Re: Parallelize the compilation using Threads

[Giuliano Belinassi <giuliano.belinassi@usp.br>]

Hi,

Since gimple-match.c takes so long to compile, I was wondering if it
might be possible to reorder the compilation so we can push its
compilation early in the dependency graph.

I've attached a graphic showing what I mean and the methodology into
PR84402 (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84402).

Maybe there is a simple change that can be made into Makefile? Or maybe
an feature to Make itself to compute the elapsed time for each file and
create a better scheduling for the next compilation?

Giuliano.

On 11/19, Richard Biener wrote:
> On Fri, Nov 16, 2018 at 8:00 PM Giuliano Augusto Faulin Belinassi
> <giuliano.belinassi@usp.br> wrote:
> >
> > Hi! Sorry for the late reply again :P
> >
> > On Thu, Nov 15, 2018 at 8:29 AM Richard Biener
> > <richard.guenther@gmail.com> wrote:
> > >
> > > On Wed, Nov 14, 2018 at 10:47 PM Giuliano Augusto Faulin Belinassi
> > > <giuliano.belinassi@usp.br> wrote:
> > > >
> > > > As a brief introduction, I am a graduate student that got interested
> > > >
> > > > in the "Parallelize the compilation using threads"(GSoC 2018 [1]). I
> > > > am a newcommer in GCC, but already have sent some patches, some of
> > > > them have already been accepted [2].
> > > >
> > > > I brought this subject up in IRC, but maybe here is a proper place to
> > > > discuss this topic.
> > > >
> > > > From my point of view, parallelizing GCC itself will only speed up the
> > > > compilation of projects which have a big file that creates a
> > > > bottleneck in the whole project compilation (note: by big, I mean the
> > > > amount of code to generate).
> > >
> > > That's true.  During GCC bootstrap there are some of those (see PR84402).
> > >
> >
> > > One way to improve parallelism is to use link-time optimization where
> > > even single source files can be split up into multiple link-time units.  But
> > > then there's the serial whole-program analysis part.
> >
> > Did you mean this: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84402 ?
> > That is a lot of data :-)
> >
> > It seems that 'phase opt and generate' is the most time-consuming
> > part. Is that the 'GIMPLE optimization pipeline' you were talking
> > about in this thread:
> > https://gcc.gnu.org/ml/gcc/2018-03/msg00202.html
> 
> It's everything that comes after the frontend parsing bits, thus this
> includes in particular RTL optimization and early GIMPLE optimizations.
> 
> > > > Additionally, I know that GCC must not
> > > > change the project layout, but from the software engineering perspective,
> > > > this may be a bad smell that indicates that the file should be broken
> > > > into smaller files. Finally, the Makefiles will take care of the
> > > > parallelization task.
> > >
> > > What do you mean by GCC must not change the project layout?  GCC
> > > happily re-orders functions and link-time optimization will reorder
> > > TUs (well, linking may as well).
> > >
> >
> > That was a response to a comment made on IRC:
> >
> > On Thu, Nov 15, 2018 at 9:44 AM Jonathan Wakely <jwakely.gcc@gmail.com> wrote:
> > >I think this is in response to a comment I made on IRC. Giuliano said
> > >that if a project has a very large file that dominates the total build
> > >time, the file should be split up into smaller pieces. I said  "GCC
> > >can't restructure people's code. it can only try to compile it
> > >faster". We weren't referring to code transformations in the compiler
> > >like re-ordering functions, but physically refactoring the source
> > >code.
> >
> > Yes. But from one of the attachments from PR84402, it seems that such
> > files exist on GCC,
> > https://gcc.gnu.org/bugzilla/attachment.cgi?id=43440
> >
> > > > My questions are:
> > > >
> > > >  1. Is there any project compilation that will significantly be improved
> > > > if GCC runs in parallel? Do someone has data about something related
> > > > to that? How about the Linux Kernel? If not, I can try to bring some.
> > >
> > > We do not have any data about this apart from experiments with
> > > splitting up source files for PR84402.
> > >
> > > >  2. Did I correctly understand the goal of the parallelization? Can
> > > > anyone provide extra details to me?
> > >
> > > You may want to search the mailing list archives since we had a
> > > student application (later revoked) for the task with some discussion.
> > >
> > > In my view (I proposed the thing) the most interesting parts are
> > > getting GCCs global state documented and reduced.  The parallelization
> > > itself is an interesting experiment but whether there will be any
> > > substantial improvement for builds that can already benefit from make
> > > parallelism remains a question.
> >
> > As I agree that documenting GCC's global states is good for the
> > community and the development of GCC, I really don't think this a good
> > motivation for parallelizing a compiler from a research standpoint.
> 
> True ;)  Note that my suggestions to the other GSoC student were
> purely based on where it's easiest to experiment with paralellization
> and not where it would be most beneficial.
> 
> > There must be something or someone that could take advantage of the
> > fine-grained parallelism. But that data from PR84402 seems to have the
> > answer to it. :-)
> >
> > On Thu, Nov 15, 2018 at 4:07 PM Szabolcs Nagy <Szabolcs.Nagy@arm.com> wrote:
> > >
> > > On 15/11/18 10:29, Richard Biener wrote:
> > > > In my view (I proposed the thing) the most interesting parts are
> > > > getting GCCs global state documented and reduced.  The parallelization
> > > > itself is an interesting experiment but whether there will be any
> > > > substantial improvement for builds that can already benefit from make
> > > > parallelism remains a question.
> > >
> > > in the common case (project with many small files, much more than
> > > core count) i'd expect a regression:
> > >
> > > if gcc itself tries to parallelize that introduces inter thread
> > > synchronization and potential false sharing in gcc (e.g. malloc
> > > locks) that does not exist with make parallelism (glibc can avoid
> > > some atomic instructions when a process is single threaded).
> >
> > That is what I am mostly worried about. Or the most costly part is not
> > parallelizable at all. Also, I would expect a regression on very small
> > files, which probably could be avoided implementing this feature as a
> > flag?
> 
> I think the the issue should be avoided by avoiding fine-grained paralellism.
> Which might be somewhat hard given there are core data structures that
> are shared (the memory allocator for a start).
> 
> The other issue I am more worried about is that we probably have to
> interact with make somehow so that we do not end up with 64 threads
> when one does -j8 on a 8 core machine.  That's basically the same
> issue we run into with -flto and it's threaded WPA writeout or recursive
> invocation of make.
> 
> >
> > On Fri, Nov 16, 2018 at 11:05 AM Martin Jambor <mjambor@suse.cz> wrote:
> > >
> > > Hi Giuliano,
> > >
> > > On Thu, Nov 15 2018, Richard Biener wrote:
> > > > You may want to search the mailing list archives since we had a
> > > > student application (later revoked) for the task with some discussion.
> > >
> > > Specifically, the whole thread beginning with
> > > https://gcc.gnu.org/ml/gcc/2018-03/msg00179.html
> > >
> > > Martin
> > >
> >
> > Yes, I will research this carefully ;-)
> >
> > Thank you