[llvm] [Float2Int] Resolve FIXME: Pick the smallest legal type that fits (PR #79158)

[Jessica Clarke via llvm-commits]

jrtc27 wrote:

This seems to be a preexisting issue that you can hit if you pick specific float values: https://godbolt.org/z/rxdqrve3d

It's just hit more often now because you're picking smaller integers. Before you had to be in the range [2^31, 2^32), converted to a double, but now it's true for [2^7, 2^8) and [2^15, 2^16), both of which also fit in a single-precision float, unlike the prior case.

https://github.com/llvm/llvm-project/pull/79158