[PATCH] [LoopReroll] Run LICM before Loop Rerolling.
hfinkel at anl.gov
hfinkel at anl.gov
Mon Feb 16 09:41:52 PST 2015
In http://reviews.llvm.org/D7674#124359, @jmolloy wrote:
> Hi Hal,
>
> Hmm yeah, good questions. I can do (1). I hadn't remembered that SE had a function for this - I was trying to avoid implementing LICM myself!
>
> But I'm not sure if I should do (1) because of (2). In my testcase, which is fairly difficult to reduce,
This is not difficult to reduce ;) -- add an assert in the loop reroller, and let bugpoint do it. We need to understand where this comes from so we can sensibly talk about adjusting the pass manager configuration.
> we end up with a stray invariant multiply which is not the cheapest thing in the world. Perhaps adding another LICM regardless of Rerolling is the right way to go after all?
I vote for both.
REPOSITORY
rL LLVM
http://reviews.llvm.org/D7674
EMAIL PREFERENCES
http://reviews.llvm.org/settings/panel/emailpreferences/
More information about the llvm-commits
mailing list