varasm: Fix up process_pending_assemble_externals [PR113182]
commitc05beab4ae240a593299c08ef8c775d91187a141
authorJakub Jelinek <jakub@redhat.com>
Fri, 12 Jan 2024 12:58:07 +0000 (12 13:58 +0100)
committerJakub Jelinek <jakub@redhat.com>
Fri, 12 Jan 2024 12:58:07 +0000 (12 13:58 +0100)
tree75a0fd8a5ea0233232bd4f61cb7f8eff47d5c97f
parent74e3e839ab2d368413207455af2fdaaacc73842b
varasm: Fix up process_pending_assemble_externals [PR113182]

John reported that on HP-UX we no longer emit needed external libcalls.

The problem is that we didn't strip name encoding when looking up
the identifiers in assemble_external_libcall and
process_pending_assemble_externals, while
assemble_name_resolve does that:
  const char *real_name = targetm.strip_name_encoding (name);
  tree id = maybe_get_identifier (real_name);

  if (id)
    {
...
      mark_referenced (id);
The intention is that assemble_external_libcall ensures the IDENTIFIER
exists for the external libcall, then for actually emitted calls
assemble_name_resolve sees those IDENTIFIERS and sets TREE_SYMBOL_REFERENCED
on them and finally process_pending_assemble_externals looks the
IDENTIFIER up again and checks its TREE_SYMBOL_REFERENCED.

But without the strip_name_encoding call, they can look up different
identifiers and those are likely never used.

In the PR, John was discussing whether get_identifier or
maybe_get_identifier should be used, I believe in assemble_external_libcall
we definitely want to use get_identifier, we need an IDENTIFIER allocated
so that it can be actually tracked, in process_pending_assemble_externals
it doesn't matter, the IDENTIFIER should be already created.

2024-01-12  John David Anglin  <danglin@gcc.gnu.org>
    Jakub Jelinek  <jakub@redhat.com>

PR middle-end/113182
* varasm.cc (process_pending_assemble_externals,
assemble_external_libcall): Use targetm.strip_name_encoding
before calling get_identifier.
gcc/varasm.cc