Re: [PATCH v5 0/2] Remove false-positive VLAs when using max()

[Rasmus Villemoes]

On 2018-03-17 19:52, Linus Torvalds wrote:
> On Sat, Mar 17, 2018 at 12:27 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>>
>> Unfortunately my 4.4 test fails quickly:
>>
>> ./include/linux/jiffies.h: In function âjiffies_delta_to_clock_tâ:
>> ./include/linux/jiffies.h:444: error: first argument to
>> â__builtin_choose_exprâ not a constant
> 
> Ok, so it really looks like that same "__builtin_constant_p() doesn't
> return a constant".
> 
> Which is really odd, but there you have it.

Not really. We do rely on builtin_constant_p not being folded too
quickly to a 0/1 answer, so that gcc still generates good code even if
the argument is only known to be constant at a late(r) optimization
stage (through inlining and all). So unlike types_compatible_p, which
can obviously be answered early during parsing, builtin_constant_p is
most of the time a yes/no/maybe/it's complicated thing. Sure, when the
argument is just a literal or perhaps even any kind of ICE, gcc can fold
it to "yes", and I think it does (though the details of when and if gcc
does that can obviously be very version-dependent, which may be some of
what we've seen). But when it's not that obvious, gcc leaves it in the
undetermined state. That's not good enough for builtin_choose_expr,
because even the type of the resulting expression depends on that first
argument, so that really must be resolved early.

So to have some kind of builtin_constant_p control a
builtin_choose_expr, it would need to be a "builtin_ice_p" or
"builtin_obviously_constant_p" that would always be folded to 0/1 as
part of evaluating ICEs.

So I don't think there's any way around creating a separate macro for
use with compile-time constants.

Rasmus