Add Preference to disable runtime invalidation #22

topolarity · 2025-11-19T22:55:49Z

More conservative version of #21 (thanks to @el-oso for the inspiration).

This makes the package (nearly) --trim compatible and also provides a nice preference for users who do not wish for HostCPUFeatures.jl to ever cause an "invalidation storm", even if it means running with the wrong CPU info.

If the CPU info measured at build (precompile) time does not match runtime, users will see a warning:

┌ Warning: Runtime invalidation was disabled, but the CPU info is out-of-date.
│ Will continue with incorrect CPU name (from build time).
└ @ HostCPUFeatures ~/repos/HostCPUFeatures.jl/src/HostCPUFeatures.jl:62

This makes the package (nearly) `--trim` compatible and also provides a nice preference for users who do not wish for HostCPUFeatures.jl to ever cause an "invalidation storm", even if it means running with the wrong CPU info.

oscardssmith · 2025-11-20T14:51:08Z

src/cpu_info_x86.jl

 fast_half() = False()

-@noinline function setfeaturefalse(s)
+@inline function setfeaturefalse(s)


why the inlining switch?

This was to improve inferrability - has_feature(Val(s)) is only inferrable if the literal value of s is available in the function, which is true in all of the callers (the argument is always a literal Symbol)

codecov · 2025-11-20T14:51:26Z

Codecov Report

❌ Patch coverage is 7.14286% with 26 lines in your changes missing coverage. Please review.
✅ Project coverage is 18.50%. Comparing base (8ebc3d2) to head (723446a).
⚠️ Report is 5 commits behind head on main.

Files with missing lines	Patch %	Lines
src/cpu_info.jl	0.00%	9 Missing ⚠️
src/cpu_info_aarch64.jl	0.00%	7 Missing ⚠️
src/cpu_info_x86.jl	0.00%	6 Missing ⚠️
src/HostCPUFeatures.jl	33.33%	4 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (8ebc3d2) and HEAD (723446a). Click for more details.

HEAD has 5 uploads less than BASE

Flag BASE (8ebc3d2) HEAD (723446a)

15 10

Additional details and impacted files

@@             Coverage Diff             @@
##             main      #22       +/-   ##
===========================================
- Coverage   36.12%   18.50%   -17.63%     
===========================================
  Files           6        6               
  Lines         191      200        +9     
===========================================
- Hits           69       37       -32     
- Misses        122      163       +41

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

tz-lom · 2025-11-20T15:10:42Z

src/cpu_info.jl

+                @debug "Defining $(has ? "presence" : "absense") of feature $feature."
+                set_feature(feature, has)
+            else
+                @warn "Runtime invalidation was disabled, but the CPU info is out-of-date.\nWill continue with incorrect CPU feature flag: $ext."


aside of this change, I expect that feature_string not always return same set of features, that will make certain features stuck in wrong state just because LLVM does not report them for that CPU

For the reference, that's the info I've collected from real and virtual machines with

@ccall jl_get_cpu_name()::String @ccall jl_get_cpu_features()::String

* "broadwell" "+prfchw,+avx,+aes,+sahf,+pclmul,+crc32,+sse4.1,+xsave,+sse4.2,+invpcid,+64bit,+cmov,+movbe,+rtm,+adx,+avx2,+bmi,+sse,+xsaveopt,+rdrnd,+cx8,+sse3,+fsgsbase,+lzcnt,+ssse3,+cx16,+bmi2,+fma,+popcnt,+f16c,+mmx,+sse2,+rdseed,+fxsr,-cldemote,-xop,-xsaves,-avx512fp16,-usermsr,-sm4,-avx512ifma,-avx512pf,-tsxldtrk,-ptwrite,-widekl,-sm3,-xsavec,-avx10.1-512,-avx512vpopcntdq,-avx512vp2intersect,-avx512cd,-avxvnniint8,-avx512er,-amx-int8,-kl,-avx10.1-256,-sha512,-avxvnni,-hreset,-movdiri,-serialize,-vpclmulqdq,-avx512vl,-uintr,-clflushopt,-raoint,-cmpccxadd,-amx-tile,-gfni,-avxvnniint16,-amx-fp16,-avx512f,-amx-bf16,-avx512bf16,-avx512vnni,-avx512bw,-pku,-clzero,-mwaitx,-lwp,-sha,-movdir64b,-wbnoinvd,-enqcmd,-prefetchwt1,-avxneconvert,-tbm,-pconfig,-amx-complex,-avxifma,-avx512bitalg,-rdpru,-clwb,-avx512vbmi2,-prefetchi,-rdpid,-fma4,-avx512vbmi,-shstk,-vaes,-waitpkg,-sgx,-avx512dq,-sse4a" * "cascadelake" "+prfchw,+avx,+aes,+sahf,+pclmul,+crc32,+xsaves,+sse4.1,+xsave,+sse4.2,+invpcid,+64bit,+xsavec,+cmov,+avx512cd,+movbe,+evex512,+adx,+avx2,+avx512vl,+clflushopt,+bmi,+sse,+xsaveopt,+rdrnd,+avx512f,+avx512vnni,+cx8,+avx512bw,+sse3,+pku,+fsgsbase,+lzcnt,+ssse3,+cx16,+bmi2,+fma,+popcnt,+f16c,+clwb,+mmx,+sse2,+rdseed,+fxsr,+avx512dq,-cldemote,-xop,-avx512fp16,-usermsr,-sm4,-avx512ifma,-avx512pf,-tsxldtrk,-ptwrite,-widekl,-sm3,-avx10.1-512,-avx512vpopcntdq,-avx512vp2intersect,-avxvnniint8,-avx512er,-amx-int8,-kl,-avx10.1-256,-avxvnni,-rtm,-hreset,-movdiri,-serialize,-sha512,-vpclmulqdq,-uintr,-raoint,-cmpccxadd,-amx-tile,-gfni,-avxvnniint16,-amx-fp16,-amx-bf16,-avx512bf16,-clzero,-mwaitx,-lwp,-sha,-movdir64b,-wbnoinvd,-enqcmd,-prefetchwt1,-avxneconvert,-tbm,-pconfig,-amx-complex,-avxifma,-avx512bitalg,-rdpru,-avx512vbmi2,-prefetchi,-rdpid,-fma4,-avx512vbmi,-shstk,-vaes,-waitpkg,-sgx,-sse4a" * "skylake-avx512" "+cx16,+sahf,+crc32,+prfchw,+bmi2,+fsgsbase,+popcnt,+aes,+xsaves,+clwb,+avx512f,+xsavec,+pku,+mmx,+rdseed,+avx512bw,+clflushopt,+xsave,+64bit,+avx512vl,+invpcid,+avx512cd,+avx,+cx8,+fma,+bmi,+rdrnd,+sse4.1,+sse4.2,+avx2,+fxsr,+sse,+lzcnt,+pclmul,+f16c,+ssse3,+cmov,+movbe,+xsaveopt,+avx512dq,+sse2,+adx,+sse3,-avx512pf,-tsxldtrk,-tbm,-avx512ifma,-sha,-fma4,-vpclmulqdq,-cldemote,-avx512bf16,-amx-tile,-raoint,-uintr,-gfni,-ptwrite,-avx512bitalg,-movdiri,-widekl,-avx512er,-avxvnni,-avx512fp16,-avx512vnni,-amx-bf16,-avxvnniint8,-avx512vpopcntdq,-pconfig,-cmpccxadd,-clzero,-amx-fp16,-lwp,-rdpid,-xop,-waitpkg,-prefetchi,-kl,-movdir64b,-sse4a,-avxneconvert,-avx512vbmi2,-serialize,-hreset,-vaes,-amx-int8,-rtm,-enqcmd,-mwaitx,-wbnoinvd,-rdpru,-avxifma,-sgx,-prefetchwt1,-avx512vbmi,-shstk,-avx512vp2intersect" * "skylake-avx512" "+prfchw,+avx,+aes,+sahf,+pclmul,+crc32,+sse4.1,+xsave,+sse4.2,+64bit,+cmov,+movbe,+sse,+rdrnd,+cx8,+sse3,+fsgsbase,+ssse3,+cx16,+fma,+popcnt,+f16c,+mmx,+sse2,+fxsr,-cldemote,-xop,-xsaves,-avx512fp16,-usermsr,-sm4,-avx512ifma,-avx512pf,-tsxldtrk,-ptwrite,-widekl,-sm3,-invpcid,-xsavec,-avx10.1-512,-avx512vpopcntdq,-avx512vp2intersect,-avx512cd,-avxvnniint8,-avx512er,-amx-int8,-kl,-avx10.1-256,-sha512,-avxvnni,-rtm,-adx,-avx2,-hreset,-movdiri,-serialize,-vpclmulqdq,-avx512vl,-uintr,-clflushopt,-raoint,-cmpccxadd,-bmi,-amx-tile,-gfni,-avxvnniint16,-amx-fp16,-xsaveopt,-avx512f,-amx-bf16,-avx512bf16,-avx512vnni,-avx512bw,-pku,-clzero,-mwaitx,-lwp,-lzcnt,-sha,-movdir64b,-wbnoinvd,-enqcmd,-prefetchwt1,-avxneconvert,-tbm,-pconfig,-amx-complex,-bmi2,-avxifma,-avx512bitalg,-rdpru,-clwb,-rdseed,-avx512vbmi2,-prefetchi,-rdpid,-fma4,-avx512vbmi,-shstk,-vaes,-waitpkg,-sgx,-avx512dq,-sse4a" * "icelake-server" "+prfchw,+avx,+aes,+sahf,+pclmul,+crc32,+sse4.1,+xsave,+sse4.2,+64bit,+cmov,+movbe,+sse,+rdrnd,+cx8,+sse3,+fsgsbase,+ssse3,+cx16,+fma,+popcnt,+f16c,+mmx,+sse2,+fxsr,-cldemote,-xop,-xsaves,-avx512fp16,-usermsr,-sm4,-avx512ifma,-avx512pf,-tsxldtrk,-ptwrite,-widekl,-sm3,-invpcid,-xsavec,-avx10.1-512,-avx512vpopcntdq,-avx512vp2intersect,-avx512cd,-avxvnniint8,-avx512er,-amx-int8,-kl,-avx10.1-256,-sha512,-avxvnni,-rtm,-adx,-avx2,-hreset,-movdiri,-serialize,-vpclmulqdq,-avx512vl,-uintr,-clflushopt,-raoint,-cmpccxadd,-bmi,-amx-tile,-gfni,-avxvnniint16,-amx-fp16,-xsaveopt,-avx512f,-amx-bf16,-avx512bf16,-avx512vnni,-avx512bw,-pku,-clzero,-mwaitx,-lwp,-lzcnt,-sha,-movdir64b,-wbnoinvd,-enqcmd,-prefetchwt1,-avxneconvert,-tbm,-pconfig,-amx-complex,-bmi2,-avxifma,-avx512bitalg,-rdpru,-clwb,-rdseed,-avx512vbmi2,-prefetchi,-rdpid,-fma4,-avx512vbmi,-shstk,-vaes,-waitpkg,-sgx,-avx512dq,-sse4a" * "sandybridge" "+avx,+aes,+sahf,+pclmul,+crc32,+sse4.1,+xsave,+sse4.2,+64bit,+cmov,+sse,+xsaveopt,+cx8,+sse3,+ssse3,+cx16,+popcnt,+mmx,+sse2,+fxsr,-prfchw,-cldemote,-xop,-xsaves,-avx512fp16,-usermsr,-sm4,-avx512ifma,-avx512pf,-tsxldtrk,-ptwrite,-widekl,-sm3,-invpcid,-xsavec,-avx10.1-512,-avx512vpopcntdq,-avx512vp2intersect,-avx512cd,-movbe,-avxvnniint8,-avx512er,-amx-int8,-kl,-avx10.1-256,-sha512,-avxvnni,-rtm,-adx,-avx2,-hreset,-movdiri,-serialize,-vpclmulqdq,-avx512vl,-uintr,-clflushopt,-raoint,-cmpccxadd,-bmi,-amx-tile,-gfni,-avxvnniint16,-amx-fp16,-rdrnd,-avx512f,-amx-bf16,-avx512bf16,-avx512vnni,-avx512bw,-pku,-fsgsbase,-clzero,-mwaitx,-lwp,-lzcnt,-sha,-movdir64b,-wbnoinvd,-enqcmd,-prefetchwt1,-avxneconvert,-tbm,-pconfig,-amx-complex,-bmi2,-fma,-avxifma,-f16c,-avx512bitalg,-rdpru,-clwb,-rdseed,-avx512vbmi2,-prefetchi,-rdpid,-fma4,-avx512vbmi,-shstk,-vaes,-waitpkg,-sgx,-avx512dq,-sse4a"

That's a good point. We probably need to tighten that up.

I'm curious to have @gbaraldi 's input on this

This is just a huge mess. some VMs hide features for example

I guess to be clear, the mess I mean is as a whole. I think this is mergeable without too many issues but I wish there was something nicer that LLVM exposed, but there isn't currently

ref: JuliaLang/julia#57049

tz-lom · 2025-11-20T15:11:52Z

src/HostCPUFeatures.jl

  reset_extra_features!()
 end
 const BASELINE_CPU_NAME = get_cpu_name()
+const allow_eval = @load_preference("allow_runtime_invalidation", false)


Why not to check that --trim is enabled in JLOptions ?

everyone who will use it with trim will have to find out this culprit themself

The main thinking is that this preference is usable by more than just JuliaC.jl, since some users may wish to be opt-in to error / warn on invalidation storms from this package. JuliaC.jl can set this preference automatically, so the end-user experience is the same.

Also technically checking JLOptions at pre-compilation time won't detect --trim properly, but something like JuliaLang/JuliaC.jl#31 would work

tz-lom · 2025-11-20T15:15:20Z

As a source of the problem I understand the approach, but it actually downgrades package HostCPUFeatures to MaybeSomeOtherHostCPUFeatures, while what we really need is to use some CompileTargetCPUFeatures in LoopVectorization etc.

Pragmatically it is "easier" upgrade standard interface of HostCPUFeatures to return features for the compile target and provide API to retrieve real Host features if that what user really wants (like printing statistics on which CPU we are running now)

topolarity · 2025-11-21T15:17:40Z

As a source of the problem I understand the approach, but it actually downgrades package HostCPUFeatures to MaybeSomeOtherHostCPUFeatures, while what we really need is to use some CompileTargetCPUFeatures in LoopVectorization etc.

That's already true if "native" is missing from your JULIA_CPU_TARGET string though - I agree that it might be a good idea to require that we are specifically operating in the non-native case though (where we already under-approximate feature flags)

Add Preference to disable runtime invalidation

723446a

This makes the package (nearly) `--trim` compatible and also provides a nice preference for users who do not wish for HostCPUFeatures.jl to ever cause an "invalidation storm", even if it means running with the wrong CPU info.

topolarity mentioned this pull request Nov 19, 2025

Add HostCPUFeatures.jl compatibility by providing allow_runtime_invalidation = false Preference JuliaLang/JuliaC.jl#80

Open

oscardssmith reviewed Nov 20, 2025

View reviewed changes

tz-lom reviewed Nov 20, 2025

View reviewed changes

Add Preference to disable runtime invalidation #22

Are you sure you want to change the base?

Add Preference to disable runtime invalidation #22

Uh oh!

Conversation

topolarity commented Nov 19, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codecov bot commented Nov 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

tz-lom commented Nov 20, 2025

Uh oh!

topolarity commented Nov 21, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

codecov bot commented Nov 20, 2025 •

edited

Loading