[llvm-dev] Your help needed: List of LLVM Open Projects 2017

Fri Feb 3 14:07:22 PST 2017

This is a little embarrassing -- I've missed this amazing discussion!

I definitely should put something about XRay in the Open Projects document.

Please see some thoughts in-lined below:

> On 17 Jan 2017, at 16:04, Xinliang David Li via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> right -- which is why I said it is complicated :)
> 

Definitely 100% complicated -- but not entirely impossible. :)

> David
> 
> On Mon, Jan 16, 2017 at 9:02 PM, Mehdi Amini <mehdi.amini at apple.com> wrote:
> That’s a very good point, the X-ray data that would be useful for the linker should be collected from a "profile-use” build to be accurate.
> 
> IIUC, that’d look like:   build with profile-gen -> run benchmarks -> build with profile-use + Xray -> run benchmarks -> build with profil-use and feed linker with X-ray data for layout optimization.
> 

Another interesting thing to think about here is the fact that you can plug in a different runtime logging implementation (what's not default in the xray runtime implementation) to collect the data needed instead of logging out full information. A "profile-collecting" log handler could be implemented in parallel to the full logging/trace mode, that could be more efficient in terms of this specific use-case.

> — 
> Mehdi
> 
>> On Jan 16, 2017, at 8:25 PM, Xinliang David Li via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>> 
>> Xray data are collected post-inlining, but the problem is that inlining decisions are different between profile-gen and profile-use compilations. This makes it hard to match call trace data from prof-gen run with post-inline callgraph of the profile-use build.
>> 

Yes, although we have been thinking about implementing some intrinsics that will allow front-ends (and even IR passes) to explicitly mark where an "entry"  or "exit" event will be potentially recorded from -- very useful for front-ends that do inlining before generating LLVM IR. We're still thinking about the semantics and implementation of these intrinsics (it's not as simple as generating the sleds in a MachineFunction pass) but we're confident a first attempt being worked on now for a custom event logging intrinsic (https://reviews.llvm.org/D27503) may pave the way for more interesting xray-specific intrinsics that do very interesting things.

>> David 
>> 
>> On Mon, Jan 16, 2017 at 7:41 PM, Sean Silva <chisophugis at gmail.com> wrote:
>> Would it make sense for xray instrumentation be part of -fprofile-generate? PGO will affect inlining decisions etc for the optimized binary, but the collected traces during the instrumented build would still have quite a bit of useful information.
>> 

I've thought about this long and hard, and I think there's a way to make the the profile generation mechanisms use the XRay infrastructure to implement the profile gathering pieces for PGO. As I mention above, we can choose at compile-time and/or at runtime which handler will be used by an XRay instrumented binary -- so not only can you turn on/off instrumentation arbitrarily, you can get profile information instead of getting the more expensive XRay logging implementation.

Of course there's work to be done here to make that happen, but most of the pieces are already there.

It may probably be better to discuss this specifically in another thread, lest we hijack this one too much. ;)

Cheers

>> -- Sean Silva
>> 
>> 
>> On Mon, Jan 16, 2017 at 4:33 PM, Xinliang David Li <xinliangli at gmail.com> wrote:
>> Google GCC records profile data (dynamic callgraph) in a special named section in ELF object file to be consumed by the plugin. Those sections will be discarded later by the linker.  
>> 
>> There are pros and cons of using xray for layout purpose.  The call trace from xray is certainly more powerful for layout purpose, but it adds addtional complexity to the optimized build process.  You would need to collect xray trace profile on the optimized binary (presumably built with PGO already) and rebuild without xray nop insertion and function layout. 
>> 
>> David
>> 
>> On Mon, Jan 16, 2017 at 3:40 PM, Sean Silva via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>> 
>> 
>> On Mon, Jan 16, 2017 at 3:34 PM, Sean Silva <chisophugis at gmail.com> wrote:
>> 
>> 
>> On Mon, Jan 16, 2017 at 3:32 PM, Sean Silva <chisophugis at gmail.com> wrote:
>> 
>> 
>> On Mon, Jan 16, 2017 at 2:31 PM, Davide Italiano <davide at freebsd.org> wrote:
>> On Mon, Jan 16, 2017 at 2:07 PM, Mehdi Amini <mehdi.amini at apple.com> wrote:
>> >
>> > On Jan 16, 2017, at 1:47 PM, Sean Silva <chisophugis at gmail.com> wrote:
>> >
>> >
>> >
>> > On Mon, Jan 16, 2017 at 1:25 PM, Davide Italiano <davide at freebsd.org> wrote:
>> >>
>> >> On Mon, Jan 16, 2017 at 12:31 PM, Sean Silva via llvm-dev
>> >> <llvm-dev at lists.llvm.org> wrote:
>> >> > Do we have any open projects on LLD?
>> >> >
>> >> > I know we usually try to avoid any big "projects" and mainly add/fix
>> >> > things
>> >> > in response to user needs, but just wondering if somebody has any ideas.
>> >> >
>> >>
>> >> I'm not particularly active in lld anymore, but the last big item I'd
>> >> like to see implemented is Pettis-Hansen layout.
>> >> http://perso.ensta-paristech.fr/~bmonsuez/Cours/B6-4/Articles/papers15.pdf
>> >> (mainly because it improves performances of the final executable).
>> >> GCC/gold have an implementation of the algorithm that can be used as
>> >> base. I'll expand if anybody is interested.
>> >> Side note: I'd like to propose a couple of llvm projects as well, I'll
>> >> sit down later today and write them.
>> >
>> >
>> >
>> > I’m not sure, can you confirm that such layout optimization on ELF requires
>> > -ffunction-sections?
>> >
>> 
>> For the non-LTO case, I think so.
>> 
>> > Also, for clang on OSX the best layout we could get is to order functions in
>> > the order in which they get executed at runtime.
>> >
>> 
>> That's what we already do for lld. We collect and order file (run a
>> profiler) and pass that to the linker that lays out functions
>> accordingly.
>> This is to improve startup time for a class of startup-time-sensitive
>> operations. The algorithm proposed by Pettis (allegedly) aims to
>> reduce the TLB misses as it tries to lay out hot functions (or
>> functions that are likely to  be called together near in the final
>> binary).
>> 
>> IIRC from when I looked at the paper a while ago, it is mostly just a "huffman tree construction" type algorithm (agglomerating based on highest probability) and assumes that if two functions are hot then they are likely to be needed together. This is not always the case.
>> 
>> E.g. consider a server that accepts RPC requests and based on those requests either does Foo or Bar which are largely disjoint. It's entirely possible for the top two functions of the profile to be one in Foo and one in Bar, but laying them out near each other doesn't make sense since there is never locality (for a given RPC, either Foo or Bar gets run). A static call graph analysis can provide the needed signals to handle this case better. 
>> 
>> 
>> Hence you said "allegedly" :) I know we've talked about this before. Just wanted to put the backstory of the "allegedly" on the list.
>> 
>> Looks like I remembered this wrong. The algorithm in section 3.2 of the paper is call-graph aware. It does do greedy coalescing like a Huffman tree construction algorithms, but constrains the available coalescing operations at each step by call graph adjacency (in fact, what it is "greedy" about is the hotness of the edges between call graph nodes and not the nodes themselves).
>> 
>> -- Sean Silva
>>  
>> 
>> -- Sean Silva
>>  
>> -- Sean Silva
>>  
>> 
>> >
>> > For FullLTO it is conceptually pretty easy to get profile data we need for
>> > this, but I'm not sure about the ThinLTO case.
>> >
>> > Teresa, Mehdi,
>> >
>> > Are there any plans (or things already working!) for getting profile data
>> > from ThinLTO in a format that the linker can use for code layout? I assume
>> > that profile data is being used already to guide importing, so it may just
>> > be a matter of siphoning that off.
>> >
>> >
>> > I’m not sure what kind of “profile information” is needed, and what makes it
>> > easier for MonolithicLTO compared to ThinLTO?
>> >
>> > Or maybe that layout code should be inside LLVM; maybe part of the general
>> > LTO interface? It looks like the current gcc plugin calls back into gcc for
>> > the actual layout algorithm itself (function call
>> > find_pettis_hansen_function_layout) rather than the reordering logic living
>> > in the linker:
>> > https://android.googlesource.com/toolchain/gcc/+/3f73d6ef90458b45bbbb33ef4c2b174d4662a22d/gcc-4.6/function_reordering_plugin/function_reordering_plugin.c
>> >
>> >
>> > I was thinking about this: could this be done by reorganizing the module
>> > itself for LTO?
>> >
>> > That wouldn’t help non-LTO and ThinLTO though.
>> 
>> This is a dimension that I think can be explored. The fact that it
>> wouldn't help with other modes of operation is completely orthogonal,
>> in particular until it's proven that this kind of optimization makes
>> sense with ThinLTO (and if it doesn't, it can be an optimization ran
>> only during full LTO).
>> 
>> --
>> Davide
>> 
>> "There are no solved problems; there are only problems that are more
>> or less solved" -- Henri Poincare
>> 
>> 
>> 
>> 
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>> 
>> 
>> 
>> 
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

-- Dean