[LLD][ELF] Eliminate symbols of merged .ARM.exidx sections.

[ikudrin]

Summary

GNU tools generate mapping symbols "$d" for .ARM.exidx sections. The symbols are added to the symbol table much earlier than the merging takes place, and after that, they become dangling. Before the patch, LLD output those symbols as SHN_ABS with the value of 0. The patch removes such symbols from the symbol table.

[psmith]

Thanks for the patch. I think what you have looks like it will work.

I'm thinking that there may be a simpler option. As the .ARM.exidx is only accessed via the linker defined start and end symbols and mapping symbols are optional for sections containing data; we can either not add any symbols from .ARM.exidx sections to the symbol table, or strip all of them rather than having to select just the ones defined in the merged sections. Maybe worth seeing if that is possible?

[grimar]

(a minor general comment about the code from me).

lld/ELF/SyntheticSections.cpp

2034 ↗

(On Diff #259934)

I think you wanted to use cast:

auto *d = cast<Defined>(s.sym);

[ikudrin]

• Reworked the patch so that the local symbols for .ARM.exidx sections are not added to the symbols table. Thanks for the suggestion, @psmith!

[psmith]

Thanks for the update, I'm happy with that. Will be worth seeing if MaskRay or George have any final comments.

[MaskRay]

lld/test/ELF/arm-exidx-mapping-symbols.s

2

///

24

,"ax"

[MaskRay]

lld/test/ELF/arm-exidx-mapping-symbols.s

16

///

lld/ELF/Writer.cpp

static bool shouldKeepInSymtab(const Defined &sym) {
   if (sym.isSection())
     return false;
 
   // If --emit-reloc or -r is given, preserve symbols referenced by relocations
   // from live sections.
   if (config->copyRelocs && sym.used)
     return true;
 
+  // Exclude local symbols pointing to .ARM.exidx sections.
+  // They are probably mapping symbols "$d", which are optional for these
+  // sections. After merging the .ARM.exidx sections, some of these symbols
+  // may become dangling. The easiest way to avoid the issue is not to add
+  // them to the symbol table from the beginning.
+  if (config->emachine == EM_ARM && sym.section &&
+      sym.section->type == SHT_ARM_EXIDX)
+    return false;
+
   if (config->discard == DiscardPolicy::None)
     return true;
   if (config->discard == DiscardPolicy::All)
     return false;
 
   // In ELF assembly .L symbols are normally discarded by the assembler.
   // If the assembler fails to do so, the linker discards them if
   // * --discard-locals is used.

lld/test/ELF/arm-exidx-mapping-symbols.s

+// REQUIRES: arm
+// Test that symbols which point to input .ARM.exidx sections are eliminated.
+// These symbols might be produced, for example, by GNU tools.
+
+// RUN: llvm-mc -filetype=obj -triple=armv7a-none-linux-gnueabi %s -o %t
+// RUN: ld.lld %t -o %t2
+// RUN: llvm-readelf -s %t2 | FileCheck %s
+
+// CHECK-NOT: $d.exidx.foo
+
+    .syntax unified
+    .section .text.foo,"axG",%progbits,foo,comdat
+foo:
+    bx lr
+
+// GNU as adds mapping symbols "$d" for .ARM.exidx sections it generates.
+// llvm-mc does not do that, so reproduce that manually.
+    .section .ARM.exidx.text.foo,"ao?",%0x70000001,.text.foo
+$d.exidx.foo:
+    .reloc 0, R_ARM_NONE, __aeabi_unwind_cpp_pr0
+    .long .text.foo(PREL31)
+    .long 0x80b0b0b0
+
+    .section .text.h
+    .global __aeabi_unwind_cpp_pr0
+__aeabi_unwind_cpp_pr0:
+    bx lr