RFC follow-up: attribute to suppress -Wvla on a single declaration - relationship to lva-larger-then

[David Brown]

On 28/02/2026 15:41, Martin Uecker via Gcc wrote:
> Am Freitag, dem 27.02.2026 um 20:54 +0100 schrieb David Brown:
>> On 27/02/2026 19:57, Martin Uecker via Gcc wrote:
>>> Am Freitag, dem 27.02.2026 um 19:40 +0100 schrieb David Brown:
>>>> On 27/02/2026 18:51, Martin Uecker via Gcc wrote:
>>
>> <snip>
>>
>>>> No, it does not trigger a warning with those values.  (It will if
>>>> someone changes MAX_POINTS to 250000, but that's the point of the
>>>> -Wvla-larger-than=512000.)
>>>
>>>
>>> You are right, thank you. I was somehow confusing this with
>>> something else.
>>>
>>
>> You showed that code in your other post:
>>
>> 	void foo(int n)
>> 	{
>> 	  if (n > 100) return;
>> 	  char buf[n];
>> 	  foo(n);
>> 	}
>>
>> I would say that definitely deserves a warning.  Either make the
>> parameter unsigned, or check if for negative values.
> 
> In any case, the warning one gets
> 
>   warning: argument to variable-length array may be too large
> 
> is then a bit confusing.
> 
> But the problem is that I like to use signed types to be able to use the
> sanitizer to find overflow eerrors, and I can also use the sanitizer to find
> negative sizes for VLAs.  So I peronally find the warning for the negative
> part unhelpful.
> 

It is always much better if a static warning can find a problem than to 
wait for the sanitizer!  But I agree the message here is a bit 
confusing.  It appears that the warning treats the array size as an 
unsigned type.  If you try setting n to -1, you can see in the warning 
it gives exactly how big it thinks n is.

There are lots of reasons to choose signed or unsigned types.  But when 
you want to be sure that your "n" is in the range 1 to 100, you really 
should check that range - not just the upper bound.


>>>> So there is scope for improvement on the wording of the warnings here!
>>>
>>> I agree.  Do you have a suggestion?
>>>
>>> Martin
>>>
>>
>> How about just :
>>
>> 	warning: Variable length array 'v' [-Wvla]
>>
>> You will only see the warning if you specifically enable it - you are
>> getting the warning because /you/ choose to enable the warning, not
>> because ISO C90 forbids it.  (And ISO C90 does not "forbid" VLAs - it
>> just doesn't have the concept.)
>>
>> The same message could be given if "n" is "const" - it does not seem to
>> me that there should be a difference here.
>>
>> Alternatively, you could have two levels of the warning - -Wvla=1 which
>> warns on VLAs with a size not known until runtime, and -Wvla=2 which
>> also warns on arrays with sizes known at compile-time but which are
>> still VLAs in C.  (Perhaps following the C++ rules, to give an
>> optimisation-independent definition.)  For consistency, -Wvla alone
>> would default to -Wvla=2.
>>
>> Then you would have two warnings :
>>
>> warning: Dynamically sized variable length array [-Wvla=1]
>>
>> warning: Const sized array is variable length array in C [-Wvla=2]
> 
> Makes sense!
> 
> Note there is a proposal to make those constants true integer constants:
> https://www.open-std.org/jtc1/sc22/wg14/www/docs/n3693.htm
> 

That seems a good idea to me.  It works fine for C++, and should also 
work fine here for C.