Deal with deprecated AVX2 feature (and possibly others)

[appleby]

It looks like the upcoming sbcl 2.0.1 release includes changes to move certain symbols out of cl:*features* and into sb-impl:+internal-features+ [1]. IIUC, any "non-public" features will continue to work (for now), but issue a warning [2].

I haven't tested it, but it looks like we use at least one such soon-to-be-deprecated feature, namely avx2. We should figure out what to do about our usage of any "internal" features.

Options include:

1) Do nothing, ignore the warnings, and hope they remain deprecated for at least a decade :-) 2) Kindly ask SBCL maintainers to keep avx2 in *features* 3) If possible, find some other combination of public features that are equivalent 4) Hack around it 5) Who cares, everybody has a modern x86-64 processor, right?

[colescott]

Update: sbcl 2.0.1 removed the avx2 feature flag.

The main issue this brings up is that the if-feature flag in the asdf file cannot tell that we have avx enabled. Maybe there is some other way of conditionalizing loading the files.

[jmbr]

The AVX2 flag seems to be alive and kicking in 2.0.3 (see https://github.com/sbcl/sbcl/commit/b01dce77d6a27d31e203adc33e484c64f2de46f0 for instance).

[appleby]

It was moved out of cl:*features* and into sb-impl:+internal-features+, and sb-int:featurep was changed to warn and return nil for internal features. Maybe whatever is checking local-target-features in the sbcl build knows to include +internal-features+?

CL-USER> (lisp-implementation-type)
"SBCL"
CL-USER> (lisp-implementation-version)
"2.0.3"
CL-USER> #+avx2 :avx2
WARNING: :AVX2 is no longer present in *FEATURES*
; No value
CL-USER> #-avx2 :avx2
WARNING: :AVX2 is no longer present in *FEATURES*
:AVX2
CL-USER> (find :avx2 *features*)
NIL
CL-USER> (find :avx2 sb-impl:+internal-features+)
:AVX2

[appleby]

In the original issue description, I incorrectly assumed that SBCL would still respect sb-impl:+internal-features+ for feature expressions, but apparently it just issues the warning.

[jmbr]

Yeah, I've been looking at featurep too. This is a nuisance indeed. I've seen a couple of instances in the code where they do a union of *features* and +internal-features+ (during bootstrap). Perhaps we could do that or we could redefine sharp-plus-minus in src/code/sharpm.lisp.