Compile time float binary representation

[Andrei Alexandrescu]

Jeremie Pelletier wrote:
> Jeremie Pelletier Wrote:
> 
>> Andrei Alexandrescu Wrote:
>>
>>> Jeremie Pelletier wrote:
>>>> Is there a way to convert a float (or double/real) to an integral number without changing its binary representation at compile time?
>>>>
>>>> I need to extract the sign, exponent and mantissa yet I cant use bit shifting.
>>>> "Error: 'R' is not of integral type, it is a real" is the error I get.
>>>>
>>>> The usual *cast(uint*)&value wont work either at compile time.
>>>>
>>>> Any suggestions?
>>> I got this to work, would it be applicable to your need?
>>>
>>> void main(string[] args)
>>> {
>>>      union A
>>>      {
>>>          double x;
>>>          int y;
>>>      }
>>>      enum A a = A(0.3);
>>>      writeln(a.x);
>>>      writeln(a.y);
>>> }
>>>
>>>
>>> Andrei
>> Good suggestion, however it doesnt work the way I need it:
>>
>> ---
>> template FloatParts(real R) {
>>     union A {
>>         real theReal;
>>         struct {
>>             ushort hi;
>>             ulong lo;
>>         }
>>     }
>>     
>>     A a = cast(A)R; // Error: cannot cast real to immutable(A)
>>     enum A a = A(R); // members cannot be used at runtime
>> }
>> ---
>>
>> I need to gather .sign, .exp and .mantissa among others (possibly a .radix too) and use them at compile time (generic metaprogramming, or else i would just fallback to runtime).
>>
>> Unless there are arithmetic methods to get these, I dont see how it can be done without changes in the compiler first.
>>
>> To Walter: is this something possible for DMD2.32?
> 
> After some toying around, I managed to get some results:
> 
> ---
> import std.metastrings;
> immutable union A {
> 	double x;
> 	int y;
> }
> pragma(msg, ToString!(A(5.2).y));
> ---
> 
> but it crashed dmd (2.031) instead of printing 3435973837 to stdout.
> 

Yah, I crashed it a couple of times with similar code now too. Could you 
please do the bugzilla honors.

(Also you may want to check std.numeric.CustomFloat. It doesn't help 
with this particular problem, but it allows defining floating point 
types with specific mantissa and exponent sizes.)


Andrei