[Mesa-dev] [PATCH] gallivm: fix [IU]MUL_HI regression

[Roland Scheidegger]

Am 08.11.2016 um 20:10 schrieb Marek Olšák:
> FYI, this doesn't fix the regression fully. (GLCTS failures with
> piglit: -t mulextended)

Maybe using shuffle isn't such a good idea then.
Not sure how well you handle them, and there's probably a problem with
scalar build contexts (initially this was restricted to 4 and 8-sized
vectors), looking at it we'd actually return a (1-sized) vector instead
of a scalar in the end... shifts/truncs have the advantage that they are
completely agnostic if it's scalars or vectors (and if it's vectors,
what kind of vectors).

Roland