whole program optimization examples?

Hello,

I was wondering if there is an example list somewhere of whole program optimizations done by LLVM based compilers?

I’m only familiar with method-level optimizations, and I’m being told wpo can deliver many great speedups.

My language is currently staticly typed JIT based and uses the JVM, and I want to move it over to LLVM so that I can have options where it can be ahead of time compiled as well.

I’m hearing bad things about LLVM’s JIT capabilities – specifically that writing your own GC is going to be a pain.

Anyways, sort of diverged there, but still looking for WPO examples!

Hayden.

While VMKit is retired, there are good examples of how to integrate
VMs and GCs into LLVM.

http://vmkit.llvm.org/

I'll let others chime in on the WPO examples. :slight_smile:

cheers,
--renato

I'm hearing bad things about LLVM's JIT capabilities -- specifically that
writing your own GC is going to be a pain.

While VMKit is retired, there are good examples of how to integrate
VMs and GCs into LLVM.

http://vmkit.llvm.org/

A better example would be the FTL project which is now part of WebKit. That's a production ready compiler while VMKit is not.

Hello,

I was wondering if there is an example list somewhere of whole program optimizations done by LLVM based compilers?

I'm only familiar with method-level optimizations, and I'm being told wpo can deliver many great speedups.

My language is currently staticly typed JIT based and uses the JVM, and I want to move it over to LLVM so that I can have options where it can be ahead of time compiled as well.

Depending on your use case (and frankly, your budget), you might want to consider Azul Zing's ReadyNow features: http://www.azulsystems.com/solutions/zing/readynow

This isn't true ahead of time compilation, but it would be a way to get most of the benefits of classic ahead of time compilation running on a standards compliant JVM.

(Keep in mind, I work for Azul. I may be slightly biased here.)

I'm hearing bad things about LLVM's JIT capabilities -- specifically that writing your own GC is going to be a pain.

Out of curiosity, where did you hear this?

We are actively working on improving the state of the world here. I'd suggest you take a look at the infrastructure patches currently up for review here: http://reviews.llvm.org/D5683

These will hopefully land within a week or two. At that point, the "gc infrastructure" part should be functional. You'd have to pick a GC (LLVM does not provide one), but you're frontend could emit barriers and statepoints (gc parseable callsites) and everything should work. (Well, modulo bugs! Which I want to know about so we can fix.)

There are a couple of options out there for pluggable GC libraries. The best well known is Boehm's conservative GC, but there are others.

Once that's in, we're planning on landing all of the late safepoint insertion logic we've been working on. This will enable full optimization of code for garbage collected languages - provided you meet a few requirements on the input IR. You can read about it here:
http://www.philipreames.com/Blog/tag/late-safepoint-placement/

And find the (slightly out of date) code here:
https://github.com/AzulSystems/llvm-late-safepoint-placement

Anyways, sort of diverged there, but still looking for WPO examples!

I'm curious to hear others take here as well. A few things that jump out at me: cross function escape analysis, alias analysis (in support of things like LICM), and cross function constant propagation. Not all of these work out of the box, but with work (sometimes on your side, sometimes an LLVM patch), interesting results can be had.

Fair warning, while getting an LLVM based JIT up and running at peak performance is a worthwhile endeavor (IMHO), it's also a fair amount of work. Getting something functional is relatively straight forward, but there's a lot of non-trivial tuning of your generated IR to really exploit the power of the optimizers well. We've talking person years of work here. Most of this is in the performance tuning phase, and depending on your point of comparison, it may be an easier or harder problem. Essentially, the closer to C performance your current runtime is, the harder you'll have to work. Getting 1/10 of C performance with an untuned LLVM based JIT is pretty easy; the closer you get to C (or JVM) performance the harder it gets.

(Disclaimer: This is me speaking off the top of my head. Take everything I just said with a grain of salt.)

Philip

Thanks, Philip for the “lay of the ground” picture. I think the situation I’m in, which represents my employment (and now personal technical curiosity) is that we’re seeing LLVM implementations show up like every other week or month, etc. and people are asking us, “well this mathematical software of yours is great, but my engineer here tells me it’s not using this LLVM thing, and I think we’re wasting cloud compute resources on by using the JVM technology” – this is how non-tech people are talking to me about this :slight_smile:

I heard the LLVM JIT situation from a bunch of my friends, one of whom was part of the Unladen Swallow effort and basically said – “Trust me, it’s not going to work, I put 2 years of my life every single day into it”.

But honestly, I personally am not familiar with writing a GC or what necessarily entails – I want to, and I can pick it up, but I spent most of time writing JVM based tooling, profilers, and byte code cachers, etc.

With regards to ReadyNow, I think at least someone on my team was looking at it.

In any case, I’ll be following your blog closely now!

Hello,

I was wondering if there is an example list somewhere of whole program optimizations done by LLVM based compilers?

I'm only familiar with method-level optimizations, and I'm being told wpo can deliver many great speedups.

My language is currently staticly typed JIT based and uses the JVM, and I want to move it over to LLVM so that I can have options where it can be ahead of time compiled as well.

As Philip kindly pointed out, WebKit uses llvm as part of a JavaScript JIT optimization pipeline. It works well for WebKit, but this was a large amount of work. It may not be the path of least resistance depending on what your requirements are.

I'm hearing bad things about LLVM's JIT capabilities -- specifically that writing your own GC is going to be a pain.

This is a fun topic and you'll probably get some good advice. :slight_smile:

Here's my take. GC in llvm is only a pain if you make the tragic mistake of writing an accurate-on-the-stack GC. Accurate collectors are only known to be beneficial in niche environments, usually if you have an aversion to probabilistic algorithms. You might also be stuck requiring accuracy if your system relies on being able to force *every* object to *immediately* move to a new location, but this is an uncommon requirement - usually it happens due to certain speculative optimization strategies in dynamic languages.

My approach is to use a Bartlett-style mostly-copying collector. If you use a Bartlett-style collector then you don't need any special support in llvm. It just works, it allows llvm to register-allocate pointers at will, and it lends itself naturally to high-throughput collector algorithms. Bartlett-style collectors come in many shapes and sizes - copying or not, mark-region or not, generational or not, and even a fancy concurrent copying example exists.

WebKit used a Bartlett-style parallel generational sticky-mark copying collector with opportunistic mark-region optimizations. We haven't written up anything about it yet but it is all open source.

Hosking's paper about the concurrent variant is here: http://dl.acm.org/citation.cfm?doid=1133956.1133963

I highly recommend reading Bartlett's original paper about conservative copying; it provides an excellent semi space algorithm that would be a respectable starting point for any VM. You won't regret implementing it - it'll simplify your interface to any JIT, not just llvm. It'll also make FFI easy because it allows the C stack to refer directly to GC objects without any shenanigans.

Bartlett is probabilistic in the sense that it may, with low probability, increase object drag. This happens rarely. On 64-bit systems it's especially rare. It's been pretty well demonstrated that Bartlett collectors are as fast as accurate ones, insofar as anything in GC land can be demonstrated (as in it's still a topic of lively debate, though I had some papers back in the day that showed some comparisons). WebKit often wins GC benchmarks for example, and we particularly like that our GC never imposes limitations on llvm optimizations. It's really great to be able to view the compiler and the collector as orthogonal components!

Thanks, Philip for the “lay of the ground” picture. I think the situation I’m in, which represents my employment (and now personal technical curiosity) is that we’re seeing LLVM implementations show up like every other week or month, etc. and people are asking us, “well this mathematical software of yours is great, but my engineer here tells me it’s not using this LLVM thing, and I think we’re wasting cloud compute resources on by using the JVM technology” – this is how non-tech people are talking to me about this :slight_smile:

I heard the LLVM JIT situation from a bunch of my friends, one of whom was part of the Unladen Swallow effort and basically said – “Trust me, it’s not going to work, I put 2 years of my life every single day into it”.

WebKit’s FTL JIT works great, so trust me, it does work. :slight_smile:

Hello,

I was wondering if there is an example list somewhere of whole program optimizations done by LLVM based compilers?

I'm only familiar with method-level optimizations, and I'm being told wpo can deliver many great speedups.

My language is currently staticly typed JIT based and uses the JVM, and I want to move it over to LLVM so that I can have options where it can be ahead of time compiled as well.

As Philip kindly pointed out, WebKit uses llvm as part of a JavaScript JIT optimization pipeline. It works well for WebKit, but this was a large amount of work. It may not be the path of least resistance depending on what your requirements are.

I'm hearing bad things about LLVM's JIT capabilities -- specifically that writing your own GC is going to be a pain.

This is a fun topic and you'll probably get some good advice. :slight_smile:

Here's my take. GC in llvm is only a pain if you make the tragic mistake of writing an accurate-on-the-stack GC. Accurate collectors are only known to be beneficial in niche environments, usually if you have an aversion to probabilistic algorithms. You might also be stuck requiring accuracy if your system relies on being able to force *every* object to *immediately* move to a new location, but this is an uncommon requirement - usually it happens due to certain speculative optimization strategies in dynamic languages.

I disagree with Filip in the details here. I don't believe that precise collectors are always better - they're not! - but the set of circumstances they're helpful for is larger than Filip admits.

However, we've had that discussion on the mailing list before, and it's far enough off topic to not be worth repeating in detail. :slight_smile:

A couple of quick reactions here: 1) PR is a bad reason to make a technical choice (unless your job depends on it) 2) If your bytecode is reasonably idiomatic, beating the JVM is a surprisingly high bar. Unless you’ve spent a lot of time tuning your existing implementation, that’s almost certainly a better use of technical resources. 3) LLVM could probably be made to work and even come out ahead, but it’s a large investment. (Keep in mind, I know nothing about your language and am speaking in generalities. Details matter here.) If your seriously interested in using LLVM, I strongly suggest attending the dev conference in a few weeks and chatting with folks in person. You’ll get a lot of valuable advice it’s hard to duplicate over email.

With the patchpoint infrastructure, shouldn’t it now be relatively straightforward to do an accurate-but-non-relocatable scan of the stack, by attaching all the GC roots as stackmap arguments to patchpoints? This is something we’re currently working on for Pyston (ie we don’t have it working yet), but I think we might get it “for free” once we finish the work on frame introspection.

I’m not aware of any high-performance conservative GC implementations that are designed to be pluggable (if there are please let us know!) – they typically seem pretty integrated with the VMs object model and language features that need to be supported. We’re spending some time right now to improve our GC situation, which is “a pain” since it’s more-or-less reinventing the wheel. It’s not made any harder by LLVM, but it’s tough in the sense that we’re not getting it for free like we would if we were on something like the JVM.

With the patchpoint infrastructure, shouldn’t it now be relatively straightforward to do an accurate-but-non-relocatable scan of the stack, by attaching all the GC roots as stackmap arguments to patchpoints?

Yes.

This is something we’re currently working on for Pyston (ie we don’t have it working yet), but I think we might get it “for free” once we finish the work on frame introspection.

I’m not aware of any high-performance conservative GC implementations that are designed to be pluggable (if there are please let us know!) – they typically seem pretty integrated with the VMs object model and language features that need to be supported.

If you want a high-performance GC, you will end up integrating it with the VM’s object model.

Take a look at the statepoint intrinsics up for review. These are essentially exactly that, with two extensions: - A semantic distinction between gc roots and deopt state (since you may want both) - Support for explicit relocation of the gc root values (this could be made optional, but is currently not) Though, you really don’t want to emit these in your frontend. You can, it’ll work, but the performance will suffer. Doing so will prevent many useful optimizations from running. Instead, you probably want to consider something like the late safepoint placement approach we’ve been pushing. Hopefully, once the statepoint stuff lands, we can get that upstreamed fairly soon. Philip

Sorry, wasn't trying to knock any of the non-pluggable GCs out there, was
just trying to attest to the OP's worry that "writing your own GC is going
to be a pain" -- even though this is a problem that's been solved
repeatedly and LLVM doesn't necessarily get in your way, there's still a
lot of work to get a good GC (let alone a whole VM) running on top of LLVM.

Take a look at the statepoint intrinsics up for review. These are essentially exactly that, with two extensions: - A semantic distinction between gc roots and deopt state (since you may want both) - Support for explicit relocation of the gc root values (this could be made optional, but is currently not) Though, you really don’t want to emit these in your frontend. You can, it’ll work, but the performance will suffer. Doing so will prevent many useful optimizations from running.

You really should be specific here. The optimizations you’re thinking of may be uninteresting to many clients.

Also you won’t lose any performance if your GC pointers are also needed for deopt (which happens to be the common case).

I really do think that this whole discussion is tragicomic. Most clients of LLVM would be best served with mostly copying GC.

-Filip

Thanks, it isn’t PR driven per se. But at least in some cases we are debating if the overhead of a VM is pulling us back.

Agreed on 2)

I’m now actively to get to this meeting in San Jose, but it appears full. We’ll see, maybe I’ll meet some folks.

Assuming you have a VM which needs safepoints to occur at some fixed interval, you need to put a safepoint poll in all loops. (Well, unless you can prove either a) the loop is bounded or b) there’s another safepoint in the loop.) Doing so, you introduce a call into the loop (using either approach). This breaks loop recognition, complicates alias analysis and thus LICM, and is otherwise bad for the optimizer. I believe LLVM should not take a position in this debate and should try to support all collectors.

Assuming you have a VM which needs safepoints to occur at some fixed interval, you need to put a safepoint poll in all loops. (Well, unless you can prove either a) the loop is bounded or b) there’s another safepoint in the loop.) Doing so, you introduce a call into the loop (using either approach). This breaks loop recognition, complicates alias analysis and thus LICM, and is otherwise bad for the optimizer.

A multithreaded high-throughout VM will have deopt safe points in any loop that isn’t proven to terminate in a timely fashion. Those safe points will take all live state and this will be superset of your GC pointers.

Also you won’t lose any performance if your GC pointers are also needed for deopt (which happens to be the common case).

I really do think that this whole discussion is tragicomic. Most clients of LLVM would be best served with mostly copying GC.

I believe LLVM should not take a position in this debate and should try to support all collectors.

It’s good to encourage people to use state-of-the-art, easy-to-implement techniques rather than unnecessarily complicated ones.

-Filip