[llvm] r256302 - Disable use list order on the gold plugin.

Mehdi Amini via llvm-commits llvm-commits at lists.llvm.org
Thu Jan 14 15:33:42 PST 2016


> On Jan 14, 2016, at 2:54 PM, Rafael Espíndola <rafael.espindola at gmail.com> wrote:
> 
> On 12 January 2016 at 17:15, Mehdi Amini <mehdi.amini at apple.com> wrote:
>> 
>>> On Dec 22, 2015, at 3:45 PM, Rafael Espindola via llvm-commits <llvm-commits at lists.llvm.org> wrote:
>>> 
>>> Author: rafael
>>> Date: Tue Dec 22 17:45:49 2015
>>> New Revision: 256302
>>> 
>>> URL: http://llvm.org/viewvc/llvm-project?rev=256302&view=rev
>>> Log:
>>> Disable use list order on the gold plugin.
>>> 
>>> It turns out that his is *really* slow. With this change the link of
>>> clang with plugin-opt=emit-llvm goes from 41 to 26 seconds.
>> 
>> plugin-opt=emit-llvm is a debug option, I’m not sure it is a good idea to “optimize” for this case and losing some reproducibility?
> 
> I have mixed feeling about it.
> 
> I honestly don't remember the specifics, but I noticed that creating a
> .bc for all of clang was noticeably slower than what it used to be.
> 
> If this option had a small impact or if use order problems were very
> common, I would agree that we should leave it enabled. But the option
> costs more than a third of the total time. I also only remember
> hitting a use list order issue once.

But this is not on the regular path, right?
You don’t need to dump the bitcode for your regular LTO build, this code path should only be used exactly when you want to do some debugging. What other use-case is there?

— 
Mehdi




> 
> So if there is strong preferences I can put it back, if possible
> behind a plugin option.
> 
> Cheers,
> Rafael



More information about the llvm-commits mailing list