On 20/06/17 05:58, Boris Zbarsky wrote: > On 6/19/17 11:22 PM, Gregory Szorc wrote: >> The decision to strip Nightly builds does not come lightly. Read 1338651 >> comment 111 and later for the ugly backstory. > > It's still really confusing to me that not stripping symbols has a significant > performance impact. That's not the case in any other build configuration I'm > aware of, and is somewhat surprising from first principles for everything > except startup performance. > > It really would be good to figure out what's actually going on there...
I agree. Stripping the symbols as a solution makes no sense to me, given that they are not expected to be loaded into the process image. >From my scan of 1338651 it appears that we've demonstrated that the same preprocessed source is compiled in both cases. But IIUC (and correct me if I'm wrong), we haven't shown that either the same code is generated, nor that there is not some different interaction with the underlying machine for the two builds. One thing we could do is profile both with VTune, to look at both * mispredicts, cache misses (D and I), alignment stalls, whatever .. in short, anything that affects IPC. * average instruction counts for some "known" hot functions, which we can reliably compare between builds. That might show cases where the compiler didn't generate identical code. I've used VTune on Linux and have some idea what it can and can't do. I have tried it on Mac, but my impression, from the Intel web site, is that it is at least available for Mac. J _______________________________________________ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
