On May 13, 2009, at 2:38 PM, Carl Witty wrote:
> On Wed, May 13, 2009 at 1:49 PM, Dag Sverre Seljebotn
> <[email protected]> wrote:
>> Yes, but these are only used in temporaries which are presumably
>> read just
>> after. Wouldn't the compiler be able to optimize it? E.g.
>>
>> short a = ..., b = ..., c;
>> long long long long int temp;
>>
>> temp = a + b;
>> c = temp;
>>
>> It seems plausible that the compiler could decide that temp could
>> be made
>> a short (?). But it definitely needs investigation to see exactly
>> which
>> sort of situations the knowledge about the size is used in.
No, I don't trust the compiler that much. Also, I'm thinking of type
inference of the kind
def foo(int a, short b):
cdef int a
cdef short b
i = a+b # in the future, this could be a declaration of a cdef
i, which should (IMHO) be an int
return i
> This doesn't work in more complicated situations.
>
> unsigned int a=..., b=..., c=..., d = ..., result;
> unsigned long long temp1, temp2, temp3, temp4, temp5, temp6;
>
> temp1 = a+1;
> temp2 = b+1;
> temp3 = c+1;
> temp4 = d+1;
> temp5 = temp1*temp2;
> temp6 = temp3*temp4;
> result = temp5/temp6;
>
> And assume that int is 32 bits, long long is 64 bits.
>
> Here C semantics guarantee that temp5 and temp6 are computed to full
> 64-bit resolution, and that the division is done in 64 bits (and then
> truncated to 32 bits). But if the temporaries were unsigned int,
> instead, then temp5 and temp6 would be truncated to 32 bits before
> doing the division. So the C compiler can't possibly optimize temp5
> and temp6 to unsigned int. (I believe it could optimize temp1 ..
> temp4 to unsigned int, as long as it still used 32x32->64 multiplies
> to compute temp5 and temp6.)
>
> I certainly have written C code that depends on the compiler exactly
> following the C semantics for overflow on unsigned types. I don't
> remember if I've written such code in Cython (probably not).
I think this is a valid point too. In particular, with python
division semantics, it must store the denominator in a temp in order
to detect whether or not to raise a ZeroDivisionError.
I think we should be using sizeof(T) as often as possible to do the
right thing, but we should still keep track of the rank. I do think
the current type system could be simplified without getting rid of
the rank though.
- Robert
_______________________________________________
Cython-dev mailing list
[email protected]
http://codespeak.net/mailman/listinfo/cython-dev
