[llvm] LAA: generalize strides over unequal type sizes (PR #108088)

[Ramkumar Ramachandra via llvm-commits]

artagnon wrote:

> > It looks like there are 2 distinct changes in the patch
> > 
> > * improved reasoning with non-matching strides (change in llvm/test/Analysis/LoopAccessAnalysis/depend_diff_types.ll?)
> > * enabling runtime checks for cases where there's no common stride.
> > 
> > Could this be split up? The first should hopefully a strict improvement, while the later is problematic because in many cases the runtime checks will always be false.
> 
> I initially just had the first change, but since it scales the strides, checking for a common stride results in regressions. That was the reason I decided to enable runtime checks when there is no common stride. Will re-investigate whether it can be split up.

Thanks, just reverted second change and had a look at the test changes once again: the regressions seem to be artificial.

https://github.com/llvm/llvm-project/pull/108088