AMD Zen 3 CPU support (Fixes #45657)

[LebedevRI]

I've omitted invpcid feature since it's clearly omitted for other CPU's.

It's a bit sad that the very same code
is being repeated in every project :/

Fixes #45657

[LebedevRI]

CC @yuyichao, thanks.

[vchuravy]

It's a bit sad that the very same code is being repeated in every project

There is some community around archspec

[LebedevRI]

Yes, please, i was just about to ask for that :)

[LebedevRI]

One caveat that i'm observing here, is that specifying JULIA_CPU_TARGET=native in Make.user
produces bad sysimage:

$ ./usr/bin/julia --cpu-target znver3
ERROR: Unable to find compatible target in system image.

... but it works with JULIA_CPU_TARGET=znver3.

I guess it's expected?

[yuyichao]

I guess it's expected?

This should not happen on the same machine.

[LebedevRI]

I guess it's expected?

This should not happen on the same machine.

Oh, i get it now. Zen 3 also implemented cet_ss (Control flow enforcement (CET) shadow stack),
it is being detected for sysimage, but i didn't list it in features.

[LebedevRI]

Now with JULIA_CPU_TARGET=native, both the --cpu-target native and the --cpu-target znver3 work.

[LebedevRI]

I'm not sure i understand those test failures.
They seem to be happening even on znver2 and AArch64 builds.

[vchuravy]

They seem to be happening even on znver2 and AArch64 builds.

Yeah not relevant to your PR. x-ref: #45613 (comment)

Stale

[yuyichao]

Oh, i get it now. Zen 3 also implemented cet_ss (Control flow enforcement (CET) shadow stack),
it is being detected for sysimage, but i didn't list it in features.

Was the thing that didn't work compiling with native and running with znver3? Yeah that'll not work if there's a mismatch. Compiling and running with native (the default) should work though no matter what feature you've listed for znver3..

[yuyichao]

LGTM. I'm just assuming that the feature set and CPU id is correct.......

[LebedevRI]

Oh, i get it now. Zen 3 also implemented cet_ss (Control flow enforcement (CET) shadow stack),
it is being detected for sysimage, but i didn't list it in features.

Was the thing that didn't work compiling with native and running with znver3? Yeah that'll not work if there's a mismatch. Compiling and running with native (the default) should work though no matter what feature you've listed for znver3..

Something like that, yes. I've debugged that by dumping the imgt.en.features and target.dis.features
in match_sysimg_targets(), looking for the mismatched bit, and comparing the bit index to the enum

julia/src/features_x86.h

Line 61 in 628209c

JL_FEATURE_DEF(shstk, 32 * 3 + 7, 0)

.

LGTM. I'm just assuming that the feature set and CPU id is correct.......

I'm confident that it is, and it works now.

[yuyichao]

You could just do a ccall(:jl_dump_host_cpu, Cvoid, ()) to see the detected featured .......

[LebedevRI]

You could just do a ccall(:jl_dump_host_cpu, Cvoid, ()) to see the detected featured .......

I could, but i didn't know i could. Still looks good:

julia> ccall(:jl_dump_host_cpu, Cvoid, ())
CPU: znver3
Features: sse3, pclmul, ssse3, fma, cx16, sse4.1, sse4.2, movbe, popcnt, aes, xsave, avx, f16c, rdrnd, fsgsbase, bmi, avx2, bmi2, rdseed, adx, clflushopt, clwb, sha, pku, shstk, vaes, vpclmulqdq, rdpid, sahf, lzcnt, sse4a, prfchw, mwaitx, xsaveopt, xsavec, xsaves, clzero, wbnoinvd

[LebedevRI]

Apologies, i'm not familiar with the julia's PR workflow.
Do i need to do anything further to get this merged?
It seems like #45613 still has not been reverted, so CI will stay red.

[LebedevRI]

@KristofferC thank you!