[PATCH] ld: Skip the unversioned symbol from LTO

[Alan Modra]

On Thu, Aug 26, 2021 at 07:52:47AM -0700, H.J. Lu wrote:
> +      if ((oldbfd->flags & BFD_PLUGIN) != 0
> +	  && h != hi
> +	  && h->versioned != unversioned
> +	  && hi->versioned == unversioned
> +	  && hi->root.type == bfd_link_hash_indirect)
> +	{
> +	  /* Skip the unversioned symbol from LTO if there is a
> +	     versioned symbol and the unversioned symbol has been
> +	     added as an indirect symbol from IR.  */
> +	  *skip = true;
> +	  return true;
> +	}

Normally a non-IR symbol definition takes precedence over an IR symbol
definition, so this looks wrong to me.

I haven't looked at what is going on here in much detail yet, but it
seems likely the correct thing to do is twiddle hi so it is no longer
indirect to an IR versioned symbol.  Then the LTO output object will
provide the correct definition, and if the LTO output also includes a
matching versioned symbol it will go back to being indirect at that
point.   Hmm, I suppose what you're doing is correct *if* the LTO
output always includes both versioned and unversioned symbols.  It
probably isn't a good idea to rely on that always being the case.

BTW,
> +	  && h != hi
> +	  && h->versioned != unversioned
> +	  && hi->versioned == unversioned
all of these tests are redundant.

The following should work.

      if ((oldbfd->flags & BFD_PLUGIN) != 0
	  && hi->root.type == bfd_link_hash_indirect)
	{
	  /* Don't leave the symbol as indirect.  */
	  hi->root.type = bfd_link_hash_undefined;
	  hi->root.u.undef.abfd = oldbfd;
	}

> +++ b/ld/testsuite/ld-plugin/pr28264-1.d
> @@ -0,0 +1,7 @@
> +Symbol table '\.dynsym' contains [0-9]+ entries:
> + +Num: +Value +Size Type +Bind +Vis +Ndx Name
> +#...
> + +[0-9]+: +[0-9a-f]+ +[0-9a-f]+ +FUNC +GLOBAL +DEFAULT .*[0-9]+ _?bar@@VERSION.1
> + +[0-9]+: +[0-9a-f]+ +[0-9a-f]+ +FUNC +GLOBAL +DEFAULT .*[0-9]+ _?foo@VERSION.1
> +#pass

I saw failures of the test on mips64-linux, mips-linux, and
x86_64-linux, depending on the version of compiler/startup/libraries
or something that affects the set of symbols in the linker hash table.
If you are unlucky, foo@VERSION.1 can come before bar@@VERSION.1

I also saw a weird fail on microblaze-linux
lto1: error: symver is only supported on ELF platforms
Maybe I need a newer microblaze compiler?

> +++ b/ld/testsuite/ld-plugin/pr28264-2.d
> @@ -0,0 +1,4 @@
> +#failif
> +#...
> + +[0-9]+: +[0-9a-f]+ +[0-9a-f]+ +FUNC +GLOBAL +DEFAULT .*[0-9]+ _?foo
> +#...

> +++ b/ld/testsuite/ld-plugin/pr28264-3.d
> @@ -0,0 +1,4 @@
> +#failif
> +#...
> + +[0-9]+: +[0-9a-f]+ +[0-9a-f]+ +FUNC +GLOBAL +DEFAULT .*[0-9]+ _?foo
> +#...

The above two files are identical, presumably one should test that bar
is not present.

-- 
Alan Modra
Australia Development Lab, IBM