[llvm] Reland "[GlobalISel][LLT] Introduce FPInfo for LLT (Enable bfloat, ppc128float and others in GlobalISel) (#155107)" (PR #188502)

[via llvm-commits]

DenisGZM wrote:

> Could you investigate what specifically causes the regression and say to which extent this could be optimized?

Well, the first obvious guess is that we added checks if extended LLT is available for every new builders and we used them in shared places like IRTranslator or LegalizerHelper instead of constexpr scalar now.

I've tried to quickly reproduce regression in some tests, but I got the opposite result, so it still need to be investigated more deeply. I'll try to look into it this weekend.

https://github.com/llvm/llvm-project/pull/188502