Hi Martin,
What criteria do we use in our team to allow/disallow.
The main advantage of using VLA (over heap memory) is performance. So
that questions that we ask during the view:
1. Is the function in "hot path" - executed repeatedly
2. What is the maximum memory that will be needed
3. In tight-loops - memory locality can be a factor
For example: this will get approved:
double foo_slices(int n, double *a)
{
double tmp[n] ; // need temporary array
... Use tmp
return ... ;
}
int bar(...)
{
for (int i=0 ; i<nnn ; i++) foo_slices(v_len[i], v_data[ii]) ;
}
In the above example - calling malloc in foo_slices can slow the
calculation, vs using VLA. As I mentioned, we found the performance
gain coming from both the speed of the allocation, and from the fact
that memory is usually allocated in L1/L2, whereas with malloc, you
may get L3.
Regarding the "code always produce warning"
The -Wvla-larger than will NOT complain if MAX_POINTS was 25000
(initial value), Will complain when header.h is changes to: #define
MAX_POINTS 250000
Keeping the -Wvla-larger-than will allow give us an error if an
"approved" usage of VLA will cross the accepted boundary. The warning
on vla-larger-than will only get trigger if the array get "too much"
bigger.
As a side note - practical experience - The -Wvla-larger require
compiling in O2 (O3 if using old GCC) - so it's much more "expensive"
build. We use it for CI build - developer are using the faster "-Og"
mode for regular development.
Hope it make sense.
Yair
On Fri, Feb 27, 2026 at 12:51 PM Martin Uecker <[email protected]> wrote:
>
> Am Freitag, dem 27.02.2026 um 12:42 -0500 schrieb Yair Lenga via Gcc:
> > Greetings!
> >
> > I'm happy to see discussion (including opposing/disagreeing opinions
> > !) on my proposal.
> >
> > Wanted to cover one of the side-item in separate email - specifically
> > keeping -Wvla-larger-than not impacted by the proposed [[gnu::vla]]
> >
> > In my projects, we use -Wvla and -Wvla-larger-than=512000 (~1MB, given
> > each of our threads is configured for 5MB thread. This is the same
> > setting we for frame-larger-than.
> >
> > Common scenario we have is:
> >
> > // In header.h
> > #define MAX_POINTS 25000
> >
> > // In code
> > int foo(void)
> > {
> > int n = MAX_POINTS ;
> > ALLOW_VLA(int v[n+1] ) ; // ALLOW_VLA uses Pragma to ignore -Wvla
> > }
> >
> > The "ALLOW_VLA" indicate that the usage was reviewed - and approved.
>
> Out of curiosity, what would are the criteria for this being allowed
> or not? Or in other words, what is the actual use case you are
> trying to address?
>
> BTW: this code would always cause a warning with -Wvla-larger-than=, so
> I assume it actually looks a bit different.
>
> Martin
>
>
> > Common issue that we have is that the header.h is modified -> #define
> > MAX_POINTS 250000. The compiler does not warn on the VLA use - because
> > we tagged it safe. Keeping -Wvla-larger-than=500KB allow us to be
> > notified when of places where VLA usage is valid, but exceeds the
> > threshold.
> >
> > Hope this use case is clean - and explain why I believe
> > -Wvla-larger-than should NOT be impacted with the proposed [[gnu::lva]
> >
> > Regards,
> > Yair
