[lld-macho] Explicitly undefine literal exported symbols

[gkm]

Summary

Symbols explicitly exported via command-line options --exported_symbol SYM and --exported_symbols_list FILE must be defined. Before this fix, lazy symbols defined in archives would be left to languish. We now force them to be included in the linked output.

[int3]

Thanks!

lld/test/MachO/export-options.s

3

to make sure we aren't adding to the same archive multiple times

45–72

I know this is existing code, but it seems kind of odd to me that the CHECK / RUN lines are below instead of above the test inputs

[gkm]

• revise according to review feedback, notably: reorg lld/test/MachO/export-options.s to move all split input files at the end.

lld/MachO/Driver.cpp

handleSymbolPatterns(args, config->exportedSymbols, OPT_exported_symbol,
                        OPT_exported_symbols_list);
   handleSymbolPatterns(args, config->unexportedSymbols, OPT_unexported_symbol,
                        OPT_unexported_symbols_list);
   if (!config->exportedSymbols.empty() && !config->unexportedSymbols.empty()) {
     error("cannot use both -exported_symbol* and -unexported_symbol* options\n"
           ">>> ignoring unexports");
     config->unexportedSymbols.clear();
   }
+  // Explicitly-exported literal symbols must be defined, but might
+  // languish in an archive if unreferenced elsewhere. Light a fire
+  // under those lazy symbols!
+  for (const CachedHashStringRef &cachedName : config->exportedSymbols.literals)
+    symtab->addUndefined(cachedName.val(), /*file=*/nullptr,
+                         /*isWeakRef=*/false);
 
   config->saveTemps = args.hasArg(OPT_save_temps);
 
   config->adhocCodesign = args.hasFlag(
       OPT_adhoc_codesign, OPT_no_adhoc_codesign,
       (config->arch() == AK_arm64 || config->arch() == AK_arm64e) &&
           config->platform() == PlatformKind::macOS);
 
if (config->timeTraceEnabled)
     }
     // Literal exported-symbol names must be defined, but glob
     // patterns need not match.
     for (const CachedHashStringRef &cachedName :
          config->exportedSymbols.literals) {
       if (const Symbol *sym = symtab->find(cachedName))
         if (isa<Defined>(sym))
           continue;
-      error("undefined symbol " + cachedName.val() +
+      error("undefined symbol: " + cachedName.val() +
             "\n>>> referenced from option -exported_symbol(s_list)");
     }
 
     createSyntheticSections();
     createSyntheticSymbols();
 
     for (const Arg *arg : args.filtered(OPT_sectcreate)) {
       StringRef segName = arg->getValue(0);

lld/test/MachO/export-options.s

 # REQUIRES: x86
 
-# RUN: split-file %s %t
+# RUN: rm -rf %t; split-file %s %t
 # RUN: llvm-mc -filetype=obj -triple=x86_64-apple-macos %t/default.s -o %t/default.o
+# RUN: llvm-mc -filetype=obj -triple=x86_64-apple-macos %t/lazydef.s -o %t/lazydef.o
+# RUN: llvm-ar --format=darwin rcs %t/lazydef.a %t/lazydef.o
 
 ## Check that mixing exported and unexported symbol options yields an error
 # RUN: not %lld -dylib %t/default.o -o /dev/null \
 # RUN:         -exported_symbol a -unexported_symbol b 2>&1 | \
 # RUN:     FileCheck --check-prefix=CONFLICT %s
 
 # CONFLICT: error: cannot use both -exported_symbol* and -unexported_symbol* options
 # CONFLICT-NEXT: >>> ignoring unexports
 
-## Check that exported literal symbol name is present in symbol table
+## Check that an exported literal name with no symbol definition yields an error
+## but that an exported glob-pattern with no matching symbol definition is OK
 # RUN: not %lld -dylib %t/default.o -o /dev/null \
 # RUN:         -exported_symbol absent_literal \
 # RUN:         -exported_symbol absent_gl?b 2>&1 | \
 # RUN:     FileCheck --check-prefix=UNDEF %s
 
-# UNDEF: error: undefined symbol absent_literal
+# UNDEF: error: undefined symbol: absent_literal
 # UNDEF-NEXT: >>> referenced from option -exported_symbol(s_list)
 # UNDEF-NOT: error: {{.*}} absent_gl{{.}}b
 
+## Check that exported literal symbols are present in output's
+## symbol table, even lazy symbols which would otherwise be omitted
+# RUN: %lld -dylib %t/default.o %t/lazydef.a -o %t/lazydef \
+# RUN:         -exported_symbol _keep_globl \
+# RUN:         -exported_symbol _keep_lazy
+# RUN: llvm-objdump --syms %t/lazydef | \
+# RUN:     FileCheck --check-prefix=EXPORT %s
+
+# EXPORT-DAG: g     F __TEXT,__text _keep_globl
+# EXPORT-DAG: g     F __TEXT,__text _keep_lazy
+
 ## Check that exported symbol is global
 # RUN: not %lld -dylib %t/default.o -o /dev/null \
 # RUN:         -exported_symbol _private_extern 2>&1 | \
 # RUN:     FileCheck --check-prefix=PRIVATE %s
 
 # PRIVATE: error: cannot export hidden symbol _private_extern
 
-#--- default.s
-
-.globl _keep_globl, _hide_globl
-_keep_globl:
-  retq
-_hide_globl:
-  retq
-.private_extern _private_extern
-_private_extern:
-  retq
-_private:
-  retq
-
 ## Check that the export trie is unaltered
 # RUN: %lld -dylib %t/default.o -o %t/default
 # RUN: llvm-objdump --macho --exports-trie %t/default | \
 # RUN:     FileCheck --check-prefix=DEFAULT %s
 
 # DEFAULT-LABEL: Exports trie:
 # DEFAULT-DAG:   _hide_globl
 # DEFAULT-DAG:   _keep_globl
 # DEFAULT-NOT:   _private_extern
 
 ## Check that the export trie is shaped by an allow list and then
 ## by a deny list. Both lists are designed to yield the same result.
 
 ## Check the allow list
 # RUN: %lld -dylib %t/default.o -o %t/allowed \
 # RUN:     -exported_symbol _keep_globl
 # RUN: llvm-objdump --macho --exports-trie %t/allowed | \
 # RUN:     FileCheck --check-prefix=TRIE %s
 # RUN: llvm-nm -m %t/allowed | \
 # RUN:     FileCheck --check-prefix=NM %s
 
 ## Check the deny list
 # RUN: %lld -dylib %t/default.o -o %t/denied \
 # RUN:     -unexported_symbol _hide_globl
 # RUN: llvm-objdump --macho --exports-trie %t/denied | \
 # RUN:     FileCheck --check-prefix=TRIE %s
 # RUN: llvm-nm -m %t/denied | \
 # RUN:     FileCheck --check-prefix=NM %s
 
 # TRIE-LABEL: Exports trie:
 # TRIE-DAG:   _keep_globl
 # TRIE-NOT:   _hide_globl
 # TRIE-NOT:   _private_extern
 
 # NM-DAG: external _keep_globl
 # NM-DAG: non-external (was a private external) _hide_globl
 # NM-DAG: non-external (was a private external) _private_extern
 
 # RUN: llvm-mc -filetype=obj -triple=x86_64-apple-macos \
 # RUN:     %t/symdefs.s -o %t/symdefs.o
 
+## Check that only string-literal patterns match
+## Check that comments and blank lines are stripped from symbol list
+# RUN: %lld -dylib %t/symdefs.o -o %t/literal \
+# RUN:     -exported_symbols_list %t/literals.txt
+# RUN: llvm-objdump --macho --exports-trie %t/literal | \
+# RUN:     FileCheck --check-prefix=LITERAL %s
+
+# LITERAL-DAG: literal_only
+# LITERAL-DAG: literal_also
+# LITERAL-DAG: globby_also
+# LITERAL-NOT: globby_only
+
+## Check that only glob patterns match
+## Check that comments and blank lines are stripped from symbol list
+# RUN: %lld -dylib %t/symdefs.o -o %t/globby \
+# RUN:     -exported_symbols_list %t/globbys.txt
+# RUN: llvm-objdump --macho --exports-trie %t/globby | \
+# RUN:     FileCheck --check-prefix=GLOBBY %s
+
+# GLOBBY-DAG: literal_also
+# GLOBBY-DAG: globby_only
+# GLOBBY-DAG: globby_also
+# GLOBBY-NOT: literal_only
+
+#--- default.s
+
+.globl _keep_globl, _hide_globl
+_keep_globl:
+  retq
+_hide_globl:
+  retq
+.private_extern _private_extern
+_private_extern:
+  retq
+_private:
+  retq
+
+#--- lazydef.s
+
+.globl _keep_lazy
+_keep_lazy:
+  retq
+
 #--- symdefs.s
 
 .globl literal_only, literal_also, globby_only, globby_also
 literal_only:
   retq
 literal_also:
   retq
 globby_only:
   retq
 globby_also:
   retq
 
 #--- literals.txt
 
   literal_only # comment
   literal_also
 
 # globby_only
   globby_also
 
-## Check that only string-literal patterns match
-## Check that comments and blank lines are stripped from symbol list
-# RUN: %lld -dylib %t/symdefs.o -o %t/literal \
-# RUN:     -exported_symbols_list %t/literals.txt
-# RUN: llvm-objdump --macho --exports-trie %t/literal | \
-# RUN:     FileCheck --check-prefix=LITERAL %s
-
-# LITERAL-DAG: literal_only
-# LITERAL-DAG: literal_also
-# LITERAL-DAG: globby_also
-# LITERAL-NOT: globby_only
-
 #--- globbys.txt
 
 # literal_only
   l?ter[aeiou]l_*[^y] # comment
 
   *gl?bby_*
-
-## Check that only glob patterns match
-## Check that comments and blank lines are stripped from symbol list
-# RUN: %lld -dylib %t/symdefs.o -o %t/globby \
-# RUN:     -exported_symbols_list %t/globbys.txt
-# RUN: llvm-objdump --macho --exports-trie %t/globby | \
-# RUN:     FileCheck --check-prefix=GLOBBY %s
-
-# GLOBBY-DAG: literal_also
-# GLOBBY-DAG: globby_only
-# GLOBBY-DAG: globby_also
-# GLOBBY-NOT: literal_only