On Saturday, January 07, 2012 04:49:27 Timon Gehr wrote:
> On 01/07/2012 12:37 AM, Jonathan M Davis wrote:
> > On Saturday, January 07, 2012 00:03:39 Alex Rønne Petersen wrote:
> >> Most likely those functions are just implemented using inline
> >> assembly,
> >> therefore not usable in CTFE.
> > 
> > Yeah, several functions in std.math use inline assembly. So, for them to be able to be used at compile time, either the compiler must be expanded to be able to run asm statements at compile time (which may or may not be planned and may or may not be reasonable), or those functions need another branch (using __cfte in an if condition) which doesn't use assembly. Or I suppose that if the extra check for __ctfe isn't considered particularly acceptable (after all, they're already using assembly)  [snip.]
> 
> If the if condition is a constant, there is no runtime overhead.

Ah, good point - though depending on what the compiler does, that may only be the case with -O. But anyone who really cares about that level of performance would be compiling with -O anyway.

- Jonathan M Davis

On 07-01-2012 01:31, H. S. Teoh wrote:
> On Sat, Jan 07, 2012 at 12:49:46AM +0100, Alex Rønne Petersen wrote:
>> On 07-01-2012 00:37, Jonathan M Davis wrote:
>>> On Saturday, January 07, 2012 00:03:39 Alex Rønne Petersen wrote:
>>>> Most likely those functions are just implemented using inline
>>>> assembly, therefore not usable in CTFE.
>>>
>>> Yeah, several functions in std.math use inline assembly. So, for them
>>> to be able to be used at compile time, either the compiler must be
>>> expanded to be able to run asm statements at compile time (which may
>>> or may not be planned and may or may not be reasonable), or those
>>> functions need another branch (using __cfte in an if condition) which
>>> doesn't use assembly. Or I suppose that if the extra check for __ctfe
>>> isn't considered particularly acceptable (after all, they're already
>>> using assembly), then separate functions meant specifically for CTFE
>>> would be necessary.
> [...]
>
>> From my limited experience, I'd say that having two versions of the
> function is probably the least painful way to go.
>
>
> [...]
>> Allowing asm in CTFE would probably be way more work than it's worth.
>> You'd basically need full-blown analysis of x86 assembly plus an
>> interpreter. Even then, x86 is not typed, so it's going to be a major
>> pain...
> [...]
>
> I admit I've no idea how the D compiler implements compile-time
> evaluation, but is it possible for the compiler to actually emit code
> for compile-time functions containing asm blocks and, say, execute it in
> a sandbox, and read the values out from the machine registers? Or does
> this create more problems than it solves?
>
>
> T
>

Executing asm at compile time is also a security risk.

-- 
- Alex