Date: Mon, 22 Jul 2024 13:19:07 -0300
On Mon, Jul 22, 2024 at 8:51 AM connor horman via Std-Proposals
<std-proposals_at_[hidden]> wrote:
>
> LL/SC is also hard to abstract over. I don't know how it is for ARM, but on RISC-V there are very strong restrictions about what you can actually do between the LL and SC (or the SC will uncontionally fail) - one of which is no branches, so the abstraction basically requires inlining to happen (or otherwise, is allowed to reliably fail for no reason - not fail spuriously, because spurious implies that the failure isn't reliable).
>
That applies for ARM as well.
Moreover, the amount of cycles spent between the LL and SC will
dramatically change how likely the SC is to succeed in a highly
contested environment.
> On Mon, 22 Jul 2024 at 07:28, Jonathan Wakely via Std-Proposals <std-proposals_at_[hidden]> wrote:
>>
>>
>>
>> On Mon, 22 Jul 2024 at 08:00, Jens Maurer via Std-Proposals <std-proposals_at_[hidden]> wrote:
>>>
>>>
>>>
>>> On 22/07/2024 00.23, Joseph Schuchart via Std-Proposals wrote:
>>> > I'm not sure that this API could be implemented everywhere
>>>
>>> And here's the problem: I don't know how you can provide ABA
>>> avoidance if all your hardware has is atomic compare-and-swap.
>>>
>>> So, this doesn't really help write portable programs.
>>>
>>> We have rather few features in C++ that aren't available on all
>>> implementations, so the introduction of such a non-portable
>>> feature needs additional rationale.
>>
>>
>> And for better or for worse, "not supported on x86" is likely to need a lot more rationale than "not supported on AVR" (for example).
>>
>>
>>>
>>>
>>> (There are plenty of hardware features that aren't available
>>> everywhere, yet we strive to make C++ interfaces on them that
>>> can be implemented everywhere, e.g. std::simd.)
>>>
>>> Jens
>>
>> --
>> Std-Proposals mailing list
>> Std-Proposals_at_[hidden]
>> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
>
> --
> Std-Proposals mailing list
> Std-Proposals_at_[hidden]
> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
<std-proposals_at_[hidden]> wrote:
>
> LL/SC is also hard to abstract over. I don't know how it is for ARM, but on RISC-V there are very strong restrictions about what you can actually do between the LL and SC (or the SC will uncontionally fail) - one of which is no branches, so the abstraction basically requires inlining to happen (or otherwise, is allowed to reliably fail for no reason - not fail spuriously, because spurious implies that the failure isn't reliable).
>
That applies for ARM as well.
Moreover, the amount of cycles spent between the LL and SC will
dramatically change how likely the SC is to succeed in a highly
contested environment.
> On Mon, 22 Jul 2024 at 07:28, Jonathan Wakely via Std-Proposals <std-proposals_at_[hidden]> wrote:
>>
>>
>>
>> On Mon, 22 Jul 2024 at 08:00, Jens Maurer via Std-Proposals <std-proposals_at_[hidden]> wrote:
>>>
>>>
>>>
>>> On 22/07/2024 00.23, Joseph Schuchart via Std-Proposals wrote:
>>> > I'm not sure that this API could be implemented everywhere
>>>
>>> And here's the problem: I don't know how you can provide ABA
>>> avoidance if all your hardware has is atomic compare-and-swap.
>>>
>>> So, this doesn't really help write portable programs.
>>>
>>> We have rather few features in C++ that aren't available on all
>>> implementations, so the introduction of such a non-portable
>>> feature needs additional rationale.
>>
>>
>> And for better or for worse, "not supported on x86" is likely to need a lot more rationale than "not supported on AVR" (for example).
>>
>>
>>>
>>>
>>> (There are plenty of hardware features that aren't available
>>> everywhere, yet we strive to make C++ interfaces on them that
>>> can be implemented everywhere, e.g. std::simd.)
>>>
>>> Jens
>>
>> --
>> Std-Proposals mailing list
>> Std-Proposals_at_[hidden]
>> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
>
> --
> Std-Proposals mailing list
> Std-Proposals_at_[hidden]
> https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
Received on 2024-07-22 16:19:19