Re: Branchless implementation for literal case – is it worth it?
dan.doel at gmail.com
Mon Apr 20 18:39:59 UTC 2015
On Mon, Apr 20, 2015 at 10:29 AM, Sven Panne <svenpanne at gmail.com> wrote:
> * On more powerful cores with heavy out-of-order execution, it's
> hard to beat a well-predicted branch.
With regard to this, I was wondering if we have a way of arranging for
good branch prediction in GHC.
For instance (and I think I've discussed this with Carter before), we've
had issues filed against vector (I think) saying that we do branches for
certain things that could be branchless. Bounds checks are the first thing
that comes to mind, for instance (I.E. we don't use orI# for the lower and
upper bound checks). However, so long as branch prediction decides that the
bounds checks should succeed, I'd expect them to not be much of a cost
anyway---and ideally it will always decide that, because we're inclined not
to care about the performance when the checks fail and the program is
However, from what I've read, there's basically no easy hinting that can
ensure this. The heuristics require moving generated code around if you
want to try and influence the branch predictor one way. And we also have no
way of indicating to GHC that branches of a case should be preferred
anyway. Unless writing things in a certain order accomplishes this?
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the ghc-devs