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

[David Laight <David.Laight@ACULAB.COM>]

From: Vincent Mailhol
> Sent: 05 December 2024 15:31
> 
> -CC: Martin Uecker <Martin.Uecker@med.uni-goettingen.de>
> +CC: Martin Uecker <muecker@gwdg.de>
> (seems that Martin changed his address)
> 
> On Thu. 5 Dec. 2024 at 03:39, David Laight <David.Laight@aculab.com> wrote:
> > > Sent: 02 December 2024 17:33
> > >
> > > From: Vincent Mailhol <mailhol.vincent@wanadoo.fr>
> > >
> > > __is_constexpr(), while being one of the most glorious one liner hack
> > > ever witnessed by mankind, is overly complex. Following the adoption
> > > of C11 in the kernel, this macro can be simplified through the use of
> > > a _Generic() selection.
> >
> > You should give credit to some of the earlier patches that do the same.
> > I'm sure there were some related ones from Linus - not applied yet.
> 
> ACK. Would adding a suggested--by Linus tag solve your concern?

I actually suspect the first patches to change __is_constexpr() to
use _Generic were from myself.
I've found a patch I send in November 2023.

> 
> > > First, split the macro in two:
> > >
> > >   - __is_const_zero(x): an helper macro; tells whether x is the
> > >     integer constant expression 0 or something else.
> > >
> > >   - is_const(x): replacement of __is_constexpr(); tells whether x is a
> > >     integer constant expression.
> > >
> > > The split serves two purposes: first make it easier to understand;
> > > second, __is_const_zero() will be reused as a building block for other
> > > is_const_*() macros that will be introduced later on.
> > >
> > > The core principle of __is_constexpr() to abuse the return type of the
> > > ternary operator remains, but all the surrounding sizeof() hack
> > > disappear.
> > >
> > > On a side note, while not relevant to the kernel, __is_constexpr()
> > > relied on the GNU extension that sizeof(void) is 1. const_expr() does
> > > not use any GNU extensions, making it ISO C compliant.
> > >
> > > __is_constexpr() is temporarily kept and will be removed once all its
> > > users get migrated to is_const() (or one of its friend).
> > >
> > > Signed-off-by: Vincent Mailhol <mailhol.vincent@wanadoo.fr>
> > > ---
> > >  include/linux/compiler.h | 41 +++++++++++++++++++++++++++++++++++++++++
> > >  1 file changed, 41 insertions(+)
> > >
> > > diff --git a/include/linux/compiler.h b/include/linux/compiler.h
> > > index a2a56a50dd85227a4fdc62236a2710ca37c5ba52..30ce06df4153cfdc0fad9bc7bffab9097f8b0450 100644
> > > --- a/include/linux/compiler.h
> > > +++ b/include/linux/compiler.h
> > > @@ -316,6 +316,47 @@ static inline void *offset_to_ptr(const int *off)
> > >  #define statically_true(x) (__builtin_constant_p(x) && (x))
> > >  #define statically_false(x) (__builtin_constant_p(x) && (x) == 0)
> > >
> > > +/*
> > > + * Whether x is the integer constant expression 0 or something else.
> > > + *
> > > + * Details:
> > > + *   - The C11 standard defines in §6.3.2.3.3
> > > + *       (void *)<integer constant expression with the value 0>
> > > + *     as a null pointer constant (c.f. the NULL macro).
> > > + *   - If x evaluates to the integer constant expression 0,
> > > + *       (void *)(x)
> > > + *     is a null pointer constant. Else, it is a void * expression.
> > > + *   - In a ternary expression:
> > > + *       condition ? operand1 : operand2
> > > + *     if one of the two operands is of type void * and the other one
> > > + *     some other pointer type, the C11 standard defines in §6.5.15.6
> > > + *     the resulting type as below:
> > > + *       if one operand is a null pointer constant, the result has the
> > > + *       type of the other operand; otherwise [...] the result type is
> > > + *       a pointer to an appropriately qualified version of void.
> > > + *   - As such, in
> > > + *       0 ? (void *)(x) : (char *)0
> > > + *     if x is the integer constant expression 0, operand1 is a null
> > > + *     pointer constant and the resulting type is that of operand2:
> > > + *     char *. If x is anything else, the type is void *.
> > > + *   - The (long) cast silences a compiler warning for when x is not 0.
> > > + *   - Finally, the _Generic() dispatches the resulting type into a
> > > + *     Boolean.
> >
> > The comment is absolutely excessive.
> > I'm sure I managed about 2 lines in one of the patches I did.
> 
> I think that Linus made it  clear in:
> 
>   https://lore.kernel.org/all/CAHk-=wgfpLdt7SFFGcByTfHdkvv7AEa3MDu_s_W1kfOxQs49pw@mail.gmail.com/
> 
> that this deserves a detailed comment.

And he wrote one in https://lore.kernel.org/all/CAHk-=wiq=GUNWJwWh1CRAYchW73UmOaSkaCovLatfDKeveZctA@mail.gmail.com/

   /*
    * iff 'x' is a non-zero constant integer expression,
    * then '!(x)' will be a zero constant integer expression,
    * and casting that to 'void *' will result in a NULL pointer.
    * Otherwise casting it to 'void *' will be just a regular 'void *'.
    *
    * The type of '0 ? NULL : (char *)' is 'char *'
    * The type of '0 ? (void *) : (char *) is 'void *'
    */
    #define const_true(x) \
        _Generic(0 ? (void *)((long)!(x)) : (char *)0, char *: 1, void *: 0)



> 
> The details block in the current __is_constexpr() is 37 lines long,
> the details block in __is_const_zero() takes 22 lines. So I would
> argue that I made things better.

The old block was too long :-)

> 
> Unless more people share your concern, I am planning to keep this comment as-is.
> 
> > > + *
> > > + * Glory to Martin Uecker <Martin.Uecker@med.uni-goettingen.de>
> >
> > IIRC Martin has agreed in the past that the accreditation can
> > be removed - especially since it refers to the 'sizeof (void)' trick.
> 
> I tried to look for such message:
> 
>   https://lore.kernel.org/all/?q=f%3A%22martin+uecker%22+__is_constexpr
> 
> but couldn't find it. Do you have the link?
> 
> @Martin, do you agree that I remove the accreditation?
> 
> > > + */
> > > +#define __is_const_zero(x) \
> > > +     _Generic(0 ? (void *)(long)(x) : (char *)0, char *: 1, void *: 0)
> > > +
> > > +/*
> > > + * Returns a constant expression while determining if its argument is a
> > > + * constant expression, most importantly without evaluating the argument.
> >
> > You need to differentiate between a 'constant integer expression'
> > and a 'compile time constant'.
> 
> OK. This one was just copied from the previous __is_constexpr(). I will apply
> "s/constant expression/constant integer expression/g" in v2.
> 
> > > + *
> > > + * If getting a constant expression is not relevant to you, use the more
> > > + * powerful __builtin_constant_p() instead.
> >
> > __builtin_constant_p() is not 'more powerful' it is testing for
> > something different.
> 
> I meant to say that __builtin_constant_p() is more powerful at
> constant folding. But I agree that the comment is not clear.
> 
> What about this?
> 
>   If getting a constant integer expression is not relevant to you, use
>   __builtin_constant_p() which not only returns true if the argument
>   is an integer constant expression, but also if it is a compile time
>   constant.

Complete f***ed tense.

It's not about 'constant folding' and 'powerful' isn't the correct word.
They are checking for two different things.

A 'constant integer expression' is defined by the C language, and is
basically something that is constant when first parsed by the compiler
(my definition) so it can pretty much only contain constants, sizeof()
and offsetof().

__builtin_constant_p() is true if the compiler decides that an expression is
constant. This can track values through inlined function calls and can
change from 'unknown' to 'true' late in the compilation.

	David

> 
> 
> Yours sincerely,
> Vincent Mailhol

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