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

[Al Viro <viro@ZenIV.linux.org.uk>]

On Fri, Mar 16, 2018 at 05:55:02PM +0000, Al Viro wrote:
> On Fri, Mar 16, 2018 at 10:29:16AM -0700, Linus Torvalds wrote:
> >    t.c: In function ‘test’:
> >    t.c:6:6: error: argument to variable-length array is too large
> > [-Werror=vla-larger-than=]
> >      int array[(1,100)];
> > 
> > Gcc people are crazy.
> 
> That's not them, that's C standard regarding ICE.  1,100 is *not* a
> constant expression as far as the standard is concerned, and that
> type is actually a VLA with the size that can be optimized into
> a compiler-calculated value.
> 
> Would you argue that in

s/argue/agree/, sorry

> void foo(char c)
> {
> 	int a[(c<<1) + 10 - c + 2 - c];
> 
> a is not a VLA?

FWIW, 6.6 starts with
	 constant-expression:
		conditional-expression
for syntax, with 6.6p3 being "Constant expression shall not contain
assignment, increment, decrement, function call or comma operators,
except when they are contained in a subexpression that is not evaluated",
with "The operand of sizeof operator is usually not evaluated (6.5.3.4)"
as a footnote.

6.6p10 allows implementation to accept other forms of constant expressions,
but arguing that such-and-such construct surely must be recognized as one,
when there are perfectly portable ways to achieve the same...

Realistically, code like that can come only from macros, and one can wrap
the damn thing into 0 * sizeof(..., 0) + just fine there.  Which will
satisfy the conditions for sizeof argument not being evaluated...