On Fri, Sep 6, 2013 at 3:13 AM, Alan Modra <amo...@gmail.com> wrote:
> The following testcase taken from the linux kernel is miscompiled on
> powerpc64-linux.
>
> /* -m64 -mcmodel=medium -O -S -fno-section-anchors */
> static int x;
>
> unsigned long
> foo (void)
> {
> return ((unsigned long) &x) - 0xc000000000000000;
> }
>
> generates
> addis 3,2,x+4611686018427387904@toc@ha
> addi 3,3,x+4611686018427387904@toc@l
> blr
>
> losing the top 32 bits of the offset. Sadly, the assembler and linker
> do not complain, which is a hole in the ABI. (@ha and _HA relocs as
> per the ABI won't complain about overflow since they might be used in
> a @highesta, @highera sequence loading a 64-bit value.)
>
> This patch stops combine merging large offsets into a symbol addend
> by copying code from reg_or_add_cint_operand to a new predicate,
> add_cint_operand, and using that to restrict the range of offsets.
> Bootstrapped and regression tested powerpc64-linux. OK to apply?
>
> * config/rs6000/predicates.md (add_cint_operand): New.
> * config/rs6000/rs6000.md (largetoc_high_plus): Restrict offset
> using add_cint_operand.
> (largetoc_high_plus_aix): Likewise.
This patch should include a testcase.
But what user feedback are you expecting if the offset is too large,
such as your example? In my test with the patch, it produces an
unrecognizable insn error, which seems less than friendly.
Thanks, David
