Good read. But a word of caution - the "JIT vs interpreter" comparisons often favor the interpreter when the JIT is inplemented as more-or-less simple inlining of the interpreter code. (Here called "copy-and-patch" but a decades-only approach). I've had fairly senior engineers try to convince me that this is true even for Java VMs. It's not in general, at least not with the right kind of JIT compiler design.
I just recently upgraded[1] a JIT that essentially compiled each bytecode separately to one that shares registers within the same basic block. Easy 40 percent improvement to runtime, as expected.
But something I hadn't expected was it also improved compilation time by 40 percent too (fewer virtual registers made for much faster register allocation).
This is an embarrassing context to admit, but here goes.
Back when Parrot was a thing and the Perl 6 people were targeting it, I profiled the prelude of Perl 6 to optimize startup time and discovered two things:
- the first basic block of the prelude was thousands of instructions long (not surprising)
- the compiler had to allocate thousands of registers because the prelude instructions used virtual registers
The prelude emitted two instructions, one right after another: load a named symbol from a library, then make it available. I forget all of the details, but each of those instructions either one string register and one PMC register. Because register allocation used the dominance frontier method, the size of the basic block and total number of all symbolic registers dominated the algorithm.
I suggested a change to the prelude emitter to reuse actual registers and avoid virtual registers and compilation sped up quite a bit.
Yeah, I expect the real advantage of a JIT is that you can perform proper register allocation and avoid a lot of stack and/or virtual register manipulation.
I wrote a toy copy-patch JIT before and I don't remember being impressed with the performance, even compared to a naive dispatch loop, even on my ~11 year old processor.
The difference between interpreters and simple JITs has narrowed partly due to two factors: better indirect branch predictors with global history, and wider execution bandwidth to absorb the additional dispatch instructions. Intel CPUs starting with Haswell, for instance, show less branch misprediction impact due to better ability to predict jump path patterns through the interpreter. A basic jump table no longer suffers as much compared to tail-calling/dispatch or a simple splicing JIT.
Exactly, and it's not just register allocatio: but for many languages also addign proper typing, some basic data flow optimization, some constant folding, and a few other things that can be done fairly quickly, without the full set of trees and progressive lowering of the operators down to instruactions.
What's odd about the "JIT vs interpreter" debate is that it keeps coming up, given that it is fairly easy to see even in toy examples.
A shame operating systems like iOS/iPadOS do not allow JIT. iPad Pro's have such fast CPU's that you cant even use fully because of decisions like this.
They allow, but Apple's policy is to lock down that ability pretty much just to Safari/WKWebView. If you could transpile/compile your program to JS or WASM and run it through one of these blessed options, it should get JIT'ted.
They do, technically, allow JIT. You need a very hard-to-obtain entitlement that lets you turn writable pages into executable read-only pages, and good luck getting that entitlement if (for some reason) your name isn’t “mobilesafari”, but the capability exists.
When you say it's "hard" to obtain--is it possible to obtain if you aren't Apple? Does Apple ever provide it to third party developers, or is there even a path to requesting it?
JIT compilation can be faster for compiled languages too, as it allows data driven inlining and devirtualization, as well as "effective constant" propogation and runtime architecture feature detection
To re-optimize compiled code blocks isn't without effort. Google has publicly spoken about AutoFDO and Propeller [0], after Meta had open sourced BOLT [1] in 2021.
AutoFDO has since been ported to Android and adopted by Yandex [3].
Sure, but the relevant comparison isn't between languages: it's between a state-of-the-art JIT implementation of one language and a likewise-state-of-the-art AOT implementation of the same language. Unfortunately there aren't many examples of this; most languages have a preferred implementation strategy that receives much more effort than the other one.
I get that, but what interpreted language do you want to write iOS apps in when there's Swift and Obj-C right there, with bespoke support and tooling from Apple?
And if you care about performance, why aren't you writing that code in native to begin with?
> This is called branch prediction, it has been the source of many fun security issues...
No, that's speculative execution you just described. Branch prediction was implemented long before out-of-order CPUs were a thing, as you need branch prediction to make the most of pipelining (eg. fetching and decoding a new instruction while you're still executing the previous one--if you predict branches, you're more likely to keep the pipeline full).
Speculative execution does not require out-of-order execution. When you predict a branch, you're speculatively executing the predicted branch. Whether you're doing it in the same order as instruction order or out of order is independent of that.
The article is talking about OoO which is why I mentioned it. My point is that branch prediction and speculative execution are different things. You can do speculative execution without a branch predictor (run both branches and throw out the one that's wrong).
You're starting them in order and you're ending (retiring) them in order, but you're not necessarily ending one instruction before you're starting the next one. For instance, in a very simple pipeline, you can start decoding the next instruction before you've completed the previous one, so you can do some work in parallel.
The fetch stage of the pipeline will have needed to predict the branch N cycles before the execute stage of the pipeline actually gets around to evaluating it, in order to continue fetching the post-branch instructions. Without branch prediction the fetch stage would need to stall until that happens, which decreases throughput. The point of branch prediction and the subsequent speculative execution is to optimistically avoid that stall.
Essentially all microarchtectural state is fodder for side channel exploits.
Static branch prediction like "predict taken if negative branch offset" doesn't leak anything, but just about any dynamically updated tables will (almost tautologically) contain statistical information about what was executed recently.
I'm not really interested in building an interpreter, but the part about scalar out of order execution got me thinking. The opcode sequencing logic of an interpreter is inherently serial and an obvious bottleneck (step++; goto step->label; requires an add, then a fetch and then a jump, pretty ugly).
Why not do the same thing the CPU does and fetch N jump addresses at once?
Now the overhead is gone and you just need to figure out how to let the CPU fetch the chain of instructions that implement the opcodes.
You simply copy the interpreter N times, store N opcode jump addresses in N registers and each interpreter copy is hardcoded to access its own register during the computed goto.
You run into the same problem a CPU does: if you have dependencies between the instructions, you can't execute ahead of time. Your processor has a bunch of hardware to efficiently resolve conflicts but your interpreter does not.
The thing you're suggesting makes sense, but it's far more efficient to do in hardware. You might say that you could do it on one of the many cores available on your modern processor, but it turns out that synchronizing them to your main thread is really inefficient -- and anyway, they're busy running your HN browser threads and your YouTube music video.
Depending on the bytecode, instructions might be variable-length, which means that you need to execute a nontrivial amount of logic to fetch more than just the next bytecode or handler. That said, I tinkered with adding a prefetch to Wizard's interpreter which basically moves the load of the next handler from the dispatch at the end to the first thing in the handler, and saw something like a 5% improvement.
From the previous article in the series, it looks like the biggest impediment to just using full llvm to compile the query is that they didn't find a good way to cache the results across invocations.
Sql server hekaton punted this problem in a seemingly effective way by requiring the client to use stored procedures to get full native compilation. Not sure though if they recompile if the table statistics indicate a different query plan is needed.
Doesn't work for every case, but I think for a lot of cases nowadays, if you are using an interpreter and its slow, you should just generate web assembly. Libraries like walrus for rust make this pretty easy to do, and wasmtime provides a serviceable standalone runtime. For my little language, recursive fib(40) executes in firefox with wasm in about 600ms. My interpreter basically can't finish it.
The paper is 10 years old. While the gap between a threaded an interpreter (a dispatch at the end of every handler) versus non-threaded (loop over switch) isn't as big as it used to be, it's still 15-30% on modern very fast interpreters. For example, I measured between 14 and 29% performance improvement for threading Wizard's interpreter[1].
Interesting paper :) I've kept choosing threaded myself, but would have put the gap in a 5-10% range. I guess the branch predictor hasn't kept up. (Also trying to resist getting nerdsniped into measuring it myself 0_0)
Good read. But a word of caution - the "JIT vs interpreter" comparisons often favor the interpreter when the JIT is inplemented as more-or-less simple inlining of the interpreter code. (Here called "copy-and-patch" but a decades-only approach). I've had fairly senior engineers try to convince me that this is true even for Java VMs. It's not in general, at least not with the right kind of JIT compiler design.
I just recently upgraded[1] a JIT that essentially compiled each bytecode separately to one that shares registers within the same basic block. Easy 40 percent improvement to runtime, as expected.
But something I hadn't expected was it also improved compilation time by 40 percent too (fewer virtual registers made for much faster register allocation).
[1] https://github.com/ZQuestClassic/ZQuestClassic/commit/68087d...
This is an embarrassing context to admit, but here goes.
Back when Parrot was a thing and the Perl 6 people were targeting it, I profiled the prelude of Perl 6 to optimize startup time and discovered two things:
- the first basic block of the prelude was thousands of instructions long (not surprising) - the compiler had to allocate thousands of registers because the prelude instructions used virtual registers
The prelude emitted two instructions, one right after another: load a named symbol from a library, then make it available. I forget all of the details, but each of those instructions either one string register and one PMC register. Because register allocation used the dominance frontier method, the size of the basic block and total number of all symbolic registers dominated the algorithm.
I suggested a change to the prelude emitter to reuse actual registers and avoid virtual registers and compilation sped up quite a bit.
Yeah, I expect the real advantage of a JIT is that you can perform proper register allocation and avoid a lot of stack and/or virtual register manipulation.
I wrote a toy copy-patch JIT before and I don't remember being impressed with the performance, even compared to a naive dispatch loop, even on my ~11 year old processor.
The difference between interpreters and simple JITs has narrowed partly due to two factors: better indirect branch predictors with global history, and wider execution bandwidth to absorb the additional dispatch instructions. Intel CPUs starting with Haswell, for instance, show less branch misprediction impact due to better ability to predict jump path patterns through the interpreter. A basic jump table no longer suffers as much compared to tail-calling/dispatch or a simple splicing JIT.
Exactly, and it's not just register allocatio: but for many languages also addign proper typing, some basic data flow optimization, some constant folding, and a few other things that can be done fairly quickly, without the full set of trees and progressive lowering of the operators down to instruactions.
What's odd about the "JIT vs interpreter" debate is that it keeps coming up, given that it is fairly easy to see even in toy examples.
Turns out one of the classic papers on this is available for those interested in this discussion - https://news.ycombinator.com/item?id=45582127
A shame operating systems like iOS/iPadOS do not allow JIT. iPad Pro's have such fast CPU's that you cant even use fully because of decisions like this.
Those operating systems allow it, but Apple does not. Agree that it is a total waste.
They allow, but Apple's policy is to lock down that ability pretty much just to Safari/WKWebView. If you could transpile/compile your program to JS or WASM and run it through one of these blessed options, it should get JIT'ted.
They do, technically, allow JIT. You need a very hard-to-obtain entitlement that lets you turn writable pages into executable read-only pages, and good luck getting that entitlement if (for some reason) your name isn’t “mobilesafari”, but the capability exists.
When you say it's "hard" to obtain--is it possible to obtain if you aren't Apple? Does Apple ever provide it to third party developers, or is there even a path to requesting it?
Yes.
Source? Is there any non-Apple app that has this entitlement?
If your app happens to be a browser that's only usable in the EU then:
https://developer.apple.com/documentation/browserenginekit/p...
I believe the Delta emulator has JIT support, but possibly only when installed as a developer.
As far as I can tell, you need to connect your phone to a PC running software which enables JIT by exploiting a feature intended for remote debugging. https://faq.altstore.io/altstore-classic/enabling-jit
[dead]
What advantage does JIT compilation have over Swift or Obj-C?
JIT compilation can be faster for compiled languages too, as it allows data driven inlining and devirtualization, as well as "effective constant" propogation and runtime architecture feature detection
It can be but it never is.
To re-optimize compiled code blocks isn't without effort. Google has publicly spoken about AutoFDO and Propeller [0], after Meta had open sourced BOLT [1] in 2021.
AutoFDO has since been ported to Android and adopted by Yandex [3].
[0] https://lwn.net/Articles/995397/
[1] https://news.ycombinator.com/item?id=40868224
[2] https://news.ycombinator.com/item?id=42896716
But is that because of JIT compilation or other decisions for how the language should work (dynamic typing, GC, etc.)?
Especially after PGO (profiling guided optimization) gets most of the way there
Is there a production JIT for a compiled language that is actually faster? I understand the theory, I don't think the practice backs it up.
I believe HotSpot is usually faster than GCJ.
Depends, what do you consider Java?
Java is certainly not the fastest language out there.
Sure, but the relevant comparison isn't between languages: it's between a state-of-the-art JIT implementation of one language and a likewise-state-of-the-art AOT implementation of the same language. Unfortunately there aren't many examples of this; most languages have a preferred implementation strategy that receives much more effort than the other one.
It speeds up interpreted languages.
I get that, but what interpreted language do you want to write iOS apps in when there's Swift and Obj-C right there, with bespoke support and tooling from Apple?
And if you care about performance, why aren't you writing that code in native to begin with?
Why would you care about faster cars when planes exist?
And emulation.
What is an architecture but a scripting language to interpret? ;)
Of course they do - this is how javascript in any site works
> that you cant even use fully because of decisions like this.
Have no clue what this means - you can pre-compile for target platforms and therefore "fully" use whichever Apple device CPU.
You can't precompile your web page's Javascript for iOS, even if you're willing to have it signed and notarized and submitted for policy review.
Apart from the fact that its JS engine is really fast, Safari accepts WebAssembly. What else would you precompile it to?
Originally, iOS Safari handled WASM but only with JIT disabled.
However the EU decreed that it must allow for fair competition, leading to it claiming that it will enable JIT for authorized developers: https://developer.apple.com/support/alternative-browser-engi...
But I'm not sure that they have done so...
Mozilla: https://github.com/mozilla/platform-tilt/issues/3
Chrome: https://issues.chromium.org/issues/42203058
> This is called branch prediction, it has been the source of many fun security issues...
No, that's speculative execution you just described. Branch prediction was implemented long before out-of-order CPUs were a thing, as you need branch prediction to make the most of pipelining (eg. fetching and decoding a new instruction while you're still executing the previous one--if you predict branches, you're more likely to keep the pipeline full).
Speculative execution does not require out-of-order execution. When you predict a branch, you're speculatively executing the predicted branch. Whether you're doing it in the same order as instruction order or out of order is independent of that.
The article is talking about OoO which is why I mentioned it. My point is that branch prediction and speculative execution are different things. You can do speculative execution without a branch predictor (run both branches and throw out the one that's wrong).
You're right, I missed the article specifically mentions Meltdown in that sentence, not Spectre.
If you're executing instructions in order, wouldn't you already know the result of the branch by the time you reach its code?
You're starting them in order and you're ending (retiring) them in order, but you're not necessarily ending one instruction before you're starting the next one. For instance, in a very simple pipeline, you can start decoding the next instruction before you've completed the previous one, so you can do some work in parallel.
The fetch stage of the pipeline will have needed to predict the branch N cycles before the execute stage of the pipeline actually gets around to evaluating it, in order to continue fetching the post-branch instructions. Without branch prediction the fetch stage would need to stall until that happens, which decreases throughput. The point of branch prediction and the subsequent speculative execution is to optimistically avoid that stall.
A good explanation of branch prediction: https://danluu.com/branch-prediction/
Essentially all microarchtectural state is fodder for side channel exploits.
Static branch prediction like "predict taken if negative branch offset" doesn't leak anything, but just about any dynamically updated tables will (almost tautologically) contain statistical information about what was executed recently.
That was a pretty interesting read.
My take is that you can get pretty far these days with a simple bytecode interpreter. Food for thought if your side project could benefit from a DSL!
As I noted in another comment, I would like to persuade you that this is not the right take-away..
I'm not really interested in building an interpreter, but the part about scalar out of order execution got me thinking. The opcode sequencing logic of an interpreter is inherently serial and an obvious bottleneck (step++; goto step->label; requires an add, then a fetch and then a jump, pretty ugly).
Why not do the same thing the CPU does and fetch N jump addresses at once?
Now the overhead is gone and you just need to figure out how to let the CPU fetch the chain of instructions that implement the opcodes.
You simply copy the interpreter N times, store N opcode jump addresses in N registers and each interpreter copy is hardcoded to access its own register during the computed goto.
You run into the same problem a CPU does: if you have dependencies between the instructions, you can't execute ahead of time. Your processor has a bunch of hardware to efficiently resolve conflicts but your interpreter does not.
The thing you're suggesting makes sense, but it's far more efficient to do in hardware. You might say that you could do it on one of the many cores available on your modern processor, but it turns out that synchronizing them to your main thread is really inefficient -- and anyway, they're busy running your HN browser threads and your YouTube music video.
Depending on the bytecode, instructions might be variable-length, which means that you need to execute a nontrivial amount of logic to fetch more than just the next bytecode or handler. That said, I tinkered with adding a prefetch to Wizard's interpreter which basically moves the load of the next handler from the dispatch at the end to the first thing in the handler, and saw something like a 5% improvement.
I was surprised to learn that postgresql does not have a compiled plan / query cache.
Also I recommend reading the previous blog post first, then this one, for additional context: https://www.pinaraf.info/2024/03/look-ma-i-wrote-a-new-jit-c...
From the previous article in the series, it looks like the biggest impediment to just using full llvm to compile the query is that they didn't find a good way to cache the results across invocations.
Sql server hekaton punted this problem in a seemingly effective way by requiring the client to use stored procedures to get full native compilation. Not sure though if they recompile if the table statistics indicate a different query plan is needed.
Doesn't work for every case, but I think for a lot of cases nowadays, if you are using an interpreter and its slow, you should just generate web assembly. Libraries like walrus for rust make this pretty easy to do, and wasmtime provides a serviceable standalone runtime. For my little language, recursive fib(40) executes in firefox with wasm in about 600ms. My interpreter basically can't finish it.
The issues with branch prediction aren't really as much of a thing in modern interpreters, I can really recommend reading https://inria.hal.science/hal-01100647/document
The paper is 10 years old. While the gap between a threaded an interpreter (a dispatch at the end of every handler) versus non-threaded (loop over switch) isn't as big as it used to be, it's still 15-30% on modern very fast interpreters. For example, I measured between 14 and 29% performance improvement for threading Wizard's interpreter[1].
[1] https://dl.acm.org/doi/10.1145/3563311
Interesting paper :) I've kept choosing threaded myself, but would have put the gap in a 5-10% range. I guess the branch predictor hasn't kept up. (Also trying to resist getting nerdsniped into measuring it myself 0_0)
Testing this in Wizard is fairly easy.
Compare the running speed of the two binaries built with different options: