Re: [+externe Mail+] RE: [PATCH] compiler.h: Explain how __is_constexpr() works

From: Uecker, Martin
Date: Fri Mar 01 2024 - 08:56:49 EST


Am Freitag, dem 01.03.2024 um 13:43 +0000 schrieb David Laight:
> From: Uecker, Martin
> > Sent: 01 March 2024 13:22
> >
> > My suggestion would also to limit explanation. Nobody should
> > write such code and if you need to, you can find explanations
> > all over the internet.
> >
> > Finally, I still think the motivation for this macro (removing
> > VLAs) is misguided if security is the goal because VLAs provide
> > precise bounds and larger worst-case fixed-size arrays do not.
> >
> > It would be better to use the compiler options that detect
> > possibly use of VLAs of unbounded size and if there a problems
> > with this, improve this on the compiler side.
>
> In kernel code (with limited stack) there has to be enough room
> for the largest possible 'VLA' so you might as well allocate one.
>
> Allowing VLA also makes it pretty much impossible to do any
> kind of static stack use analysis.

If you limit VLAs to a certain maximum size, then you could use
this for analysis and it would not be worse than using worst case
fixed-size array on the stack. But you can also use the *precise*
run-time bound of the VLA if your static analysis is smart enough.
You can also use the precise run-time bound for run-time bounds
checking. It is strictly more expressive to use VLAs (or dependent
types in general) and therefor *good* for static analysis.

> The fine IBT tags can be used identify valid indirect calls
> which pretty much only leaves recursion stopping full static
> stack analysis - and that could be banned except for a few
> limited cases where 1 level could be permittd.
>
> is_constexpr() has other uses - there are places where
> __builtin_constant_p() isn't strong enough.
> Particularly if you need to use builtin_choose_expr()
> or _Generic() to get select a type.
>
> For instance, if you can a constant value between 0 and MAXINT
> it is safe to cast to/from unsigned in order change any
> implicit integer promotion cast that may be grief some.

glad to hear it is useful.

Martin

>
> David
>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)