[llvm] [InstCombine] Avoid breaking FMA pattern when hoisting freeze (PR #141934)

[Harrison Hao via llvm-commits]

harrisonGPU wrote:

> I was going to ask about this on Discouse. The way that InstCombine moves freeze instructions seems a bit arbitrary to me. But adding extra checks like this, to avoid putting freeze in the middle of any pattern that _might_ be recognized by instruction selection, feels like even more of a hack.

But I think once we add a freeze here, we can't recognize the FMA pattern across the freeze.

https://github.com/llvm/llvm-project/pull/141934