Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

CPU feature detection in core #3469

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

Conversation

Amanieu
Copy link
Member

@Amanieu Amanieu commented Aug 3, 2023

Rendered

This RFC moves the is_*_feature_detected macros into core, but keeps the logic for actually doing feature detection (which requires OS support) in std.

@ehuss ehuss added the T-libs-api Relevant to the library API team, which will review and decide on the RFC. label Aug 4, 2023
@programmerjake
Copy link
Member

programmerjake commented Aug 4, 2023

I think it would be useful to have the is_*_feature_detected macros instead run code like:

macro_rules! is_x86_feature_detected {
    ("sse") => { feature_detect(X86Feature::SSE) };
    ("sse2") => { feature_detect(X86Feature::SSE2) };
    ("avx") => { feature_detect(X86Feature::AVX) };
    ("avx2") => { feature_detect(X86Feature::AVX2) };
    // ...
}
#[repr(u16)]
pub enum X86Feature {
    // arbitrary order since I didn't bother to look it up,
    // we'd probably want to base these off `cpuid` bit positions
    // or something similar, since that makes `rust_cpu_feature_detect` much simpler
    SSE,
    SSE2,
    AVX,
    AVX2,
    // ... lots of elided features
    AVX512F, // assume this is the last one
}
impl X86Feature {
    pub const MAX: Self = Self::AVX512F; // assume this is the last one
}
#[derive(Copy, Clone, Eq, PartialEq, Debug, Default)]
pub struct X86Features(pub [usize; Self::ARRAY_SIZE]);
impl X86Features {
    pub const ARRAY_SIZE: usize = X86Feature::MAX as usize / usize::BITS as usize + 1;
}
extern "Rust" {
    // this should have some kind of weak linking or something
    fn rust_cpu_feature_detect() -> X86Features;
}
#[inline]
pub fn feature_detect(feature: X86Feature) -> bool {
    const Z: AtomicUsize = AtomicUsize::new(0);
    static CACHE: [AtomicUsize; X86Features::ARRAY_SIZE] = [Z; X86Features::ARRAY_SIZE];
    static CACHE_VALID: AtomicBool = AtomicBool::new(false);
    #[cold]
    fn fill_cache() {
        for (cache, v) in CACHE.iter().zip(&unsafe { rust_cpu_feature_detect() }.0) {
            cache.store(*v, Ordering::Relaxed);
        }
        CACHE_VALID.store(true, Ordering::Release);
    }
    // intentionally only use atomic store/load to avoid needing cmpxchg or similar for cpus without support for that
    if !CACHE_VALID.load(Ordering::Acquire) {
        fill_cache();
    }
    let index = feature as usize;
    let bit = index % usize::BITS as usize;
    let index = index / usize::BITS as usize;
    (CACHE[index].load(Ordering::Relaxed) >> bit) & 1 != 0
}

@Lokathor
Copy link
Contributor

Lokathor commented Aug 4, 2023

That's approximately how std_detect works already, could you explain what in particular is different about your example code?

@programmerjake
Copy link
Member

programmerjake commented Aug 4, 2023

That's approximately how std_detect works already, could you explain what in particular is different about your example code?

it's not what @Amanieu proposed? in particular the code has the initialization driven by the threads calling is_*_feature_detected instead of depending on main or similar to fill it in and hope that's before you need it.

it also needs much less cache space for platforms without atomic fetch_or

@Lokathor
Copy link
Contributor

Lokathor commented Aug 4, 2023

Ah, well when you put it like that I see the difference :3


## Using a lang item to call back into `std`

Instead of having `std` "push" the CPU features to `core` at initialization time, an alternative design would be for `core` to "pull" this information from `std` by calling a lang item defined in `std`. The problem with this approach is that it doesn't provide a clear path for how this would be exposed to no-std programs which want to do their own feature detection.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A lang item with two "grades" - weak and strong, core defines a weak version of the lang item so users are not required to provide their own lang item definition, but anyone can override it with a strong version of the same lang item (libstd will do that as well).

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(A lang item can have a stable alias like #[panic_handler].)

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Weak lang items is already used as name for lang items like #[panic_handler] that need not be defined when it is used, but does need to be defined when linking if it was used anywhere. Maybe use preemptible lang items as name here?

text/0000-core_detect.md Outdated Show resolved Hide resolved
text/0000-core_detect.md Outdated Show resolved Hide resolved
text/0000-core_detect.md Outdated Show resolved Hide resolved
@Amanieu
Copy link
Member Author

Amanieu commented Aug 4, 2023

I think it would be useful to have the is_*_feature_detected macros instead run code like:

This is what I described in alternative designs here.

The main issues are:

  • this requires a much larger API surface, which makes stabilization more difficult.
  • this is actually slightly slower since it requires an atomic check that the CPU features have been initialized.

it also needs much less cache space for platforms without atomic fetch_or

This could also be solved by requiring that mark_*_feature_as_detected is only called before multiple threads access the CPU features (since it's unsafe anyways). This works for static initializers, even for dynamic libraries, since they are executed before any other code.

@programmerjake
Copy link
Member

I think it would be useful to have the is_*_feature_detected macros instead run code like:

This is what I described in alternative designs here.

The main issues are:

  • this requires a much larger API surface, which makes stabilization more difficult.

we don't have to have all that API surface stabilized, we can just stabilize the existence of rust_cpu_feature_detect and X86Features with just ARRAY_SIZE and the field .0 and stabilizing the location of each feature bit (copying cpuid locations would make that relatively uncontroversial), which imho is a rather small API surface for what it does.

  • this is actually slightly slower since it requires an atomic check that the CPU features have been initialized.

we'll likely want an atomic check anyway so we can be sure all the features are correctly synchronized with each other.
e.g. if we check is_aarch64_feature_detected!("v8.1a") followed by is_aarch64_feature_detected!("v8.2a") then, assuming the function computing the cpu features always returns the same value within each program run, !v8_1a && v8_2a should never be true. if we don't have a acquire-load from the same location as the function computing cpu features release-stored to after writing all features to memory, then we can end up with inconsistent features since relaxed-loads to different locations can be reordered with each other.

it also needs much less cache space for platforms without atomic fetch_or

This could also be solved by requiring that mark_*_feature_as_detected is only called before multiple threads access the CPU features (since it's unsafe anyways). This works for static initializers, even for dynamic libraries, since they are executed before any other code.

the problem with that is that users are often specifically trying to avoid static initializers due to ordering issues and stuff:
rust-lang/rust#111921

@lyphyser
Copy link

There's also the option of including all detection code in libcore, using raw system calls in assembly if OS interaction is needed.

It makes libcore OS-specific, but I think that might be fine as long as it doesn't use any library.

@bjorn3
Copy link
Member

bjorn3 commented Aug 10, 2023

Apart from Linux most OSes don't allow direct syscalls. Go tried, but had to revert back to using libc on most platforms after macOS changed the ABI of one syscall breaking every Go program in existence and OpenBSD added an exploit mitigation that makes your process crash if you try to do a syscall while outside of libc.

@lyphyser
Copy link

lyphyser commented Aug 10, 2023

Apart from Linux most OSes don't allow direct syscalls. Go tried, but had to revert back to using libc on most platforms after macOS changed the ABI of one syscall breaking every Go program in existence and OpenBSD added an exploit mitigation that makes your process crash if you try to do a syscall while outside of libc.

I guess on those systems libcore can just link the C library. After all, it seems that no useful program can exist on those systems without linking the C library (since without system calls, the only things a program can do are to loop infinitely, waste a number of CPU cycles and then crash, or try to exploit the CPU or kernel), so might as well link it from libcore.

There's also the issue that in some cases CPU feature detection needs to access the ELF auxiliary values, but argv and envp are passed to static initializers, and the ELF auxiliary entries are guaranteed to be after the environment pointers which are after the argument pointers (at least on x86-64, but I guess all ABIs are like that), so there should be no need to call getauxval to access them.

@YurySolovyov
Copy link

might be worth thinking about something like AVX10 which uses point versioning like 10.1, 10.2, etc.

@the8472
Copy link
Member

the8472 commented Aug 21, 2023

AVX10 will be more complicated than that because cores can have different capabilities. We currently have no way to represent those differences. We don't even have thread pinning in std which would be a prerequisite to make that work.

@YurySolovyov
Copy link

I thought the whole idea of AVX10 was to give cores identical capabilities 🤔

text/0000-core_detect.md Outdated Show resolved Hide resolved
@the8472
Copy link
Member

the8472 commented Aug 22, 2023

I thought the whole idea of AVX10 was to give cores identical capabilities 🤔

https://cdrdv2.intel.com/v1/dl/getContent/784267 Introduction section on page 14 as of revision 2.0 of the document

The same instructions will be supported across all cores. But maximum register width may vary across cores. Specifically P-cores may support ZMM registers while E-cores may be limited to YMM.

And in the CPUID table on page 15 I'm not seeing anything to query the minimum set across all cores instead of the current CPU...

@Amanieu Amanieu added the I-libs-api-nominated Indicates that an issue has been nominated for prioritizing at the next libs-api team meeting. label Jul 23, 2024
@m-ou-se
Copy link
Member

m-ou-se commented Jul 30, 2024

@rfcbot merge

@rfcbot
Copy link
Collaborator

rfcbot commented Jul 30, 2024

Team member @m-ou-se has proposed to merge this. The next step is review by the rest of the tagged team members:

No concerns currently listed.

Once a majority of reviewers approve (and at most 2 approvals are outstanding), this will enter its final comment period. If you spot a major issue that hasn't been raised at any point in this process, please speak up!

See this document for info about what commands tagged team members can give me.

@rfcbot rfcbot added proposed-final-comment-period Currently awaiting signoff of all team members in order to enter the final comment period. disposition-merge This RFC is in PFCP or FCP with a disposition to merge it. labels Jul 30, 2024
@joshtriplett
Copy link
Member

Could this RFC please add an unresolved question about the efficiency of having (say) a hundred of these at startup, and whether this can be reasonably optimized?

@rfcbot rfcbot added the final-comment-period Will be merged/postponed/closed in ~10 calendar days unless new substational objections are raised. label Jul 30, 2024
@rfcbot
Copy link
Collaborator

rfcbot commented Jul 30, 2024

🔔 This is now entering its final comment period, as per the review above. 🔔

@rfcbot rfcbot removed the proposed-final-comment-period Currently awaiting signoff of all team members in order to enter the final comment period. label Jul 30, 2024
@Amanieu Amanieu removed the I-libs-api-nominated Indicates that an issue has been nominated for prioritizing at the next libs-api team meeting. label Aug 6, 2024
@rfcbot rfcbot added finished-final-comment-period The final comment period is finished for this RFC. and removed final-comment-period Will be merged/postponed/closed in ~10 calendar days unless new substational objections are raised. labels Aug 9, 2024
@rfcbot
Copy link
Collaborator

rfcbot commented Aug 9, 2024

The final comment period, with a disposition to merge, as per the review above, is now complete.

As the automated representative of the governance process, I would like to thank the author for their work and everyone else who contributed.

This will be merged soon.

@programmerjake
Copy link
Member

programmerjake commented Aug 9, 2024

in unresolved questions, can you add the problem that some features imply other features (e.g. avx2 implies avx) and we should try to make updates synchronized so we don't see a transient state where avx2 is enabled but avx is disabled.

(I would have posted this on the tracking issue, but there isn't one yet...)

@Lokathor
Copy link
Contributor

Lokathor commented Aug 9, 2024

The implementation of feature detection by std sounds separate from if core-only code can call upon feature detection and expect other linked code to provide it.

@programmerjake
Copy link
Member

The implementation of feature detection by std sounds separate from if core-only code can call upon feature detection and expect other linked code to provide it.

If this is in reply to #3469 (comment), I meant that the proposed implementation is just reading from some atomics in core, so we should synchronize updates to those atomics to prevent transient inconsistent features.

@clarfonthey
Copy link
Contributor

Feels like multiple problems could be solved at once if the features were represented using some kind of bitflags-type struct that was applied over a slice of atomics. Things like transitive features (avx2 -> avx) are pretty naturally representable this way, since you can just make each feature set the bits it affects. If you make these relations between features clear to the people who need to set them, they should in theory be able to do a minimum number of set calls, rather than say, setting 100 flags at once like people were worried about.

@Lokathor
Copy link
Contributor

@programmerjake yes that was my intention.

My point is that this RFC is extending an existing and stable part of std to also work in core, and that's all it's doing. Any concerns about how the std detection system works should be tracked separately in other issues or rfcs or whatever appropriate location.

@programmerjake
Copy link
Member

@programmerjake yes that was my intention.

My point is that this RFC is extending an existing and stable part of std to also work in core, and that's all it's doing.

no, it's also adding the core::arch::mark_*_feature_as_detected APIs...my concern is about how those are implemented, since if they just naively fetch_or the one bit for each feature and don't think about which order to set the features in, it's very easy to end up reading avx2 as true since it happened to be set first but avx as false since that was set slightly later.

@clarfonthey
Copy link
Contributor

clarfonthey commented Aug 10, 2024

@programmerjake yes that was my intention.
My point is that this RFC is extending an existing and stable part of std to also work in core, and that's all it's doing.

no, it's also adding the core::arch::mark_*_feature_as_detected APIs...my concern is about how those are implemented, since if they just naively fetch_or the one bit for each feature and don't think about which order to set the features in, it's very easy to end up reading avx2 as true since it happened to be set first but avx as false since that was set slightly later.

Right, which is why I mentioned bitfields. If your atomics are integers instead of bools, then you can just set both bits at once and it works as expected.

Of course, this does require explicitly specifying these interdependencies between features, but I'd argue we should be doing that even without this RFC. It doesn't make sense for avx to be ever unavailable when avx2 is available, and I'd imagine many people already rely on this.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
disposition-merge This RFC is in PFCP or FCP with a disposition to merge it. finished-final-comment-period The final comment period is finished for this RFC. T-libs-api Relevant to the library API team, which will review and decide on the RFC. to-announce
Projects
None yet
Development

Successfully merging this pull request may close these issues.