[llvm] [VPlan] Don't use the legacy cost model for loop conditions (PR #156864)

[John Brawn via llvm-commits]

john-brawn-arm wrote:

> Looks like there's another potential crash:

It looks like this is a consequence of the change I made due to https://github.com/llvm/llvm-project/pull/156864#discussion_r2602864311. The umax doesn't exist in the vplan, which would normally cause planContainsAdditionalSimplifications to return true, but because it's used only in the exit condition we skip it, so planContainsAdditionalSimplifications returns false and then we fail the assertion.

I'll undo this change, which means we'll be back to planContainsAdditionalSimplifications returning true in more situations. I think this is fine, and possible we should be thinking of just getting rid of this check that the two cost models give the same result except in the cases where they don't, as it feels like it's been a lot of work for something that's going to be increasingly less useful as the vplan cost model gets more accurate.

https://github.com/llvm/llvm-project/pull/156864