Some comments from the maintainer of parrot, p2 and perl B::CC which do similar things:
tracelets instead of basic blocks analysis sounds interesting, but php is still doomed by not allowing optional types. In-house code can easily be optimized by explicit types. The AUTOLOAD problem is a big one, and I am just planning to tackle it, but came to the same design decisions mostly. We are compiling modules, files as this is easiest to handle. My p2 jit has no type guards and seperate specialized methods yet, I rather support optional early binding, a jitted method cache and small tagged data, which doesn't fill up the cache that much. It outperforms java and clr by far, just luajit is ahead.
With the static B::CC, type inference has the same problem as php, but has the same performance advantages as hhpc, but I added special syntax for typed and sized arrays, and to disallow too much runtime magic. The current production compiler at Cpanel only uses better data layout to get its performance boost at startup and overall memory usage. Readonly strings and hash keys mostly. Perfect hashes not yet.
IMHO most important is smaller data and ops overhead, not the optimizer.
I'm migrating to HHVM across my whole stack. Massive performance gains, only issues have been with some plugins that I was able to pretty easily work around.
Good stuff! Uncached, HHVM outperforms my cached php-fpm sites.
I've been doing the same and so far have seen big performance gains as well. Using HHVM also has the benefit of being able to "tack on" any Hack code in the future if (for example) you have some important code which could benefit from static typing.
Question to the authors: Any reason why hippyvm is not included for comparison in the paper? It does usually outperform hhvm on those benchmarks (but not on real world use cases which is maybe a good reason to include real world use cases more into such papers).
This is awesome! I am not familiar with how compilers/VMs are generally implemented, but the tracelets and guards idea strikes me as very general - in particular, it looks like this approach could be applied to provide type inference to any dynamically typed language. Has this kind of thing been tried before?
In terms of optimization, is it possible to create tracelets that are not continuous regions in code? Roughly, if you identify two non contiguous tracelets, having the same inputs, and can guarantee the inputs havent changed in between, then you could merge them together. Because bigger tracelets would mean less guards and better performance.
As far as I know, you could try tracelets with any dynamic language.
And yes, you could try to share tracelets whose bodies are identical, but, unless their successor tracelets are also identical, you'd need to "dynamicise" the dispatch so you go down Path 1 when you're really tracelet 1 and Path 2 when you're really tracelet 2. We normally chain tracelets together by either falling through (if it's unconditional and the successor tracelet could be placed right next to it) or with jmp or branch instructions.
The tracing JIT was removed from Spidermonkey because it was brittle and didn't perform as well as the traditional JIT compiler(s) (Jaeger..., Ion...) that replaced it.
I assume latent type means simply the implicit typing. i.e. you declare a variable with the value, and the type is inferred through the value of the variable rather than the explicit type declared.
As usual, the abstract really isn't enough to draw big conclusions from. The concept of a latent type applies to arbitrary expressions, not just variables. All the values flowing through a PHP program implicitly share the same union type; they can be floats, strings, arrays, etc. The latent types are the narrower types that can actually flow through the program in practice.
To be clear, it is more general than just things like:
$a = 0; // $a is an int!
It includes learning that:
g(foo() . bar());
foo() and bar() return strings. Since none of this information is marked syntactically in PHP, and since it might actually be undecidable because of dynamic control flow in the callees, dynamic binding, etc., you really need to see the program run to do this stuff.
I'm wondering: how does this compare speed-wise to other real-world JITs. For example, node.js (V8). Are there any tricks used in your VM which are not used by V8, and conversely?
a. Does the HHVM JIT do anything that LuaJIT doesn't? I assume you are familiar with LuaJIT, as it is mentioned in the paper - and from a quick scan, the only two things I didn't recognize from LuaJIT were the refcount optimizations (not required by Lua GC) and guard relaxation.
b. Is HHIR tied to Php, or is it usable as a general purpose JIT backend? LuaJIT's IR is, (unfortunately for other languages) tied very strongly to Lua semantics.
It's a shame. As far as I can tell, the only proven multi-language JIT for FOSS platforms is still the JVM? (and Mono maybe?)
Despite a lot of that's gone in to Parrot, PyPy, v8, etc, none of these VMs seem to have really taken off beyond the language they were intended for. Somewhat more sad is that pretty much all de-facto default runtimes for popular dynamic languages (except Javascript) are still interpreters... the cost of being portable.
I wouldn't say the JVM is a proven multi-language JIT, FOSS or not.
It works for multiple languages that go the extra mile - e.g. Jython pays very dearly in performance because of the object model mismatch between Python and Java. And I haven't looked closely recently, but when Clojure first came out, the impedance mismatch between Clojure's persistent data structures and Java's mutable ones also had a ridiculous performance cost.
Mono is perhaps more deserving of that title, especially together with IKVM, (so, everything JVM), F#, IronPython and Boo. But note that everything is still shackled to the underlying object model.
But if anything, LLVM is the only proven multi-language JIT for FOSS platforms; it's method-at-a-time, which is great for staticly typed languages (whether those types are declared upfront or inferred), not so much for extremely dynamic languages.
Does anyone use Guile for anything other than Scheme? I thought all of the other languages were pretty much just half-baked proofs of concept (like Parrot, for that matter).
I'm using HHVM in production for a few weeks now and all I can say is I'm very happy with the results.
It's running stable (compiled under CentOS7) and I was able to cut the response times in half (~145ms with php-fpm, ~70ms HHVM).
The same idea has been done for psyco (tracelets there are called basic-block-at-a-time compilation). PyPy uses a tracing JIT instead to achieve better performance in cases where psyco was too hard.
tracelets instead of basic blocks analysis sounds interesting, but php is still doomed by not allowing optional types. In-house code can easily be optimized by explicit types. The AUTOLOAD problem is a big one, and I am just planning to tackle it, but came to the same design decisions mostly. We are compiling modules, files as this is easiest to handle. My p2 jit has no type guards and seperate specialized methods yet, I rather support optional early binding, a jitted method cache and small tagged data, which doesn't fill up the cache that much. It outperforms java and clr by far, just luajit is ahead.
With the static B::CC, type inference has the same problem as php, but has the same performance advantages as hhpc, but I added special syntax for typed and sized arrays, and to disallow too much runtime magic. The current production compiler at Cpanel only uses better data layout to get its performance boost at startup and overall memory usage. Readonly strings and hash keys mostly. Perfect hashes not yet. IMHO most important is smaller data and ops overhead, not the optimizer.