Re: [PATCH v5 15/18] whpx: arm64: implement -cpu host

[Pierrick Bouvier <pierrick.bouvier@linaro.org>]

On 8/19/25 9:40 AM, Daniel P. Berrangé wrote:
> On Tue, Aug 19, 2025 at 09:25:22AM -0700, Pierrick Bouvier wrote:
>> On 8/19/25 9:08 AM, Daniel P. Berrangé wrote:
>>> On Tue, Aug 19, 2025 at 08:48:16AM -0700, Pierrick Bouvier wrote:
>>>> On 8/19/25 8:12 AM, Daniel P. Berrangé wrote:
>>>>> On Tue, Aug 19, 2025 at 04:06:45PM +0100, Peter Maydell wrote:
>>>>>> On Tue, 19 Aug 2025 at 16:04, Pierrick Bouvier
>>>>>> <pierrick.bouvier@linaro.org> wrote:
>>>>>>>
>>>>>>> On 8/19/25 6:24 AM, Peter Maydell wrote:
>>>>>>>> On Fri, 8 Aug 2025 at 07:55, Mohamed Mediouni <mohamed@unpredictable.fr> wrote:
>>>>>>>> Can you follow the QEMU coding style, please (here and elsewhere)?
>>>>>>>> Variables and function names should be all lower case,
>>>>>>>> and variable declarations go at the start of a C code
>>>>>>>> block, not in the middle of one.
>>>>>>>>
>>>>>>>
>>>>>>> In some cases, including in this function, I feel that the rule to
>>>>>>> declare variables at the start of a block is not really helpful, and is
>>>>>>> more related to legacy C than a real point nowadays.
>>>>>>> As well, it sometimes forces to reuse some variables between various sub
>>>>>>> blocks, which definitely can create bugs.
>>>>>>>
>>>>>>> Anyway, I'm not discussing the existing QEMU coding style, but just
>>>>>>> asking if for the current context, is it really a problem to declare
>>>>>>> variable here?
>>>>>>
>>>>>> The point of a coding style is to aim for consistency. QEMU
>>>>>> is pretty terrible at being consistent, but we should try.
>>>>>> The rule about variables at start of block is not because
>>>>>> some compilers fail to compile it, but because we think
>>>>>> it's overall more readable that way.
>>>>>
>>>>> There are also potential[1] functional problems with not declaring
>>>>> at the start of block, because if you have a "goto cleanup" which
>>>>> jumps over the line of the declaration, the variable will have
>>>>> undefined state when the 'cleanup:' block is running. This is
>>>>> something which is very subtle and easily missed when reading the
>>>>> code flow.
>>>>>
>>>>
>>>> This has nothing to do with where variables are declared, but where they are
>>>> assigned. The same issue can happen whether or not it's declared at the
>>>> start of a block.
>>>>
>>>> I suspect we use -ftrivial-auto-var-init precisely because we force
>>>> variables to be declared at start of the scope, i.e. where they don't have
>>>> any value yet. So, instead of forcing an explicit initialization or rely on
>>>> compiler warnings for uninitialized values, it was decided to initialize
>>>> them to 0 by default.
>>>>
>>>> If we declared them at the point where they have a defined semantic value,
>>>> this problem would not exist anyway, out of the goto_cleanup situation,
>>>> which has the same fundamental issue in both cases.
>>>
>>> It really isn't the same issue when you compare
>>>
>>>     void bar(void) {
>>>       char *foo = NULL;
>>>
>>>       if (blah)
>>>          goto cleanup:
>>>
>>>     cleanup:
>>>       if (foo)
>>>          ....
>>>     }
>>>
>>> vs
>>>
>>>     void bar(void) {
>>>       if (blah)
>>>          goto cleanup:
>>>
>>>       char *foo = NULL;
>>>
>>>       ...some code...
>>>
>>>     cleanup:>      if (foo)
>>>          ....
>>>     }
>>>
>>> The late declaration of 'foo' is outright misleading to reviewers.
>>>
>>> Its initialization at time of declaration gives the impression
>>> that 'foo' has well defined value in the 'cleanup' block, when
>>> that is not actually true. In big methods it is very easy to
>>> overlook an earlier 'goto' that jumps across a variable declaration
>>> and initialization.
>>>
>>
>> "Big" method is probably the issue there. If it's not possible to follow
>> control flow in a given function, it's a strong hint there is a problem with
>> its size, independently of any standard.
> 
> Certainly some methods are too big & deserve refactoring, but that's a
> non-trivial investment, and it isn't always a clearcut win to split
> code out into a bunch of arbitrarily short methods. You may solve the
> goto/initialization problem, but make other things harder as you often
> still have to fully page all the code into mind to understand it.
>

I am still looking for an example of where breaking down a big function 
in smaller logical chunks has reduced the readability for anyone, but I 
never met one so far in my professional life.
Usually it comes with the additional benefit that you need to *name* 
things explicitely, which is usually better than add a comment about them.
The point of breaking down code is explicitely to remove the need to 
keep things in mind and assume functions do what they are named for.

I respect the difference about tastes concerning readability and code 
structure, and I know the context and era from which QEMU codebase comes 
from.

However, arguing that variables should be at the start of a block 
because of a potential goto_cleanup situation is not a good argument.

>>> Even if not all methods have this problem, the coding standards
>>> guide us into the habit of writing code that is immune from this
>>> kind of problem. That habit only forms reliably if we apply the
>>> coding standards unconditionally, rather than selectively.
>>>
>>
>> That's right, but humanly enforced coding standard are usually a waste of
>> time for everyone (reviewers and developers).
> 
> Human enforced standards are absolutely better than a free-for-all. Over
> time contributors will gain familiarity with the project standards and
> largely comply without enforcement being required. If contributors
> repeatedly ignore coding standards, it will disincentivise reviewers
> from looking at their patches.
>

Sure.
As well, incessant pushbacks and nitpicking from those same reviewers 
can disincentivise people to send any patch.
But maybe the whole point is simply to keep people out of their lawn, or 
reduce the amount of patches they need to process daily, who knows.

>> How many messages and exchanges on the mailing list could we save by using
>> something like clang-format on the codebase, and force it to be "clean" as
>> part of the CI? There would be no more discussion, as there would be only
>> one single and objective source of truth.
> 
> I would really love if it we could apply clang-format to everything, but
> that has a non-trivial impact on maint when done on a large pre-existing
> codebase like QEMU. Cherry-picking to upstream stable or distros would
> be immensely painful, verging on impossible, after a bulk reformat. For
> any new codebase I'd go for clang-format every time.
>

Maybe we could organize a conversation about this, because the benefits 
are worth making cherry-picking a little bit harder. In this case, all 
the community benefits from this, while blocking this penalizes everyone 
except the stable maintenance part and downstream forks.
As well, it would be a once in a lifetime price to pay.

> With regards,
> Daniel