[PATCH 02/10] compiler.h: add is_const() as a replacement of __is_constexpr()

Martin Uecker muecker at gwdg.de
Sat Dec 7 19:19:37 UTC 2024


Am Samstag, dem 07.12.2024 um 10:26 -0800 schrieb Linus Torvalds:
> On Sat, 7 Dec 2024 at 05:07, Martin Uecker <muecker at gwdg.de> wrote:
> > 
> > VLA use *less* stack than a fixed size arrays with fixed bound.
> 
> Not really. You end up with tons of problems, not the least of which
> is how to actually analyze the stack size. It also gets *very* nasty
> to have code that declares the VLA size using an argument that is then
> checked afterwards - and if you have a strong preference for
> "declarations before code", you end up with *horrific* issues.
> 
> And even if you are super-careful, and you solved the analysis
> problem, in practice VLAs will cause huge stack issues simply due to
> code generation issues.  The compiler will end up doing extra
> alignment and extra frame handling and saving, to the point where any
> advantages the VLA would bring is completely dwarfed by all the
> disadvantages.

But that all seem solvable issues on the compiler side.  If you 
said the maximum stack size for arrays we tolerate is X,
then a compiler could tell you if 

a) this is not guaranteed in a specific situation (-Wvla-larher-than)
and 
b) transform the array automatically to fixed size array
of size X *or* something smaller when it can show this.

Because now you do the exact same thing manually while losing
precise bounds checking.

Martin

> 
> We went through this. We are so *much* better off without VLAs that
> it's not even funny.
> 
> Now when the compiler says "your stack size is big", you just look
> "Oh, that struct should be allocated with kmalloc, not on the stack".
> Boom. Done.
> 
>             Linus



More information about the Intel-gfx mailing list