On Wed, 9 Dec 2015, Richard Biener wrote:
>
> Well, not really in the end - the actual testcase would need code
> hoisting to make the refs equal enough for ifcvt. Making the testcase
> simpler doesn't seem to capture the issue so for now without one
> (I'll keep trying).
>
> Bootstrapped and t
This time with a simplified testcase - the remaining issue with
the original one is if-conversion looking at DR_REF and nothing
hoisting the address compute of a[i+1] in both arms making it
see they are the same (yeah, SCEV to the rescue, but not in this
patch).
This patch re-organizes things so
Well, not really in the end - the actual testcase would need code
hoisting to make the refs equal enough for ifcvt. Making the testcase
simpler doesn't seem to capture the issue so for now without one
(I'll keep trying).
Bootstrapped and tested on x86_64-unknown-linux-gnu.
Richard.
2015-12-09
