[cfe-dev] [RFC] C++17 hardware constructive / destructive interference size

[David Blaikie via cfe-dev]

[+echristo because he's been thinking about some of these things
(especially those highlighted in (1)) since implementing the target
attribute support & looking at how to build code optimized for specific
subtargets]

On Wed, May 30, 2018 at 12:39 PM James Y Knight via cfe-dev <
cfe-dev at lists.llvm.org> wrote:

>
>
> On Tue, May 29, 2018 at 4:39 PM JF Bastien <jfbastien at apple.com> wrote:
>
>>
>> On May 25, 2018, at 3:12 PM, James Y Knight <jyknight at google.com> wrote:
>>
>>
>>
>> On Fri, May 25, 2018 at 5:30 PM JF Bastien <jfbastien at apple.com> wrote:
>>
>>>
>>>
>>> On May 25, 2018, at 2:23 PM, James Y Knight <jyknight at google.com> wrote:
>>>
>>> My own employer doesn't make ABI stability promises for that code, and
>>> thus is fine with changing the value anytime it feels like. That's not a
>>> generically viable strategy for a value provided by the standard library.
>>>
>>> Additionally, before I sent that email, I looked at a number of the
>>> uses, and it appeared as though a great many could be easily modified to
>>> use a runtime-determined alignment.
>>>
>>>
>>> That would be useful feedback on the paper… prior to it getting into
>>> C++17. The committee’s POV voting the paper in was that having a constexpr
>>> value was something we wanted, and so that’s what we have. At this point in
>>> time I’d like to focus on implementing C++17 as it is, and / or filing DRs
>>> as required.
>>>
>>
>> Sure. I'm not on the committee. Even if I was, I certainly don't know
>> that I would have identified the problem...
>>
>> But now that it has been identified, there's a choice of what to do. And
>> not implementing the function (and presumably filing a DR saying so) is
>> seeming like a pretty reasonable option.
>>
>>
>> The committee discussed ABI issues (Jacksonville 2016) and decided that
>> they’d rather have them than have a proliferation of #define SOMETHING 64.
>> That discussion occurred with Google folks in the room, it might be higher
>> bandwidth to consult with them? The notes are unfortunately quite sparse
>> for that discussion.
>>
>> The libc++ community shouldn't decline to implement a feature without
>> bringing concrete feedback to the committee. Without such feedback, I’d
>> like to move forward with an implementation plan, because we should offer
>> full C++17 support regardless of our distaste for specific features. I’ve
>> received good feedback on the thread so far, I’m happy to leave the
>> discussion open for a bit, talk to committee people next week in
>> Rapperswil, and unless feedback goes the committee’s way I’d like to pursue
>> an implementation. Does this sound fair?
>>
>
> There's been 3 options discussed so far -- I'm not sure which (#1 or #2)
> you're now proposing to implement.
>
> 1. Return an subtarget-dependent value, depending on the exact CPU model
> selected at compile-time.
>   Good: Allows for better memory-usage/performance.
>   Bad: Potential risk of ODR violations/ABI issues, due to dependency on
> cpu tuning flags.
>   Bad: Potential risk of same across versions of the compiler, if the
> default generic cpu tuning is changed.
>
> 2. Choose a single "good enough" constant value for each platform.
>   Good: eliminate ABI/ODR issues.
>   Bad: value might be too conservative for users' desires.
>     e.g. returning 128 for hardware_destructive_interference_size when 64
> would've been sufficient will waste memory.
>   Bad: Future CPU changes might invalidate the constant generic value,
> requiring either that it be changed (introducing an ABI issue again), or
> remain incorrect forever.
>     e.g. most ARM chips have had 64-byte cache-lines for a while now, so
> that would've seemed the only reasonable number to choose on ARM up until
> recently. But, now, some of the newest CPUs have apparently switched to
> 128-byte cache-lines; should we change to 128?
>
> (Or, 2b: YOLO, 64 bytes should be good enough for all platforms!)
>
> 3. Decline to implement at all.
>   Good: avoid these issues.
>   Bad: users who need it must do something themselves, e.g. choose some
> arbitrary value e.g. 64.
>
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20180531/024aae54/attachment.html>