Ridiculous compile time for alg.find_end about cmcstl2 HOT 5 CLOSED

Is this related to your changes to begin/end?

from cmcstl2.

Yes - it wouldn't compile at all before the memoization changes (#11). It would fill the 7G of virtual memory in my VirtualBox and the Linux kernel OOM killer would kill the compiler. It's probably a similar issue with multiple redundant constraints on the many find_end overloads that could be worked around with some clever memoization.

from cmcstl2.

1c46215 is a 40% speedup - still slower than I'd like.

from cmcstl2.

cf29332 is memoization taken to the extreme - pushed as far as possible without obscuring subsumption relationships between concepts. The concepts are effectively decomposed into conjunctions of memoizable traits. alg.find_end compiles in 15 seconds.

It is, however, not strictly conforming. Since the expressions are all hidden from the compiler, it won't know that e.g. an overload that requires a == b is strictly less specialized than an overload that requires EqualityComparable<decltype(a), decltype(b)>(). I hate to commit something so "loosely" conformant, but we're rapidly outscaling the compiler's capabilites.

from cmcstl2.

Merged the memoization branch to master in 8650909. For the time being, we need a working reference implementation more than we need strict conformance.

from cmcstl2.