On 2/3/2017 3:03 PM, Walter Lee wrote:
Hi,
In looking at PR 79365 I found that the problem is actually in the
combiner. The combiner sometimes applies scalar optimizations to
vector context where they are invalid. i.e. (a > b) >> 1 can optimize
to 0 if ">" is a scalar op but not if it is a vector op. The reason
this shows up on tile* and not other architectures is because tile*
uses the same register file for both scalars and vectors, so the
combiner sees these mixed mode expressions on tile* that would go
through memory on other architectures.
I have found two places where this improper optimization occurs.
Patch below.
Ok to commit?
I forgot to mention this fix has been validated on TILEPro/TILE-Gx. I also
sanity checked it against x86.
Thanks,
Walter
