Signed word lengths and indexes

[Andrei Alexandrescu]

Jérôme M. Berger wrote:
> Jérôme M. Berger wrote:
>> Walter Bright wrote:
>>> Jérôme M. Berger wrote:
>>>>     Actually, that problem already occurs in C. I've had problems when
>>>> porting code from x86 to x86_64 because some unsigned operations
>>>> don't behave the same way on both...
>>> How so? I thought most 64 bit C compilers were specifically designed to
>>> avoid this problem.
>> 	I can't isolate it to a minimal test case, but at my job, we make
>> an image processing library. Since negative image dimensions don't
>> make sense, we decided to define width and height as "unsigned int".
>> Now, we have code that works fine on 32-bit platforms (x86 and arm)
>> but segfaults on x86_64. Simply adding an (int) cast in front of the
>> image dimensions in a couple of places fixes the issue (tested with
>> various versions of gcc on linux and windows).
>>
> 	Gotcha! See the attached test case. I will post the explanation for
> the issue as a reply to give everyone a chance to try and spot the
> error...
> 
> 		Jerome
> 

Whoa! That's indeed unfortunate. Allow me some more whoring for TDPL:

==============
\indexes{surprising behavior!of unary \lstinline{-}}%
One surprising  behavior of  unary minus is  that, when applied  to an
unsigned value,  it still yields  an unsigned value (according  to the
rules in~\S~\vref{sec:typing-of-ops}).  For example,\sbs @-55u@ is\sbs
@4_294_967_241@, which is\sbs \ccbox{uint.max - 55 + 1}.

\indexes{unsigned type, natural number, two's complement, overflow}%
The fact that unsigned types are  not really natural numbers is a fact
of  life.   In\sbs\dee  and  many other  languages,  two's  complement
arithmetic with  its simple overflow  rules is an  inescapable reality
that cannot be abstracted away.  One way to think \mbox{of} @-val@ for
any integral value~@val@ is to  consider it a short form \mbox{of}$\,$
\cc{\~val + 1};  in other words, flip every bit in  @val@ and then add
@1@  to  the result.   This  manipulation  does  not raise  particular
questions about the signedness of~@val at .
==============

(This heavily adorned text also shows what sausage making looks like...)


Andrei