Why does (unsigned short)-1.3 == 0 and not 65535? (But only if optimizing with GCC 4.x)

[Bob Plantz]

I think this is a good illustration of why one should not use type 
casting to perform algorithmic operations. You should operate on the 
date and test the resulting value to ensure it will fit within the new type.

I'm assuming here that in the actual code it is not known whether d is 
positive or negative. But type casting a possibly negative value into an 
unsigned value is simply asking for your code to break.

I'm a big fan of writing code that explicitly shows each step in an 
algorithm one statement at a time. If the operations are in a loop, and 
testing shows that the loop is slowing things down too much, then 
obscure and tricky code may be in order, but only with good commenting. 
The maintenance programmer, perhaps you at a later date, will appreciate 
being able to follow simple code.

--Bob

On 3/24/2010 1:58 PM, Ian Lance Taylor wrote:
> Jack Lloyd<lloyd@randombit.net>  writes:
>
>    
>> int main()
>>     {
>>     double d = -1.3;
>>     unsigned short x = (unsigned short)d;
>>     printf("%d\n", x);
>>     }
>>      
> The C99 standard says that when converting floating point to integer,
> the value is first truncated toward zero.  If the value of the
> integral part cannot be represented by the integer type, the behaviour
> is undefined.  This means that converting a negative floating point
> value<= -1.0 to an unsigned integer type is undefined, and the code
> may act unpredictably.
>
>
>    
>> Secondly, can someone tell me if there is a warning I can use to tell
>> me when this is happening? The code compiles cleanly with (or without)
>> optimizations using "-Wall -Wextra -Wconversion". In this specific
>> case, I can get the behavior I want using
>>    (unsigned short)(int)d;
>>      
> Adding the cast works because -1 can be represented in the type int,
> and conversions from signed integer types to unsigned integer types
> are defined to do the obvious thing.
>
> Unfortunately, while gcc warns about conversions which have no
> explicit cast, I don't think there is any option to warn about an
> explicit cast which happens to be invalid.
>
> Ian
>