[Bug rtl-optimization/80401] [7 regression] gcc.target/powerpc/dimode_off.c and gcc.target/powerpc/pr79038-1.c fail starting with r246764

[meissner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80401

Michael Meissner <meissner at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2017-04-13
                 CC|                            |meissner at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #3 from Michael Meissner <meissner at gcc dot gnu.org> ---
While the underlying bug is due to an LRA change, the fix for each of these in
the rs6000 backend is likely to be different.

The first failure (dimode_off) is one of these tests that are so hard-wired to
the compiler that it likely needs to change every so often.

Yes, it can become a regression, but it only happens on a power8/power9 system
and the offset of a load/store is not divisible by 4.  Due to alignment issues,
the DImode load/store is almost always divisible by 4, unless you are loading
up unaligned memory, or specifically add odd numbers to a char * pointer and
convert it to a long pointer.  These should be rare in practice.

The underlying reason is the LD and STD instructions use the bottom 2 bits for
encoding the instruction, while the LFD and STFD instructions allows a full
16-bit offset.

The second failure listed is 79308-1.c that makes sure an optimization I put in
still works.  However, since the optimization is converting a char/short value
to __float128 on a Power9