On Mon, Jan 23, 2012 at 1:29 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> profiledbootstrap currently fails on powerpc64-linux (64-bit). The problem
> is that the linker errors on .toc section references from within
> non-allocated sections. Alan changed the linker some months ago to
> not die with assertion failure, but error out on these, and the attached
> testcase reduced from cp/parser.c during bootstrap shows that we still
> can hit it in some cases. The problem here is that var-tracking.c
> (adjust_insn) isn't able to delegitimize it, because the UNSPEC_TOCREL
> isn't added there to r2 register, but to a debug_expr (which only afterwards
> is found to contain r2 value). So the UNSPEC_TOCREL makes it through to
> mem_loc_descriptor, where it is already properly delegitimized into a
> (mem (symbol_ref (".LC0")) where .LC0 is a constant pool (.toc) symbol.
> We in some places try to prefer avoid_constant_pool_reference values over
> the mems, in this case avoid_constant_pool returns (symbol_ref "w"), which
> is a non-local symbol which isn't const_ok_for_output though. So
> mem_loc_descriptor on the avoid_constant_pool_reference value fails and we
> try the original (mem (symbol_ref (".LC0")) - on other targets this is still
> useful, if we know the right value is in say some .got entry etc. and we
> can reference that location from within the .debug_info, we win.
> Unfortunately the ppc64 linker doesn't like this, so we need to reject
> those. The following patch adds a target hook and rejects there from
> emitting into .debug_info/.debug_loc any SYMBOL_REFs that are in the
> .toc section.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, profiledbootstrapped
> on powerpc64-linux (both --with-cpu=default32 and defaulting to 64-bit,
> the last one failed without this patch), powerpc64-linux regtest is in
> progress (but the testcase in the patch has been tested already to
> fail without the patch and succeed with the patch).
> Ok for trunk?
>
> 2012-01-23 Jakub Jelinek <ja...@redhat.com>
>
> PR target/51957
> * target.def (const_not_ok_for_debug_p): New hook.
> * doc/tm.texi.in (TARGET_CONST_NOT_OK_FOR_DEBUG_P): New hook
> documentation.
> * doc/tm.texi: Regenerated.
> * dwarf2out.c (const_ok_for_output_1): If
> targetm.const_not_ok_for_debug_p returns true, fail.
> * config/rs6000/rs6000.c (rs6000_const_not_ok_for_debug_p): New
> function.
> (TARGET_CONST_NOT_OK_FOR_DEBUG_P): Redefine.
>
> * gcc.dg/pr51957-1.c: New test.
> * gcc.dg/pr51957-1.h: New file.
> * gcc.dg/pr51957-2.c: New test.
Okay.
Thanks, David
