[llvm] BranchProbabilityInfo: Report branch_weights 1, 0 for invoke (PR #72947)

[via llvm-commits]

WenleiHe wrote:

Not sure if this is worth adding a new calcFixedWeights...

One can always contrive a case where `0xfffff : 1` is actually close to ground truth than `1:0` for exception. If this is just for non-PGO case, it's not possible to make a generally good guess anyways.

However even if there is something that is generally better, can that be achieved with current calcEstimatedHeuristics? I see two ways of doing it, either increase `DEFAULT` or make `UNWIND ` 0. Having an inconsistent way of arriving at a branch probability seems unnecessary IMO.

https://github.com/llvm/llvm-project/pull/72947