[polly] r244606 - Revise the simplification of regions

[Michael Kruse via llvm-commits]

2015-08-12 8:47 GMT+02:00 Tobias Grosser <tobias at grosser.es>:
> Hi Michael,
>
> thank you for this detailed analysis. I played a little with what you
> said and came up with the attached patch which I believe solves
> the issue without introducing any special handling. What do you think?

So you chose solution 2)

In the meanwhile I was looking into pifft. It is the same issue with
the difference that the PHI node is used in a different BB of the
scop. This caused an undef value because the using BB generated a load
for that value, but the write for that PHI value would never be
generated because its inside the region.

I was also looking into solution number 2 since I was afraid 1) would
break more things that it would fix. But your patch works well for
pifft as well. Are you going to commit it? I will add a second test
case for the pifft case.

Michael