Re: [PATCH] c++: Implement C++26 P2809R3 - Trivial infinite loops are not Undefined Behavior

[Jason Merrill <jason@redhat.com>]

On 4/3/24 15:16, Jakub Jelinek wrote:
> On Wed, Apr 03, 2024 at 12:58:12PM -0400, Jason Merrill wrote:
>> On 4/3/24 12:42, Jakub Jelinek wrote:
>>> On Wed, Apr 03, 2024 at 12:07:48PM -0400, Jason Merrill wrote:
>>>> Using std::is_constant_evaluated directly in a loop condition is, as the
>>>> paper says, unlikely and "horrendous code", so I'm not concerned about
>>>> surprising effects, though I guess we should check for it with
>>>> maybe_warn_for_constant_evaluated.
>>>
>>> Ok, though guess the question is what to say about it.
>>> Because unlike the existing cases in maybe_warn_for_constant_evaluated
>>> where it always evaluates to true or always to false depending on where,
>>> in the trivial empty iteration statements it evaluates to always true or
>>> always false depending or sometimes true, sometimes false, depending on
>>> if the condition is a constant expression that evaluates to true (then it is
>>> always true), or if in immediate function (also always true), or if not
>>> in constexpr function (then always false), or in constexpr function (then
>>> it might be true or false).
>>> Not sure how exactly to word that.
>>> Maybe just say that it is horrendous code to use std::is_constant_evaluated
>>> () in trivial empty iteration statement conditions ;)
>>
>> Maybe if the condition constant-evaluates to true, warn something like
>> "%<std::is_constant_evaluated%> always constant-evaluates to true in the
>> condition of a trivially empty iteration statement"?
> 
> Given the mails on core about this, isn't it actually the case that not all
> "trivial infinite loops" actually loop at all or if they loop, don't loop
> infinitely?
> I.e. what I wrote in the patch would be wrong and that generally we need to
> use ANNOTATE_EXPR in some cases?

Apparently so.

> As in, for trivially empty iteration statements try to evaluate the
> maybe_const_value evaluate condition with mce_true, if it yields true,
> try to maybe_const_value evaluate condition with mce_false, if it also
> yields true, replace it with true, but if the former returns true and
> the latter isn't a constant expression or returns false mark the loop with
> ANNOTATE_EXPR as infinite loop regardless of -ffinite-loops and actually
> keep invoking the condition in there?

Sure.

Jason