On 2014-11-12 at 11:22:25 +0100, Simon Peyton Jones wrote:
Great stuff. But why do primetest, rsa, and mandel allocate so much more? You mention different accounting, but I've found that it is often worth really understanding what is happening because you discover "Oh, yes, that's egregiously bad, and easily fixed".
Oh, but I still plan to investigate those allocation numbers a bit more to give a more exhaustive explaination (or even fix them)! But I've already got a few ideas where to look: For instance,for the old integer-gmp I did that trick to to avoid single-limb allocations (by using a C stack-allocated temp buffer instead) so much as to avoid them altogether, and I may have lost some of those opportunities due to the different design. Cheers, hvr