[llvm-dev] llvm (the middle-end) is getting slower, December edition

[Daniel Berlin via llvm-dev]

>
>
>>
> LVI is one of those analyses with quadratic runtime, but has a cutoff to
> its search depth so that it is technically not quadratic. So increased
> inlining could easily exacerbate it more than non-"quadratic" passes.
> (increased inlining would also cause a general slowdown too).
>
>
LVI is only quadratic because of the way we've built it (it's actually
worse than quadratic,but let's just normalize to "quadratic").
Non-lazy versions are not quadratic, and you could likely build an
incremental lazy version that is also not quadratic in practice.
At some point, none of this will change unless people hold the line
somewhere. Compilers usually get slower 0.1% at a time, not in huge leaps
and bounds.
Without people saying "If we really want to get this case, in a thing not
really designed to get that case sanely, we need to stop and think about
it", it doesn't get thought about.

Obviously, you can go too far into that extreme, but i think we are still
too far on one side of that one (at least, in most places in LLVM)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161217/3549f873/attachment.html>