[PATCH] Properly detect working jobserver in gcc driver.

[Martin Liška]

On 8/2/19 11:55 AM, Richard Biener wrote:
> On Fri, Aug 2, 2019 at 11:19 AM Martin Liška <mliska@suse.cz> wrote:
>>
>> On 8/2/19 11:15 AM, Jan Hubicka wrote:
>>>> On Fri, Aug 2, 2019 at 10:50 AM Jakub Jelinek <jakub@redhat.com> wrote:
>>>>>
>>>>> On Fri, Aug 02, 2019 at 10:47:10AM +0200, Martin Liška wrote:
>>>>>>> Can you strace if other fds are opened and not closed in the spot you had it
>>>>>>> before?  Advantage of doing it there is that it will not be done for all the
>>>>>>> -E/-S/-c compilations when the linker is not spawned.
>>>>>>
>>>>>> I've used the same trick which you used and I'm attaching the output.
>>>>>> I believe it's fine, I can't see any opened fd by GCC.
>>>>>
>>>>> LGTM.
>>>>
>>
>> Hi.
>>
>>>> Btw, we discussed yesterday on the phone and the conclusion was to
>>>> make -flto auto-detect a job-server (but not fall back to # of threads)
>>>> and add -flto=auto to auto-detect a job-server and fall back to # of threads.
>>>> That basically makes -flto=jobserver the default behavior which means
>>>> we should document -flto=1 as a way to override jobserver detection.
>>>
>>> And concerning to the yesterday discussion, my preference would still be
>>> -flto to first try presence of jobserver and default to number of
>>> threads otherwise. It seems like user friendly default to me and other
>>> tools with reasonable parallelism support usually behaves this way, too.
>>
>> I also like the default as Honza defined.
>>
>>>
>>> Sure one can use -flto=auto everywhere but then one needs to use
>>> different options for differnt compilers.  It would be nice to make
>>> -flto to do right hting for majority of users including those like me
>>> who cut&paste command lines from Make output and execute them by hand.
>>
>> Note that our ambition is to ideally backport the patches to our gcc9 package.
>> The auto-detection of job-server with -flto is a behavior change that will happen
>> in the gcc9 package.
>>
>> To be honest I don't like the invention of 'auto' value for -flto command.
>> That would be useful just for gcc9 right now :/
> 
> Well.  We teached people they need to use -flto=N or -flto=jobserver to
> get parallelism.  Like we teached people they need to use -O2 to get
> any optimization.
> 
> Other compilers behave differently.  So what.
> 
> Auto-detecting threads is nice.  Suddenly trashing your system
> because you happen to invoke two links in parallel is not.
> Which is at least why changing this kind of behavior for 9.2 (or 9.3)
> vs. 9.1 is out of the question.  And I'm not sure it is OK for 10.
> 
> As with defaulting to use a job-server when present that's more
> reasonable.
> 
> For cut&pasting you then can use -flto=auto.
> 
> But it's just my opinion - I surely can adapt.

Hi.

Based on what was written, I feel OK with -flto=auto. I would
like to see jobserver detection to be only enabled with the 'auto'
option value. That will make it consistent in GCC 9.x branch.

Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

Ready to be installed?
Thanks,
Martin

> 
> Richard.
> 
>> Martin
>>
>>>
>>> The fork bombing is IMO relatively rare and it was not what Jakub ran
>>> into (it was broken jobserv detection).
>>> After all whole distro built with Martin's orriginal approach of
>>> -flto=<nthreads> which forkbombs on every possible occasion.
>>>
>>> Said that, I could live with more conservative defaults.
>>> Honza
>>>
>>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Add-flto-auto-option-value.patch
Type: text/x-patch
Size: 3815 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20190805/ff07e128/attachment.bin>