGraph;
   llvm::DenseMap<ConcatInputSection *, FDE> fdes;
   std::vector<AliasSymbol *> aliases;
 
 private:
   llvm::once_flag initDwarf;
   template <class LP> void parseLazy();

lld/MachO/InputFiles.cpp

for (auto *cmd : findCommands<linker_option_command>(hdr, LC_LINKER_OPTION)) {
     StringRef data{reinterpret_cast<const char *>(cmd + 1),
                    cmd->cmdsize - sizeof(linker_option_command)};
     parseLCLinkerOption(LCLinkerOptions, this, cmd->count, data);
   }
 }
 
 SmallVector<StringRef> macho::unprocessedLCLinkerOptions;
 ObjFile::ObjFile(MemoryBufferRef mb, uint32_t modTime, StringRef archiveName,
-                 bool lazy, bool forceHidden, bool compatArch)
-    : InputFile(ObjKind, mb, lazy), modTime(modTime), forceHidden(forceHidden) {
+                 bool lazy, bool forceHidden, bool compatArch,
+                 bool builtFromBitcode)
+    : InputFile(ObjKind, mb, lazy), modTime(modTime), forceHidden(forceHidden),
+      builtFromBitcode(builtFromBitcode) {
   this->archiveName = std::string(archiveName);
   this->compatArch = compatArch;
   if (lazy) {
     if (target->wordSize == 8)
       parseLazy<LP64>();
     else
       parseLazy<ILP32>();
   } else {

lld/MachO/LTO.cpp

for (unsigned i = 0; i < maxTasks; ++i) {
 
     auto filePath = outputFilePath(i);
     uint32_t modTime = 0;
     if (!config->ltoObjPath.empty()) {
       saveOrHardlinkBuffer(objBuf, filePath, cachePath);
       modTime = getModTime(filePath);
     }
     ret.push_back(make<ObjFile>(
-        MemoryBufferRef(objBuf, saver().save(filePath.str())), modTime, ""));
+        MemoryBufferRef(objBuf, saver().save(filePath.str())), modTime,
+        /*archiveName=*/"", /*lazy=*/false,
+        /*forceHidden=*/false, /*compatArch=*/true, /*builtFromBitcode=*/true));
   }
 
   return ret;
 }

lld/MachO/SymbolTable.cpp

if (auto *defined = dyn_cast<Defined>(s)) {
         dupSymDiags.push_back({make_pair(srcLoc1, srcFile1),
                                make_pair(srcLoc2, srcFile2), defined});
       }
 
     } else if (auto *dysym = dyn_cast<DylibSymbol>(s)) {
       overridesWeakDef = !isWeakDef && dysym->isWeakDef();
       dysym->unreference();
     } else if (auto *undef = dyn_cast<Undefined>(s)) {
-      // Preserve the original bitcode file name (instead of using the object
-      // file name).
-      if (undef->wasBitcodeSymbol)
+      if (undef->wasBitcodeSymbol) {
+        auto objFile = dyn_cast<ObjFile>(file);
+        if (!objFile) {
+          // The file must be a native object file, as opposed to potentially
+          // being another bitcode file. A situation arises when some symbols
+          // are defined thru `module asm` and thus they are not present in the
+          // bitcode's symbol table. Consider bitcode modules `A`, `B`, and `C`.
+          // LTO compiles only `A` and `C`, since there's no explicit symbol
+          // reference to `B` other than a symbol from `A` via `module asm`.
+          // After LTO is finished, the missing symbol now appears in the
+          // resulting object file for `A`, which  prematurely resolves another
+          // prevailing symbol with `B` that hasn't been compiled, instead of
+          // the resulting object for `C`. Consequently, an incorrect
+          // relocation is generated for the prevailing symbol.
+          assert(isa<BitcodeFile>(file) && "Bitcode file is expected.");
+          std::string message =
+              "The pending prevailing symbol(" + name.str() +
+              ") in the bitcode file(" + toString(undef->getFile()) +
+              ") is overridden by a non-native object (from bitcode): " +
+              toString(file);
+          error(message);
+        } else if (!objFile->builtFromBitcode) {
+          // Ideally, this should be an object file compiled from a bitcode
+          // file. However, this might not hold true if a LC linker option is
+          // used. In case LTO internalizes a prevailing hidden weak symbol,
+          // there's a situation where an unresolved prevailing symbol might be
+          // linked with the corresponding one from a native library, which is
+          // loaded later after LTO. Although this could potentially result in
+          // an ODR violation, we choose to permit this scenario as a warning.
+          std::string message = "The pending prevailing symbol(" + name.str() +
+                                ") in the bitcode file(" +
+                                toString(undef->getFile()) +
+                                ") is overridden by a post-processed native "
+                                "object (from native archive): " +
+                                toString(file);
+          warn(message);
+        } else {
+          // Preserve the original bitcode file name (instead of using the
+          // object file name).
           file = undef->getFile();
         }
+      }
+    }
     // Defined symbols take priority over other types of symbols, so in case
     // of a name conflict, we fall through to the replaceSymbol() call below.
   }
 
   // With -flat_namespace, all extern symbols in dylibs are interposable.
   // FIXME: Add support for `-interposable` (PR53680).
   bool interposable = config->namespaceKind == NamespaceKind::flat &&
                       config->outputType != MachO::MH_EXECUTE &&

lld/test/MachO/symbol-resolution-error.ll

+; REQUIRES: x86
+; RUN: rm -rf %t; split-file %s %t
+
+; RUN: opt --thinlto-bc %t/f.ll -o %t/f.o
+; RUN: llvm-ar rcs %t/libf.a %t/f.o
+; RUN: opt --thinlto-bc %t/q.ll -o %t/q.o
+; RUN: llvm-ar rcs %t/libq.a %t/q.o
+; RUN: opt --thinlto-bc %t/b.ll -o %t/b.o
+; RUN: llvm-ar rcs %t/libb.a %t/b.o
+; RUN: opt --thinlto-bc %t/m.ll -o %t/m.o
+
+; RUN: not %no-fatal-warnings-lld -lSystem %t/libf.a %t/libq.a %t/libb.a %t/m.o -o %t/test.out 2>&1 | FileCheck %s
+
+; We can't read symbols that are set by `module asm` in the bitcode's symbol table.
+; LTO compiles only f.ll, b.ll, and m.ll without q.ll due to missing reference to `q_asm` in f.ll.
+; However, later `q_asm` appears in the resulting object for `f.ll`, which triggers resolving
+; a ODR symbol `q_odr` from `q.ll,` as opposed to one expected from `b.ll`.
+; As a result, we have invalid relocation for `q_odr` as `q.ll` is neither compiled, nor linked.
+
+; CHECK: error: The pending prevailing symbol(_q_odr) in the bitcode file({{.*}}libb.a(b.o)) is overridden by a non-native object (from bitcode): {{.*}}libq.a(q.o)
+
+;--- f.ll
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+module asm ".no_dead_strip _q_asm"
+
+define i64 @f() {
+  ret i64 1
+}
+
+;--- q.ll
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+@q_asm = global i64 1
+
+@q_odr = linkonce_odr hidden unnamed_addr constant i64 0
+
+
+;--- b.ll
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+@q_odr = linkonce_odr hidden unnamed_addr constant i64 1
+
+define i64 @b() {
+  ret i64 ptrtoint (i64* @q_odr to i64)
+}
+
+;--- m.ll
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+declare i64 @f()
+declare i64 @b()
+
+define i64 @main() {
+  %call1 = call i64 @f()
+  %call2 = call i64 @b()
+  %add = add nsw i64 %call1, %call2
+  ret i64 %add
+}

lld/test/MachO/symbol-resolution-warning.ll

+; REQUIRES: x86
+; RUN: rm -rf %t; split-file %s %t
+
+; RUN: opt %t/f.ll -o %t/f.o
+; RUN: llvm-ar rcs %t/libf.a %t/f.o
+; RUN: llc -filetype=obj %t/q.ll -o %t/q.o
+; RUN: llvm-ar rcs %t/libq.a %t/q.o
+; RUN: llc -filetype=obj %t/m.ll -o %t/m.o
+
+; RUN: %no-fatal-warnings-lld -dylib -lSystem -L%t %t/libf.a %t/m.o -o %t/test.out 2>&1 | FileCheck %s
+
+; We can't read symbols that are set by module asm in the bitcode's symbol table.
+; LTO internalizes `odr` in `f.ll`, and thus the prevailing `odr` remains unresolved.
+; `q.ll` is loaded after LTO is finished during post-processing LC linker options.
+; The unresolved prevailing `odr` is now resolved with `q.ll`.
+
+; CHECK: warning: The pending prevailing symbol(_odr) in the bitcode file({{.*}}libf.a(f.o)) is overridden by a post-processed native object (from native archive): {{.*}}libq.a(q.o)
+
+;--- f.ll
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+!0 = !{!"-lq"}
+!llvm.linker.options = !{!0}
+
+define i64 @f()  {
+  %1 = call i64 @odr()
+  %2 = call i64 @q()
+  %3 = add i64 %1, %2
+  ret i64 %3
+}
+
+define weak hidden i64 @odr() noinline {
+  ret i64 1
+}
+
+declare i64 @q()
+
+;--- q.ll
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+define i64 @q() {
+  %1 = call i64 @odr()
+  ret i64 %1
+}
+
+define linkonce_odr hidden i64 @odr() noinline {
+  ret i64 2
+}
+
+;--- m.ll
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.15.0"
+
+declare i64 @f()
+
+define i64 @m() {
+  %1 = call i64 @f()
+  ret i64 %1
+}