http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45233
--- Comment #6 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-11-07
09:28:17 UTC ---
(In reply to comment #5)
> (In reply to comment #4)
> > (In reply to comment #3)
>
> FWIW gcc-4.2.1 (Apple local) fails and clang passes the test (although clang's
> asm at version 2.9 is a bit long-winded c.f. GCC's ... but that's quite an old
> version).
>
> > > or not properly restoring the UNSPEC during reload (it's its duty, not
> > > IRAs).
> >
> > I'll check this too.
>
> Thanks Richi, this seems to be the problem - it appears that the reload
> legitimizer does not consider that the item might be 'undefined' (in the
> mach-o, local-to-the-file sense) - and, therefore, there was no mechanism to
> recreate the necessary indirection.
>
> so, this is a fix - which bootstraps and regtests:
>
> I'd very much like an opinion as to whether it's the _right_ fix ... this is
> an
> area with which I am not familiar and therefore could have missed some other
> guard that's needed.
>
> It's clearly a very infrequent (perhaps even non-existent) occurrence outside
> the particular asm test-case (which makes me wonder still if combine came to
> the right decision to remove the refs in the first place).
Combine also asks the target whether the result is valid, so yes.
> Index: gcc/config/rs6000/rs6000.c
> ===================================================================
> --- gcc/config/rs6000/rs6000.c (revision 181027)
> +++ gcc/config/rs6000/rs6000.c (working copy)
> @@ -6185,7 +6185,21 @@ rs6000_legitimize_reload_address (rtx x, enum mach
> #if TARGET_MACHO
> if (flag_pic)
> {
> - rtx offset = machopic_gen_offset (x);
> + rtx offset;
> +
> + /* Reload might present us with a case where there is a reference to
> + an undefined entity. */
> + if (!machopic_symbol_defined_p (x) && !MACHO_DYNAMIC_NO_PIC_P)
> + {
> + offset = gen_rtx_SYMBOL_REF (Pmode,
> + machopic_indirection_name (x,
> + false));
> + SYMBOL_REF_DATA (offset) = SYMBOL_REF_DATA (x);
> + machopic_define_symbol (gen_const_mem (Pmode,offset));
> + x = offset;
> + }
> +
> + offset = machopic_gen_offset (x);
> x = gen_rtx_LO_SUM (GET_MODE (x),
> gen_rtx_PLUS (Pmode, pic_offset_table_rtx,
> gen_rtx_HIGH (Pmode, offset)), offset);
I suppose you should post this to gcc-patches@.
