Aaron Sherman <[EMAIL PROTECTED]> wrote: > What might be interesting is to compare Parrot to Parrot doing this with > and without a hyper-operator. That's all I was trying to say.
I'd posted that as well. Here again with an O3 build of parrot: $ time parrot ih.imc 6060 real 0m0.214s user 0m0.150s sys 0m0.040s $ time parrot -j i.imc 6060 real 0m1.793s user 0m1.750s sys 0m0.040s Both are using the same natural int array. But the builtin does use list internals to directly access list chunk memory. This is of course only possible for natural types in our aggregates, where no vtable is needed to access the aggregates data members. But the numbers do indicate, that number crunchers just want that. > As for the DOD: you have an excellent point, but it extends far beyond > the hyper-operators. I'm starting to think that front-ends like the > Python compiler or the Perl 6 compiler are going to need controls over > the DOD for just the reasons you cite. After all, they know when they > are about to start doing some large looping operation that's all highly > constrained with respect to allocation. It would make sense to gather > the resources they need, lock down DOD, do what they need to do and then > unlock the DOD... Well, it's unlikely that we can expose all the details the more that such details may change. We could have a generalized version of such an operation though: i_need_now_x_pmcs_and_wont_dispose_any_start 100000 # ... deep clone code or loop i_need_now_x_pmcs_and_wont_dispose_any_end er EOPCODETOOLONG :) But as soon as there are any temps created inside that code, this might already fail miserably because of resource outage. OTOH a compiler now already can emit: sweep 1 sweepoff ... deep clone or some such sweepon leo
