Disable arm-gepopt by default
mzolotukhin at apple.com
Mon Apr 20 21:22:17 PDT 2015
Hi Tim and others,
So, the entire discussion is moving around the assumption that this pass is beneficial in some cases, but we don’t have such data. Please note that I don’t want to say that the assumption is wrong, but as far as I understand, only Hao observed some gains. Thus, until we hear from him, we have real slow downs versus presumable gains.
As for the looking into the code later - I’m not an expert in this area, and I don’t even have a motivational example for this optimization, so I’d barely be a right person for it. I’d be glad to assist with collecting some additional data/running tests if that might help though.
> On Apr 20, 2015, at 8:40 PM, Tim Northover <t.p.northover at gmail.com> wrote:
> On 20 April 2015 at 20:34, Gerolf Hoflehner <ghoflehner at apple.com> wrote:
>> It shouldn’t come across as ‘extreme’.
> The "extreme" part wasn't so much the tentative disabling as "Could
> you and/or Hao take this on after the code is disabled/backed out?". I
> realise it wasn't an ultimatum, but it could very easily turn into
> "we're washing our hands of this", which I'd object to.
> I assume the pass was committed with some kind of metric that it
> improved, and now we have a counter-example. It's down to everyone to
> decide what the best way to proceed is.
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
More information about the llvm-commits