We also learned to legalize v2f32 FMIN/FMAX nodes later on.
I tried to also support v3f32, which might also happen often, but they
were already widened to v4f32 by the generic split-and-rebuild-vector
legalization.
This happens in the HINT benchmark, where the SLP-vectorizer created
v2f32 fcmp/select code. The "correct" solution would have been to
teach the vectorizer cost model that v2f32 isn't legal (because really,
it isn't), but if we can vectorize we might as well do so.
I know the surrounding tests are already written like this, but I would like something more verbose that check we just generate this one instruction (+ret).