On Thu, Dec 25, 2014 at 11:55 PM, Adve, Vikram Sadanand
> Diego, Teresa, David,
> Sorry for my delayed reply; I left for vacation right after sending my
message about this.
> Diego, it wasn't explicit from your message whether LLVM LTO can handle
Firefox-scale programs, which you said GCC can handle. I assumed that's
what you meant, but could you confirm that? I understand that neither can
handle the very large Google applications, but that's probably not a
near-term concern for a project like the one Charles is embarking on.
Vikram, LLVM can handle Firefox size programs. Honza wrote two good
articles about LTO.
Comparison with LLVM is described in the second article. It took about
40min to finish building Firefox with llvm using lto and -g. The
following is a quote:
"This graph shows issues with debug info memory use. LLVM goes up to
35GB. LLVM developers are also working on debug info merging
improvements (equivalent to what GCC's type merging is) and the
situation has improved in last two releases until the current shape.
Older LLVM checkouts happily run out of 60GB memory & 60GB swap on my
> I'd be interested to hear more about the LTO design you folks are working
on, whenever you're ready to share the details.
We will share the details as soon as we can -- possibly some time in Jan
> I read the GCC design docs on LTO, and I'm curious how similar or
different your approach will be. For example, the 3-phase approach of
WHOPR is fairly sophisticated (it actually follows closely some research
done at Rice U. and IBM on scalable interprocedural analysis, in the same
group where Preston did his Ph.D.).
In Google, we care mostly about peak optimization performance. Peak
Optimization is basically PGO + CMO. For cross-module optimization
(CMO) to be usable for large applications, small memory footprint is
just one aspect of it, and fast build time is equally important. Peak
optimization is not only used in release build but in developer
workflow too. This means build time with CMO should be close to O2
time as much as possible. It is important to compiler engineers too
-- you don't want to wait for more than 20min to hit a breakpoint in
debugging a compiler problem :)
For this reason, GCC LTO is not used in Google. Instead, the much more
scalable solution called LIPO is widely used for CMO:
LIPO by design requires PGO.
While LIPO is scalable, it has its own limitation that prevents the
compiler from maximizing the benefit of CMO. The new design is
intended to solve the problem with more very aggressive objectives.
The new design is pretty simple and shares the basic principles of
LIPO without requiring PGO (though it still works best with PGO). It
still fits in LTO framework, so that toolchain support change is
minimized. For now, without giving details, I can share some of the
objectives of the new design:
* The build should be almost fully parallelizable (at both process
level and build machine node level)
* The build should scale to programs with *any/unlimited* size
(measured in number of TUs). It should handle programs 10x, 100x the
size of Firefox.
* The build time should be very close to non-LTO build, and can be
considered to be turned on *by default* for O2 or at least O3
* When turned on the by default, it can eliminate the need for
users to put inline functions in header files (thus greatly help
improving parsing time)
* Most of the benefit of CMO comes from cross module inlining and
cross module indirect call promotions. By default, the design only
enables these two, but it is still compatible with whole program
analysis which can be turned on with additional option.
> For now, I would like to introduce you all to Charles, so that he has
access to people working on this issue, which will probably continue to be
a concern for his project. I have copied you on my reply to him.
thanks for introduction! I am interested in knowing more about
> Thanks for the information.
> --Vikram S. Adve
> Visiting Professor, Computer Science, EPFL
> Professor, Department of Computer Science
> University of Illinois at Urbana-Champaign
> [email protected]
> On Dec 16, 2014, at 3:48 AM, Teresa Johnson wrote:
>> On Fri, Dec 12, 2014 at 1:59 PM, Diego Novillo
>>> On 12/12/14 15:56, Adve, Vikram Sadanand wrote:
>>>> I've been asked how LTO in LLVM compares to equivalent capabilities
>>>> in GCC. How do the two compare in terms of scalability? And
>>>> robustness for large applications?
>>> Neither GCC nor LLVM can handle our (Google) large applications.
>>> just too massive for the kind of linking done by LTO.
>>> When we built GCC's LTO, we were trying to address this by creating a
>>> partitioned model, where the analysis phase and the codegen phase are
>>> to allow working on partial callgraphs
>>> This allows to split and parallelize the initial bytecode generation
>>> final optimization/codegen. However, the analysis is still implemented
>>> single process. We found that we cannot even load summaries, types and
>>> symbols in an efficient way.
>>> It does allow for programs like Firefox to be handled. So, if by "big"
>>> need to handle something of that size, this model can doit.
>>> With LLVM, I can't even load the IR for one of our large programs on a
>>> with 64Gb of RAM.
>>>> Also, are there any ongoing efforts or plans to improve LTO in LLVM
>>>> in the near future?
>>> Yes. We are going to be investing in this area very soon. David and
>>> (CC'd) will have details.
>> Still working out the details, but we are investigating a solution
>> that is scalable to very large programs. We'll share the design in the
>> near future when we have more details worked out so that we can get
>> Teresa Johnson | Software Engineer | [email protected] | 408-460-2413