[Bug ld/26262] New: ld/plugin.c:plugin_notice weak

[i at maskray dot me]

https://sourceware.org/bugzilla/show_bug.cgi?id=26262

            Bug ID: 26262
           Summary: ld/plugin.c:plugin_notice weak
           Product: binutils
           Version: unspecified
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: ld
          Assignee: unassigned at sourceware dot org
          Reporter: i at maskray dot me
  Target Milestone: ---

This can be reproduced with clang 10 (likely also 9; and all master commits
before https://reviews.llvm.org/D83967 ). LLVMgold.so is needed.

% cat a.c
int main() {}
% /tmp/Debug/bin/clang -fuse-ld=bfd -fprofile-generate -flto a.c
-Wl,-plugin-opt=save-temps -Wl,-y,__llvm_profile_raw_version
-Wl,-plugin-opt=save-temps
/usr/bin/ld.bfd: /tmp/a-e575a9.o (symbol from plugin): definition of
__llvm_profile_raw_version
/usr/bin/ld.bfd:
/tmp/Debug/lib/clang/12.0.0/lib/linux/libclang_rt.profile-x86_64.a(InstrProfiling.c.o):
definition of __llvm_profile_raw_version

/tmp/a-e575a9.o (symbol from plugin) provides a definition in a prevailing
comdat group.
libclang_rt.profile-x86_64.a(InstrProfiling.c.o) is an weak definition which
should not override the IR definition.

% cat a.out.resolution.txt 
/tmp/a-e575a9.o
-r=/tmp/a-e575a9.o,main,plx
-r=/tmp/a-e575a9.o,__llvm_profile_raw_version,l  # should be 'plx' instead of
'l'
-r=/tmp/a-e575a9.o,__llvm_profile_filename,plx

Gold and LLD are correct.

% /tmp/Debug/bin/clang -fuse-ld=gold -fprofile-generate -flto a.c
-Wl,-plugin-opt=save-temps -Wl,-y,__llvm_profile_raw_version
/tmp/a-668c9c.o: definition of __llvm_profile_raw_version
/tmp/Debug/lib/clang/12.0.0/lib/linux/libclang_rt.profile-x86_64.a(InstrProfiling.c.o):
definition of __llvm_profile_raw_version
a.out.o: definition of __llvm_profile_raw_version
% cat a.out.resolution.txt 
/tmp/a-668c9c.o
-r=/tmp/a-668c9c.o,main,plx
-r=/tmp/a-668c9c.o,__llvm_profile_raw_version,plx
-r=/tmp/a-668c9c.o,__llvm_profile_filename,plx

% /tmp/Debug/bin/clang -fuse-ld=lld -fprofile-generate -flto a.c
-Wl,-plugin-opt=save-temps -Wl,-y,__llvm_profile_raw_version
/tmp/a-afb841.o: definition of __llvm_profile_raw_version
<internal>: reference to __llvm_profile_raw_version
lto.tmp: definition of __llvm_profile_raw_version
% cat a.out.resolution.txt
/tmp/a-afb841.o
-r=/tmp/a-afb841.o,main,plx
-r=/tmp/a-afb841.o,__llvm_profile_raw_version,plx
-r=/tmp/a-afb841.o,__llvm_profile_filename,plx


The ld bug is in ld/plugin.c:plugin_notice . ld somehow drops
bfd_link_hash_defined (provided by "/tmp/a-e575a9.o (symbol from plugin)") in
favor of bfd_link_hash_undefweak. Then bfd_link_hash_undefweak is overridden by
the weak definition in libclang_rt.profile-x86_64.a(InstrProfiling.c.o)

      /* Otherwise, it must be a new def.
         Ensure any symbol defined in an IR dummy BFD takes on a
         new value from a real BFD.  Weak symbols are not normally
         overridden by a new weak definition, and strong symbols
         will normally cause multiple definition errors.  Avoid
         this by making the symbol appear to be undefined.  */
      else if (((h->type == bfd_link_hash_defweak
                 || h->type == bfd_link_hash_defined)
                && is_ir_dummy_bfd (sym_bfd = h->u.def.section->owner))
               || (h->type == bfd_link_hash_common
                   && is_ir_dummy_bfd (sym_bfd = h->u.c.p->section->owner)))
        {
          h->type = bfd_link_hash_undefweak;
          h->u.undef.abfd = sym_bfd;
        }

The intended behavior is to let "/tmp/a-e575a9.o (symbol from plugin)" provide
the definition.

-- 
You are receiving this mail because:
You are on the CC list for the bug.