[llvm] [LV] Fix runtime-VF logic when generating RT-checks (PR #130118)

[Florian Hahn via llvm-commits]

================
@@ -5,7 +5,6 @@ target triple = "aarch64-unknown-linux-gnu"
 
 ; Test case where the minimum profitable trip count due to runtime checks
 ; exceeds VF.getKnownMinValue() * UF.
-; FIXME: The code currently incorrectly is missing a umax(VF * UF, 28).
----------------
fhahn wrote:

I think this FIXME is wrong. We already generate the umax considering UF * VF. This is for the minimum iteration check, and UF*VF is needed for correctness.

the constant is the minimum iteration count needed to be profitable, but not for correctness. 

AFAICT the only reason the constant gets changed is due to the runtime checks being more expensive now, due to duplicated RT VF computations?

the actual memory checks seem to use the same runtime VF as originally?

https://github.com/llvm/llvm-project/pull/130118