Disable arm-gepopt by default
ghoflehner at apple.com
Mon Apr 20 20:34:42 PDT 2015
It shouldn’t come across as ‘extreme’. Based on the data we have (performance, code size, disabling of test cases at commit time) we hope to come to a mutual agreement that disabling/backout is the way to go. If there is data to the contrary we need to look into. But if there is no gain from an optimization for anyone, no none should have to deal with it.
> On Apr 20, 2015, at 7:39 PM, Tim Northover <t.p.northover at gmail.com> wrote:
> On 20 April 2015 at 19:03, Gerolf Hoflehner <ghoflehner at apple.com> wrote:
>> from a birds view this optimization needs more work, quite possibly more so
>> in other passes rather than the algorithm itself. In our environment we
>> haven’t seen *any* gains from it. Could you and/or Hao take this on after
>> the code is disabled/backed out?
> To be fair, the chances are we've only tested it on a single CPU
> family; and from what I see in this thread only in artificial
> benchmarks. I have a great deal of sympathy with James's position, and
> think "we're disabling it, you & Hao get to pick up the pieces" is a
> bit extreme.
More information about the llvm-commits