Proposal: Allow std::hardware_(destructive|constructive)_interference_size

[Anton Bikineev]

The constants (cppreference) are useful to avoid false sharing. They're easier to use than relying on platform-specific cacheline-size macros or sysconf or whatnot. Should be safe to use.

(The story: after profiling false sharing hits with 'perf c2c' I found some places in Chromium where the constants could be useful).

[Jeremy Roman]

Makes sense to me. I would expect usage to be extremely niche, but I don't think risk of misuse is particularly high.

On Fri, Sep 16, 2022 at 4:00 AM 'Anton Bikineev' via cxx <c...@chromium.org> wrote:

The constants (cppreference) are useful to avoid false sharing. They're easier to use than relying on platform-specific cacheline-size macros or sysconf or whatnot. Should be safe to use.

(The story: after profiling false sharing hits with 'perf c2c' I found some places in Chromium where the constants could be useful).

--
You received this message because you are subscribed to the Google Groups "cxx" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cxx+uns...@chromium.org.
To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/cxx/CAFd1ZwcJkRUt40gn%3DVm-Ud%2B3G1bWvuhw-3hxsRh8Yddg--j8Jw%40mail.gmail.com.

[Peter Kasting]

Lgtm

On Fri, Sep 16, 2022, 4:00 AM 'Anton Bikineev' via cxx <c...@chromium.org> wrote:

The constants (cppreference) are useful to avoid false sharing. They're easier to use than relying on platform-specific cacheline-size macros or sysconf or whatnot. Should be safe to use.

(The story: after profiling false sharing hits with 'perf c2c' I found some places in Chromium where the constants could be useful).

--

[K. Moon]

Have we gotten to the point yet where we should be default-allowing C++17 features? Or no?

To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/cxx/CAAHOzFCqmMwV_GiOQ7Em-rDRJ-bofAdby2L4-s0%2BkKECDB0a%3Dw%40mail.gmail.com.

[Peter Kasting]

No, that's still a year out.

PK

[Peter Kasting]

Update: I discovered today that libc++ does not implement these for clang, only for gcc. I will be moving these to the banned list for now. Best workaround I know of is ABSL_CACHELINE_SIZE in third_party/abseil-cpp/absl/base/optimization.h.

PK

[Peter Kasting]

Update: The version of clang in Chromium should now support the macros libc++ uses to implement these features. Accordingly, I think it should be safe to unban-and-convert-to these (though I haven't double-checked yet).

PK

[Peter Kasting]

On Monday, June 10, 2024 at 5:09:30 PM UTC-7 Peter Kasting wrote:

Update: The version of clang in Chromium should now support the macros libc++ uses to implement these features. Accordingly, I think it should be safe to unban-and-convert-to these (though I haven't double-checked yet).

Update update: The saigo toolchain, at least, is not yet rolled here. Googlers can follow https://b.corp.google.com/issues/40210191.

PK