What is the status regarding min/max intrinsics?

Project Infrastructure LLVM Dev List Archives
1

The llvm.smin, llvm.smax, llvm.umin and llvm.umax intrinsics were
introduced quite some time ago.

As far as I known we still don't canonicalize IR to use those
intrinsics (but I suppose that is the end goal), nor do we
canonicalize by doing a reverse transform into a select.

Are there any known limitations in support for the min/max intrinsics
that prevents us from using them or trying to canonicalize IR into
using the intrinsics?
(and if so, shouldn't we canonicalize into a select-pattern if that
form still is "better")

Main reason for asking is that I'd like to know if it is safe
to simplify clang/opt to use the min/max intrinsics directly
instead of emitting a select-kind-of-pattern. Or is it a huge risk
that I end up with regressions due to optimizers not understanding
the min/max intrinsics yet?

Regards,
Björn

2

We’re getting closer to canonicalizing to the intrinsics…
But there are still some missing optimizations (mostly in instcombine I think) that we are hoping to address to reduce the chance of regressions:

This is the most current view of what’s left:
https://reviews.llvm.org/D98152

We’ve also been tracking the work with:
https://llvm.org/PR46915

3

Thanks! That’s the information I was looking for.

(I should probably have searched Bugzilla before asking, but did not consider that there would be a tracking ticket for this.)

/Björn