[LLVMdev] fix for loop scale limiting in BFI

[Diego Novillo]

On Sun, Mar 29, 2015 at 2:02 PM, Duncan P. N. Exon Smith
<dexonsmith at apple.com> wrote:

> Your change to `convertFloatingToInteger()` looks basically correct to
> me.  If we need to saturate at one end or the other, it makes more sense
> to be saturating at the bottom.

Thanks. Special casing infinite loops was the problem here. Saturating
everything else down to 1 in normal circumstances is not a big deal in
the cases I tried.

I've sent http://reviews.llvm.org/D8718 for review. The change should
be performance neutral, but I'm running a batch of SPEC and internal
tests to make sure.


Thanks.  Diego.