[llvm] [VPlan] Don't apply predication discount to non-originally-predicated blocks (PR #160449)

[David Sherwood via llvm-commits]

david-arm wrote:

Having thought about this just now, I also suspect that when tail-folding due to low trip counts (i.e. TC <= 16) we probably clamp the max VF to the next largest power-of-2 value. So we would never choose a max VF of 16 for TC=5 - we'd probably choose a max VF of 8. There will be very few cases where the cost divisor goes above 1. We also don't know at this point what the interleave count will be.

https://github.com/llvm/llvm-project/pull/160449