Ben:

First, thanks for that enumeration. Very helpful.

On Fri, Apr 4, 2014 at 7:13 PM, Ben Kloosterman <[email protected]> wrote:

> I see no strong reason for self modification of code .
>

Unfortunately the data on that is *extremely* compelling. I agree that
there aren't many algorithms where the complexity is justified, but in the
cases where it *is* justified it's *very* justified. As one example, I'd
cite the classic implementation of bitblt, which is implemented by
generating and then running the case-specific code.

The question, I think, is whether these cases generalize sufficiently that
some form of generalized support is in order. There is *lots* of
quantifiable benefit from compile time specialization (a la templates).
There are also a large number of use case arguments from JIT compilation
for dynamic specialization.

I can't point to comprehensively convincing arguments about statically
*guided* dynamic specialization, but it's very hard to imagine that linking
that chain together would fail to yield significant benefits.


Jonathan
_______________________________________________
bitc-dev mailing list
[email protected]
http://www.coyotos.org/mailman/listinfo/bitc-dev

Reply via email to