[MLton] Experimental 64-bit binary package (& SVN sources)

[Matthew Fluet]

Daniel C. Wang wrote:
> Matthew Fluet wrote:
> {stuff deleted}
>>    Word64.  Unfortunately, we don't have a *bitcast* from Real32 to
>>    Word32; naively emitting
>>      Real32 r = ...;
>>      Word32 w = (Word32)r;
>>    results in coercion that attempts to preserve the numeric meaning
>>    of r, rather than the bits of r.  Simply more evidence that we need
>>    real/word bitcast primitives.
> 
> Need them in the MLton IR or the C-code?

Both.  We've suggested having such primitives for better Basis Library 
implementation (e.g., implementing PackReal<N> in terms of PackWord<N>, 
and for floating-point predicates (isnan, etc.)); for export through 
MLton.* structures (in particular, for providing efficient hash 
functions on Real<N>.real); and now for supporting packing of 
Real<N>.real values in datatype representations.

Of course, once you have a primitive, you need codegen support for 
implementing it; for the C-codegen, that means C-code.

 > This I think does the "right"
> thing in ANSIC.
> typedef  union {
>             Word32 asWord32;
>             Reald32 asReal32;
>             /* ... */
>               } univ_t;
> inline Real32bitstoWord32bits(Word32 w) {
> univ_t v;
> v.asWord32 = w;
> return  v.asReal32;
> }

Technically, this is undefined in C.  (It is only defined to read the 
most recently written union member.)  However, it is a widely supported 
idiom.  We use this idiom to handle mis-aligned reads and writes of 
Real64.real on HPPA and Sparc.  (One only gets the mis-aligned accesses 
if compiled with -align 4.)