[llvm] [GlobalIsel] Push cast through build vector (PR #104634)

[Thorsten Schütt via llvm-commits]

tschuett wrote:

> > I believe what we found was that if there were two truncs it would be OK to leave it scalarized, but for 4+ the vector trunc is generally better so long as the inputs come from a vector.
> 
> @tschuett I think it would be a good idea to try this - by bailing out if the buildvector has >= 4 sources. That should help mitigate the more serious regressions that are occuring.

I am somehow afraid that after legalization large vectors are split in smaller ones and the heuristic fails to recognize bad cases.

https://github.com/llvm/llvm-project/pull/104634