[patch] Partial fix for PR 18048

[Dale Johannesen]

On Oct 27, 2004, at 9:03 AM, Zdenek Dvorak wrote:
>> But the real question is why is &a->data[0] not pulled out of the loop
>> in the first place.
>
> &a->data[0] is a constant (symbol reference + offset), so moving it 
> out of loop would
> very likely just cause the register pressure to increase.

Of course it increases the register pressure, but it is still 
beneficial to hoist in
some cases.  It should be considered as a hoisting candidate.  (mgrid 
is not
the only specmark where not enough candidates are being hoisted on ppc;
just the easiest to analyze, as it spends 99% of its time in one loop.  
I could
report others as PRs but my expectation is that fixing mgrid will fix 
most of
the rest as well.)