Re: [XEN PATCH][for-4.19 v4 1/8] xen/include: add macro ISOLATE_LOW_BIT

[Jan Beulich <jbeulich@xxxxxxxx>]

On 17.11.2023 01:43, Stefano Stabellini wrote:
> On Thu, 16 Nov 2023, Jan Beulich wrote:
>> On 16.11.2023 11:02, Nicola Vetrini wrote:
>>> On 2023-11-16 09:26, Jan Beulich wrote:
>>>> On 31.10.2023 11:20, Jan Beulich wrote:
>>>>> On 31.10.2023 11:03, Nicola Vetrini wrote:
>>>>>> On 2023-10-31 09:28, Nicola Vetrini wrote:
>>>>>>> On 2023-10-31 08:43, Jan Beulich wrote:
>>>>>>>> On 30.10.2023 23:44, Stefano Stabellini wrote:
>>>>>>>>> On Mon, 30 Oct 2023, Jan Beulich wrote:
>>>>>>>>>> On 27.10.2023 15:34, Nicola Vetrini wrote:
>>>>>>>>>>> --- a/xen/include/xen/macros.h
>>>>>>>>>>> +++ b/xen/include/xen/macros.h
>>>>>>>>>>> @@ -8,8 +8,14 @@
>>>>>>>>>>>  #define DIV_ROUND(n, d) (((n) + (d) / 2) / (d))
>>>>>>>>>>>  #define DIV_ROUND_UP(n, d) (((n) + (d) - 1) / (d))
>>>>>>>>>>>
>>>>>>>>>>> -#define MASK_EXTR(v, m) (((v) & (m)) / ((m) & -(m)))
>>>>>>>>>>> -#define MASK_INSR(v, m) (((v) * ((m) & -(m))) & (m))
>>>>>>>>>>> +/*
>>>>>>>>>>> + * Given an unsigned integer argument, expands to a mask where
>>>>>>>>>>> just the least
>>>>>>>>>>> + * significant nonzero bit of the argument is set, or 0 if no 
>>>>>>>>>>> bits
>>>>>>>>>>> are set.
>>>>>>>>>>> + */
>>>>>>>>>>> +#define ISOLATE_LOW_BIT(x) ((x) & -(x))
>>>>>>>>>>
>>>>>>>>>> Not even considering future Misra changes (which aiui may require
>>>>>>>>>> that
>>>>>>>>>> anyway), this generalization of the macro imo demands that its
>>>>>>>>>> argument
>>>>>>>>>> now be evaluated only once.
>>>>>>>>>
>>>>>>>>> Fur sure that would be an improvement, but I don't see a trivial 
>>>>>>>>> way
>>>>>>>>> to
>>>>>>>>> do it and this issue is also present today before the patch.
>>>>>>>>
>>>>>>>> This was an issue here for MASK_EXTR() and MASK_INSR(), yes, but 
>>>>>>>> the
>>>>>>>> new
>>>>>>>> macro has wider use, and there was no issue elsewhere so far.
>>>>>>>>
>>>>>>>>> I think it
>>>>>>>>> would be better to avoid scope-creeping this patch as we are 
>>>>>>>>> already
>>>>>>>>> at
>>>>>>>>> v4 for something that was expected to be a trivial mechanical 
>>>>>>>>> change.
>>>>>>>>> I
>>>>>>>>> would rather review the fix as a separate patch, maybe sent by you 
>>>>>>>>> as
>>>>>>>>> you probably have a specific implementation in mind?
>>>>>>>>
>>>>>>>> #define ISOLATE_LOW_BIT(x) ({ \
>>>>>>>>     typeof(x) x_ = (x); \
>>>>>>>>     x_ & -x_; \
>>>>>>>> })
>>>>>>>>
>>>>>>>> Hard to see the scope creep here. What I would consider scope creep 
>>>>>>>> I
>>>>>>>> specifically didn't even ask for: I'd like this macro to be
>>>>>>>> overridable
>>>>>>>> by an arch. Specifically (see my earlier naming hint) I'd like to 
>>>>>>>> use
>>>>>>>> x86's BMI insn BLSI in the context of "x86: allow Kconfig control 
>>>>>>>> over
>>>>>>>> psABI level", when ABI v2 or higher is in use.
>>>>>>>
>>>>>>> I appreciate you suggesting an implementation; I'll send a v5
>>>>>>> incorporating it.
>>>>>>
>>>>>> There's an issue with this approach, though: since the macro is used
>>>>>> indirectly
>>>>>> in expressions that are e.g. case labels or array sizes, the build 
>>>>>> fails
>>>>>> (see [1] for instance).
>>>>>> Perhaps it's best to leave it as is?
>>>>>
>>>>> Hmm. I'm afraid it's not an option to "leave as is", not the least 
>>>>> because
>>>>> - as said - I'm under the impression that another Misra rule requires
>>>>> macro arguments to be evaluated exactly once. Best I can think of 
>>>>> right
>>>>> away is to have a macro for limited use (to address such build issues)
>>>>> plus an inline function (for general use). But yes, maybe that then 
>>>>> indeed
>>>>> needs to be a 2nd step.
>>>>
>>>> While I've committed this patch (hoping that I got the necessary 
>>>> context
>>>> adjustment right for the 
>>>> automation/eclair_analysis/ECLAIR/deviations.ecl
>>>> change), I'd like to come back to this before going further with users 
>>>> of
>>>> the new macro: I still think we ought to try to get to the single
>>>> evaluation wherever possible. The macro would then be used only in 
>>>> cases
>>>> where the alternative construct (perhaps an isolate_lsb() macro, living
>>>> perhaps in xen/bitops.h) cannot be used. ISOLATE_LSB() would then want 
>>>> to
>>>> gain a comment directing people to the "better" sibling. Thoughts?
>>>
>>> Having the users in place would help me estimate the remaining work that 
>>> needs to be done on this rule and see if my local counts match up with 
>>> the counts in staging.
>>
>> By "having the users in place", you mean you want other patches in this
>> and the dependent series to be committed as-is (except for the name
>> change)? That's what I'd like to avoid, as it would mean touching all
>> those use sites again where the proposed isolate_lsb() could be used
>> instead. I'd rather see all use sites be put into their final shape
>> right away.
> 
> This request is coming a bit late and also after all the patches have
> been reviewed already. I for one am not looking forward to review them
> again.
> 
> That said, if you could be more specified maybe it could become
> actionable:
> 
> - do you have a pseudo code implementation of the "better" macro you
>   would like to propose?

May I remind you that I made this request (including a draft implementation)
before already, and Nicola then merely found that the evaluate-once form
simply cannot be used everywhere? Anybody could have thought of the option
of "splitting" the macro. After all I hope that there is no disagreement on
macro arguments better being evaluated just once, whenever possible.

> - do you have an list of call sites you would like to be changed to use
>   the "better" macro?

No, I don't have a list. But the pattern is pretty clear: The "better" form
ought to be used wherever it actually can be used.

> Also, you might remember past discussions about time spent making
> changes yourself vs. others doing the same. This is one of those cases
> that it would be faster for you to make the change and send a patch than
> explaining someone else how to do it, then review the result (and
> review it again as it probably won't be exactly as you asked the first
> time.)
> 
> If you don't want the call sites to be changes twice, may I suggest you
> provide a patch on top of Nicola's series, I review and ack your patch,
> and Nicola or I rebase & resend the series so that the call sites are
> only changes once as you would like? I think that's going to be way
> faster.

I'll see if I can find time to do so. I don't normally work on top of
other people's uncommitted patches, though ... So I may also choose to go
a slightly different route. (You realize though that all still pending
patches using the new macro need touching again anyway, don't you?)

Jan