[llvm-dev] InstCombine, graph rewriting, Equality saturation
Daniel Berlin via llvm-dev
llvm-dev at lists.llvm.org
Mon Sep 11 09:06:18 PDT 2017
Or the analysis need to be thought about and extended to not require fix
pointing to get the same end result out of the other end of llvm.
In situations like this, the real question is "does the same kind of
optimized code come out the other end", not "do we perfectly match every
little optimization on a per-pass basis".
Or at least, that's the answer i got about GVN.
(because trying to replace a do-everything pass that fixpoints with split
up passes rarely can match it perfectly Too many second order effects. It
shouldn't be your goal)
On Mon, Sep 11, 2017 at 8:50 AM, Sean Silva <chisophugis at gmail.com> wrote:
> On Mon, Sep 11, 2017 at 8:14 AM, Daniel Neilson via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>> Just thinking out loud…. I’m really not familiar with the vast majority
>> of what instcombine does, so please excuse me if my naiveté is too obvious.
>> I can’t help but notice all of the uses of ‘and’ in Daniel B’s description
>> of what instcombine is right now:
>> > On Sep 8, 2017, at 11:27 PM, Daniel Berlin via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>> > FWIW, Before getting to "how to do it", I think InstCombine needs an
>> actual goal.
>> > Right now it's "do a bunch of instruction combination and
>> canonicalization and random kinds of semi-local optimization with some
>> weird analysis and other stuff in the middle.
>> > Iterate this as hard as we can"
>> > Nobody can really draw any real dividing line for what should be there
>> or not except based on how easy or expensive it is to shove it in.
>> > That's a recipe for pass creep.
>> This makes me wonder… is it sensible to be talking about splitting up
>> instcombine into multiple separate passes? Would such a thing even be
>> possible? For example, split by functionality into separate passes that
>> each do one of:
>> * instruction combinations
>> * canonicalization
>> * semi-local optimizations
>> * etc…
>> Or something like that.
>> As separate passes, each would probably have a natural way to be
>> implemented effectively and those implementations might vary.
> One obstacle to that is that currently instcombine has an internal
> fixed-point iteration that it does.
> So when splitting it into separate passes we would need to either add
> fixed-point iteration to the pass manager running the separate instcombine
> passes (extending the pass management in this way is doable and has been
> explored in the past, e.g. https://www.youtube.com/watch?v=c7iP43an5_Q )
> or demonstrate/measure that we don't regress without the fixed-point
> iteration across separate instcombine passes.
> -- Sean Silva
>> Daniel Neilson, Ph.D.
>> Azul Systems
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev