To be more specific...

GHC used to inline the code you're exercising, which allows it to specialize to the comparison function, which allows it to keep all the numbers unboxed. But the current version of GHC doesn't do this, and if I force it to inline, the compiler will just die with a 'simplifier ticks exhausted' message.

I've been thinking about ways to avoid this situation, and have some ideas (one of which would work for your use case). But I guess in the mean time I should turn off use of these functions.

On Thu, Aug 21, 2014 at 7:18 PM, Bryan O'Sullivan <bos@serpentine.com> wrote:

On Thu, Aug 21, 2014 at 3:51 PM, Alexander Pakhomov <ker0sin@ya.ru> wrote:

Which GHC version do you use?
With 7.6.3 I have all the same BUT Optimal'/10 260 ns. It's better than 1200 ns
but still bad. Also there are INLINEs everywhere at vector-algorithms.
Seems to be compiler problem.

There's a lot of sensitivity in the compiler to exactly how and which rules are specified. I already brought up this regression with Dan a few weeks ago, and we didn't come up with a good way to deal with it. I don't think it's quite fair to call it a compiler problem, so much as just a change that caused a previously-good-but-fragile behaviour to now not be so good.