fix target/31361

[Richard Guenther]

On 3/27/07, Richard Henderson <rth@redhat.com> wrote:
> The bug is that the cpu considers the entire 128-bit register
> to be the shift count, and detects excessive shift and forces
> the result to be zero.  So we really need a TImode input for
> the shift count; we can't allow garbage in the upper bits of
> an SImode input.
>
> Tested i686-linux; I'll commit to 4.[12] when testing finishes.

Does that reliably work with 32bit HWI?  I wonder if you can trick
the compiler into producing a TImode constant with that patch...

Richard.