-Wsign-promo (Re: [BUILD FAILS DEBUG MODE] with test_strings_valuex.cxx)

[Stephan Bergmann]

On 01/19/2013 03:12 PM, Michael Stahl wrote:
> On 19/01/13 11:28, Lubos Lunak wrote:
>>   -Wsign-promo is a rather pointless warning these days (the section in the gcc
>> manpage is a funny read and not only because it talks about Cfront). I've

Some of the funniness is likely due to 
<http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56067> "Removal of -Wsynth 
from doc didn't remove example."

>> added more overloads to silence it, but I rather wonder why we have this
>> explicitly enabled at all.
>>
>>   My hypothesis is like this:
>> - the idea behind the warning is just nonsense (who cares to what integer type
>> the value is promoted)
>
> overflow for unsigned integral types is defined by the C and C++
> standards, while overflow for signed types is explicitly undefined.
> some implementations therefore assume that it does not occur and may
> remove tests on signed integers from the code that could only evaluate
> to true in case of an overflow in order to improve the all-important
> SPEC benchmark scores.
>
> -Wsign-promo is specifically about overload resolution, i'm not sure if
> that overflow problem would be relevant in this case but this is C++ so
> i'm never sure about anything :)

No, undefined behavior due to signed overflow is not relevant here.  For 
the integral promotion involved in overload resolution, no overflow can 
happen.  (And the overflow that can happen with integral conversion, say 
a truncation from long to int, is implementation-defined rather than 
undefined, btw.)

Stephan