[lld-macho] Coalesce local symbol aliases along with their aliased we…

…ak def

This supersedes {D139069}. In some ways we are now closer to ld64's
behavior: we previously only did this coalescing for private-label
symbols, but now we do it for all locals, just like ld64. However, we no
longer generate weak binds when a local alias to a weak symbol is
referenced. This is merely for implementation simplicity; it's not clear
to me that any real-world programs depend on us emulating this behavior.

The problem with the previous approach is that we ended up with
duplicate references to the same symbol instance in our InputFiles,
which translated into duplicate symbols in our output. While we could
work around that problem by performing a dedup step before emitting the
symbol table, it seems cleaner to not generate duplicate references in
the first place.

Numbers for chromium_framework on my 16 Core Intel Mac Pro:

             base           diff           difference (95% CI)
  sys_time   2.243 ± 0.093  2.231 ± 0.066  [  -2.5% ..   +1.4%]
  user_time  6.529 ± 0.087  6.080 ± 0.050  [  -7.5% ..   -6.3%]
  wall_time  6.928 ± 0.175  6.474 ± 0.112  [  -7.7% ..   -5.4%]
  samples    26             31

Yep, that's a massive win... because it turns out that {D140606} and
{D139069} caused a regression (of about the same size.) I just didn't
think to measure them back then. I'm guessing all the extra symbols we
have been emitting did not help perf at all...

Reviewed By: lgrey

Differential Revision: https://reviews.llvm.org/D145455

lld/MachO/InputFiles.cpp

@@ -849,15 +849,12 @@ void ObjFile::parseSymbols(ArrayRef<typename LP::section> sectionHeaders,
     // We populate subsections by repeatedly splitting the last (highest
     // address) subsection.
     llvm::stable_sort(symbolIndices, [&](uint32_t lhs, uint32_t rhs) {
-      // Put private-label symbols that have no flags after other symbols at the
-      // same address.
-      StringRef lhsName = getSymName(nList[lhs]);
-      StringRef rhsName = getSymName(nList[rhs]);
-      if (nList[lhs].n_value == nList[rhs].n_value) {
-        if (isPrivateLabel(lhsName) && isPrivateLabel(rhsName))
-          return nList[lhs].n_desc > nList[rhs].n_desc;
-        return !isPrivateLabel(lhsName) && isPrivateLabel(rhsName);
-      }
+      // Put extern weak symbols after other symbols at the same address so
+      // that weak symbol coalescing works correctly. See
+      // SymbolTable::addDefined() for details.
+      if (nList[lhs].n_value == nList[rhs].n_value &&
+          nList[lhs].n_type & N_EXT && nList[rhs].n_type & N_EXT)
+        return !(nList[lhs].n_desc & N_WEAK_DEF) && (nList[rhs].n_desc & N_WEAK_DEF);
       return nList[lhs].n_value < nList[rhs].n_value;
     });
     for (size_t j = 0; j < symbolIndices.size(); ++j) {
@@ -883,17 +880,8 @@ void ObjFile::parseSymbols(ArrayRef<typename LP::section> sectionHeaders,
       if (!subsectionsViaSymbols || symbolOffset == 0 ||
           sym.n_desc & N_ALT_ENTRY || !isa<ConcatInputSection>(isec)) {
         isec->hasAltEntry = symbolOffset != 0;
-        // If we have an private-label symbol that's an alias, and that alias
-        // doesn't have any flags of its own, then we can just reuse the aliased
-        // symbol. Our sorting step above ensures that any such symbols will
-        // appear after the non-private-label ones. See weak-def-alias-ignored.s
-        // for the motivation behind this.
-        if (symbolOffset == 0 && isPrivateLabel(name) && j != 0 &&
-            sym.n_desc == 0)
-          symbols[symIndex] = symbols[symbolIndices[j - 1]];
-        else
-          symbols[symIndex] = createDefined(sym, name, isec, symbolOffset,
-                                            symbolSize, forceHidden);
+        symbols[symIndex] = createDefined(sym, name, isec, symbolOffset,
+                                          symbolSize, forceHidden);
         continue;
       }
       auto *concatIsec = cast<ConcatInputSection>(isec);

lld/MachO/SymbolTable.cpp

@@ -61,6 +61,38 @@ struct DuplicateSymbolDiag {
 SmallVector<DuplicateSymbolDiag> dupSymDiags;
 } // namespace
 
+// Move symbols at \p fromOff in \p fromIsec into \p toIsec, unless that symbol
+// is \p skip.
+static void transplantSymbolsAtOffset(InputSection *fromIsec,
+                                      InputSection *toIsec, Defined *skip,
+                                      uint64_t fromOff, uint64_t toOff) {
+  // Ensure the symbols will still be in address order after our insertions.
+  auto insertIt = llvm::upper_bound(toIsec->symbols, toOff,
+                                    [](uint64_t off, const Symbol *s) {
+                                      return cast<Defined>(s)->value < off;
+                                    });
+  llvm::erase_if(fromIsec->symbols, [&](Symbol *s) {
+    auto *d = cast<Defined>(s);
+    if (d->value != fromOff)
+      return false;
+    if (d != skip) {
+      // This repeated insertion will be quadratic unless insertIt is the end
+      // iterator. However, that is typically the case for files that have
+      // .subsections_via_symbols set.
+      insertIt = toIsec->symbols.insert(insertIt, d);
+      d->isec = toIsec;
+      d->value = toOff;
+      // We don't want to have more than one unwindEntry at a given address, so
+      // drop the redundant ones. We We can safely drop the unwindEntries of
+      // the symbols in fromIsec since we will be adding another unwindEntry as
+      // we finish parsing toIsec's file. (We can assume that toIsec has its
+      // own unwindEntry because of the ODR.)
+      d->unwindEntry = nullptr;
+    }
+    return true;
+  });
+}
+
 Defined *SymbolTable::addDefined(StringRef name, InputFile *file,
                                  InputSection *isec, uint64_t value,
                                  uint64_t size, bool isWeakDef,
@@ -82,18 +114,27 @@ Defined *SymbolTable::addDefined(StringRef name, InputFile *file,
           defined->referencedDynamically |= isReferencedDynamically;
           defined->noDeadStrip |= noDeadStrip;
         }
-        // FIXME: Handle this for bitcode files.
-        if (auto concatIsec = dyn_cast_or_null<ConcatInputSection>(isec))
+        if (auto concatIsec = dyn_cast_or_null<ConcatInputSection>(isec)) {
           concatIsec->wasCoalesced = true;
+          // Any local symbols that alias the coalesced symbol should be moved
+          // into the prevailing section. Note that we have sorted the symbols
+          // in ObjFile::parseSymbols() such that extern weak symbols appear
+          // last, so we don't need to worry about subsequent symbols being
+          // added to an already-coalesced section.
+          if (defined->isec)
+            transplantSymbolsAtOffset(concatIsec, defined->isec,
+                                      /*skip=*/nullptr, value, defined->value);
+        }
         return defined;
       }
 
       if (defined->isWeakDef()) {
-        // FIXME: Handle this for bitcode files.
         if (auto concatIsec =
                 dyn_cast_or_null<ConcatInputSection>(defined->isec)) {
           concatIsec->wasCoalesced = true;
-          concatIsec->symbols.erase(llvm::find(concatIsec->symbols, defined));
+          if (isec)
+            transplantSymbolsAtOffset(concatIsec, isec, defined, defined->value,
+                                      value);
         }
       } else {
         std::string srcLoc1 = defined->getSourceLocation();
@@ -382,7 +423,7 @@ struct UndefinedDiag {
 };
 
 MapVector<const Undefined *, UndefinedDiag> undefs;
-}
+} // namespace
 
 void macho::reportPendingDuplicateSymbols() {
   for (const auto &duplicate : dupSymDiags) {

lld/test/MachO/local-alias-to-weak.s

@@ -0,0 +1,149 @@
+# REQUIRES: x86
+## This test checks that when we coalesce weak definitions, their local symbol
+## aliases defs don't cause the coalesced data to be retained. This was
+## motivated by MC's aarch64 backend which automatically creates `ltmp<N>`
+## symbols at the start of each .text section. These symbols are frequently
+## aliases of other symbols created by clang or other inputs to MC. I've chosen
+## to explicitly create them here since we can then reference those symbols for
+## a more complete test.
+##
+## Not retaining the data matters for more than just size -- we have a use case
+## that depends on proper data coalescing to emit a valid file format. We also
+## need this behavior to properly deduplicate the __objc_protolist section;
+## failure to do this can result in dyld crashing on iOS 13.
+##
+## Finally, ld64 does all this regardless of whether .subsections_via_symbols is
+## specified. We don't. But again, given how rare the lack of that directive is
+## (I've only seen it from hand-written assembly inputs), I don't think we need
+## to worry about it.
+
+# RUN: rm -rf %t; split-file %s %t
+# RUN: llvm-mc -filetype=obj -triple=x86_64-apple-darwin %t/weak-then-local.s -o %t/weak-then-local.o
+# RUN: llvm-mc -filetype=obj -triple=x86_64-apple-darwin %t/local-then-weak.s -o %t/local-then-weak.o
+# RUN: llvm-mc -filetype=obj -triple=x86_64-apple-darwin %t/no-subsections.s -o %t/no-subsections.o
+# RUN: llvm-mc -filetype=obj -triple=x86_64-apple-darwin %t/no-dead-strip.s -o %t/no-dead-strip.o
+
+# RUN: %lld -lSystem -dylib %t/weak-then-local.o %t/local-then-weak.o -o %t/test1
+# RUN: llvm-objdump --macho --syms --section="__DATA,__data" --weak-bind %t/test1 | FileCheck %s
+# RUN: %lld -lSystem -dylib %t/local-then-weak.o %t/weak-then-local.o -o %t/test2
+# RUN: llvm-objdump --macho --syms --section="__DATA,__data" --weak-bind %t/test2 | FileCheck %s
+
+## Check that we only have one copy of 0x123 in the data, not two.
+# CHECK:       Contents of (__DATA,__data) section
+# CHECK-NEXT:  0000000000001000  23 01 00 00 00 00 00 00 00 10 00 00 00 00 00 00 {{$}}
+# CHECK-NEXT:  0000000000001010  00 10 00 00 00 00 00 00 {{$}}
+# CHECK-EMPTY:
+# CHECK-NEXT:  SYMBOL TABLE:
+# CHECK-NEXT:  0000000000001000 l     O __DATA,__data _alias
+# CHECK-NEXT:  0000000000001008 l     O __DATA,__data _ref
+# CHECK-NEXT:  0000000000001000 l     O __DATA,__data _alias
+# CHECK-NEXT:  0000000000001010 l     O __DATA,__data _ref
+# CHECK-NEXT:  0000000000001000  w    O __DATA,__data _weak
+# CHECK-NEXT:  0000000000000000         *UND* dyld_stub_binder
+# CHECK-EMPTY:
+## Even though the references were to the non-weak `_alias` symbols, ld64 still
+## emits weak binds as if they were the `_weak` symbol itself. We do not. I
+## don't know of any programs that rely on this behavior, so I'm just
+## documenting it here.
+# CHECK-NEXT:  Weak bind table:
+# CHECK-NEXT:  segment  section            address     type       addend   symbol
+# CHECK-EMPTY:
+
+# RUN: %lld -lSystem -dylib %t/local-then-weak.o %t/no-subsections.o -o %t/sub-nosub
+# RUN: llvm-objdump --macho --syms --section="__DATA,__data" %t/sub-nosub | FileCheck %s --check-prefix SUB-NOSUB
+
+## This test case demonstrates a shortcoming of LLD: If .subsections_via_symbols
+## isn't enabled, we don't elide the contents of coalesced weak symbols if they
+## are part of a section that has other non-coalesced symbols. In contrast, LD64
+## does elide the contents.
+# SUB-NOSUB:       Contents of (__DATA,__data) section
+# SUB-NOSUB-NEXT:  0000000000001000    23 01 00 00 00 00 00 00 00 10 00 00 00 00 00 00
+# SUB-NOSUB-NEXT:  0000000000001010    00 00 00 00 00 00 00 00 23 01 00 00 00 00 00 00
+# SUB-NOSUB-EMPTY:
+# SUB-NOSUB-NEXT:  SYMBOL TABLE:
+# SUB-NOSUB-NEXT:  0000000000001000 l     O __DATA,__data _alias
+# SUB-NOSUB-NEXT:  0000000000001008 l     O __DATA,__data _ref
+# SUB-NOSUB-NEXT:  0000000000001010 l     O __DATA,__data _zeros
+# SUB-NOSUB-NEXT:  0000000000001000 l     O __DATA,__data _alias
+# SUB-NOSUB-NEXT:  0000000000001000  w    O __DATA,__data _weak
+# SUB-NOSUB-NEXT:  0000000000000000         *UND* dyld_stub_binder
+
+# RUN: %lld -lSystem -dylib %t/no-subsections.o %t/local-then-weak.o -o %t/nosub-sub
+# RUN: llvm-objdump --macho --syms --section="__DATA,__data" %t/nosub-sub | FileCheck %s --check-prefix NOSUB-SUB
+
+# NOSUB-SUB:       Contents of (__DATA,__data) section
+# NOSUB-SUB-NEXT:  0000000000001000    00 00 00 00 00 00 00 00 23 01 00 00 00 00 00 00
+# NOSUB-SUB-NEXT:  0000000000001010    08 10 00 00 00 00 00 00 {{$}}
+# NOSUB-SUB-EMPTY:
+# NOSUB-SUB-NEXT:  SYMBOL TABLE:
+# NOSUB-SUB-NEXT:  0000000000001000 l     O __DATA,__data _zeros
+# NOSUB-SUB-NEXT:  0000000000001008 l     O __DATA,__data _alias
+# NOSUB-SUB-NEXT:  0000000000001008 l     O __DATA,__data _alias
+# NOSUB-SUB-NEXT:  0000000000001010 l     O __DATA,__data _ref
+# NOSUB-SUB-NEXT:  0000000000001008  w    O __DATA,__data _weak
+# NOSUB-SUB-NEXT:  0000000000000000         *UND* dyld_stub_binder
+
+## Verify that we don't drop any flags that the aliases have (such as
+## .no_dead_strip). This is a regression test. We previously had subsections
+## that were mistakenly stripped.
+
+# RUN: %lld -lSystem -dead_strip %t/no-dead-strip.o -o %t/no-dead-strip
+# RUN: llvm-objdump --macho --section-headers %t/no-dead-strip | FileCheck %s \
+# RUN:   --check-prefix=NO-DEAD-STRIP
+# NO-DEAD-STRIP: __data        00000010
+
+#--- weak-then-local.s
+.globl _weak
+.weak_definition _weak
+.data
+_weak:
+_alias:
+  .quad 0x123
+
+_ref:
+  .quad _alias
+
+.subsections_via_symbols
+
+#--- local-then-weak.s
+.globl _weak
+.weak_definition _weak
+.data
+_alias:
+_weak:
+  .quad 0x123
+
+_ref:
+  .quad _alias
+
+.subsections_via_symbols
+
+#--- no-subsections.s
+.globl _weak
+.weak_definition _weak
+.data
+_zeros:
+.space 8
+
+_weak:
+_alias:
+  .quad 0x123
+
+#--- no-dead-strip.s
+.globl _main
+
+_main:
+  ret
+
+.data
+.no_dead_strip l_foo, l_bar
+
+_foo:
+l_foo:
+  .quad 0x123
+
+l_bar:
+_bar:
+  .quad 0x123
+
+.subsections_via_symbols