Re: [PATCH] target/arm: Provide always-false kvm_arm_*_supported() stubs for usermode

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

On 7/17/25 10:12 AM, Peter Maydell wrote:
> On Thu, 17 Jul 2025 at 18:05, Pierrick Bouvier
> <pierrick.bouvier@linaro.org> wrote:
>> On 7/17/25 9:56 AM, Peter Maydell wrote:
>>> I'm not really a fan of relying on the compiler to fold stuff
>>> away -- it's fragile and there's no guarantee the compiler
>>> will actually do it. In this example it would be really easy
>>> for somebody coming along to tidy this up later to put the
>>> nested if()s back into a single if() condition and reintroduce
>>> the problem, for instance.
>>>
>>
>> There are various places where we already relied on that, including
>> before the single-binary work ({accel}_allowed).
>>
>> For the fragile aspect of it, that's why CI exists. Building all
>> binaries with clang --enable-debug should ensure no regression can be
>> introduced.
> 
> But our CI can't check all versions of gcc and clang at
> all optimisation levels. The reason it's fragile is because
> there's never a guarantee that the compiler does it or what
> situations it will work vs not work. It's not a feature
> provided by the language or the compiler that you can fold
> away references to functions that don't exist; it just
> happens to work.
>

Sure, it's always a compromise, and it's overkill to try to cover all 
the {os, compiler, version} possibilities.

That said, checking debug builds with available clang version in latest 
ubuntu LTS + latest clang version published should already cover a lot 
of cases, including new warnings that might appear.

I would be happy to help contribute on QEMU CI side, but in my personal 
experience trying to approach concerned developers, either I didn't have 
any answer, or received a "No, our minutes are precious".
I know your point of view (Engineers are already paid, minutes are not), 
but it still doesn't make sense to me. I hope one day we can move away 
from this paradigm, and really make a good usage of cpu time instead of 
people time for testing things.

> I know we already rely on constant-folding in various
> places: I'm not super enthusiastic about those either.
> 
> -- PMM