Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Writing a basic x86-64 JIT compiler from scratch in stock Python

csl.name

Post reply on HN
Print view View original

41–50 of 52 posts

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#41

Post by csl »

defined wrote: post #38

This is a very interesting and well-written article, but I am puzzled by the technique being described as "JIT compilation". Isn't JIT compilation the process of compiling some sort of intermediate code into machine code at runtime ? For example, compiling JVM bytecode into machine code during execution, and using the cached machine code transparently on subsequent calls to that segment of code. Not to detract from t…

I do see your point, but I tried to make it clear that this is compilation in an extremely restricted sense: It specializes and executes a function at runtime. And I wanted to cover the core technique of doing just that.

Perhaps I should have included a small example of compiling a linear list of abstract instructions into machine code. But that again would consist of compiling small templates in a coherent way, just like many actual compilers do.

Anyway, point taken, and maybe I'll expand the article or follow up on it.

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#42

Post by csl »

shakna wrote: post #25
csl wrote: post #19

Earlier quoted context omitted.

Is it really that friendly when starting out, though? I found it brilliant, but quite the challenging read. I'd rather just read chapters four and five of SICP [1] and move on to Queinnec as a follow-up on more advanced topics. [1]: The book is free as well: https://mitpress.mit.edu/sicp/full-text/book/book-Z-H-4.html...

To be honest, I found LISP to be a lot easier than SICP, but that's probably just some learning styles stuff at play. SICP being freely available in both book form, and the lectures [0] makes it one of the best resources out there. (I just never found it's section on compilers that helpful. Again, probably learning styles.) [0] https://www.youtube.com/playlist?list=PLB63C06FAF154F047

Maybe I need to reread LISP, then. I agree that SICP is laser-focused on Scheme, and thus leaves out general compilation techniques.

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#43

Post by i_don_t_know »

lboasso wrote: post #20
giancarlostoro wrote: post #12

Earlier quoted context omitted.

I really want to be able to wrap my head around compilers but I'm not sure what a good "user friendly" resource is that wont stuff lots of fancy compiler buzzwords in my face and become white noise to me.

I would recommend "Compiler Construction" by Niklaus Wirth. This small book ( The book is freely available here: http://people.inf.ethz.ch/wirth/CompilerConstruction/index.h... For a more comprehensive example, you can have a look at my compiler for the full Oberon language targeting the JVM. It is based on the principles of that book: https://github.com/lboasso/oberonc

I second the recommendation. Wirth's compiler book is a wonderful, concise and very readable introduction.

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#44

Post by MaxBarraclough »

rdc12 wrote: post #39
defined wrote: post #38

This is a very interesting and well-written article, but I am puzzled by the technique being described as "JIT compilation". Isn't JIT compilation the process of compiling some sort of intermediate code into machine code at runtime ? For example, compiling JVM bytecode into machine code during execution, and using the cached machine code transparently on subsequent calls to that segment of code. Not to detract from t…

For me the key concept that comes to mind for a JIT compiler, is profiling at runtime to decide what to compile and how aggressively which this article doesn't deal with. The other big thing is how to make the profiling and compilation not be too expensive to negate itself. The article author seems to be using GCC not at runtime, but more as a generator for templates (at the machine code level) for the rest of the so…

Mixed-mode execution (interpreter+JIT when justified) isn't the only way to go.

.Net doesn't work that way, for instance. Unlike Java's HotSpot, .Net never interprets CIL instructions. They're always compiled down to native code for execution.

At least, that was true in 2009, according to Jon Skeet - https://stackoverflow.com/a/1255832/2307853

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#45

Post by fasquoika »

one_thawt wrote: post #13

In my opinion, the most tedious part of writing this sort of thing from scratch is the x86/x86-64 instruction encoding. Grab the Intel/AMD manuals or an opcode database and have fun. Other projects to study: LLVM, dynasm, asmjit, peachpy, luajit jit, Intel XED, etc.

>dynasm, ... luajit jit

Aren't these the same thing?

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#46

Post by boomlinde »

ATsch wrote: post #31
ben-schaaf wrote: post #28

Earlier quoted context omitted.

Unbinding a name does do one thing which might call `__del__`: It decreases the reference count by 1.

In CPython, yes. But this is an implementation detail.

I learned this with PyPy (or maybe it was Jython?). My usual bad habit for small scripts to depend on open files to close when they went out of scope would suddenly exhausted all file descriptors. Ouch.

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#47

Post by defined »

MaxBarraclough wrote: post #44
rdc12 wrote: post #39

Earlier quoted context omitted.

For me the key concept that comes to mind for a JIT compiler, is profiling at runtime to decide what to compile and how aggressively which this article doesn't deal with. The other big thing is how to make the profiling and compilation not be too expensive to negate itself. The article author seems to be using GCC not at runtime, but more as a generator for templates (at the machine code level) for the rest of the so…

Mixed-mode execution (interpreter+JIT when justified) isn't the only way to go. .Net doesn't work that way, for instance. Unlike Java's HotSpot, .Net never interprets CIL instructions. They're always compiled down to native code for execution. At least, that was true in 2009, according to Jon Skeet - https://stackoverflow.com/a/1255832/2307853

> .Net never interprets CIL instructions.

I wasn't aware of this. To be clear, it's not 100% accurate to say "never", because it depends on the .NET environment, but point taken.

I now understand that .NET (on desktop, anyway) compiles all intermediate code (CIL) to native code on every run. This is a way different kind of JIT than Java HotSpot, which executes intermediate code (bytecode) initially, profiles intermediate code execution, and only compiles the hot spots to native code after a while.

It seems like what "JIT" means has evolved since the time it meant "JVM HotSpot".

There's a very interesting snippet of an interview with Anders Hejlsberg at https://stackoverflow.com/a/1255828/4158187, which exposes his reasoning.

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#48

Post by defined »

csl wrote: post #41
defined wrote: post #38

This is a very interesting and well-written article, but I am puzzled by the technique being described as "JIT compilation". Isn't JIT compilation the process of compiling some sort of intermediate code into machine code at runtime ? For example, compiling JVM bytecode into machine code during execution, and using the cached machine code transparently on subsequent calls to that segment of code. Not to detract from t…

I do see your point, but I tried to make it clear that this is compilation in an extremely restricted sense: It specializes and executes a function at runtime. And I wanted to cover the core technique of doing just that. Perhaps I should have included a small example of compiling a linear list of abstract instructions into machine code. But that again would consist of compiling small templates in a coherent way, just…

My subsequent comment acknowledges that JIT means more than compilation at runtime. In .NET desktop, it means compilation of all CIL to native on every run, just before execution. To me, that's just executing the second pass of a compiler at launch time, but Anders Hejlsberg says it's JIT[1], and I'm not going to argue :)

[1]: https://stackoverflow.com/a/1255828/4158187

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#49

Post by MaxBarraclough »

defined wrote: post #47
MaxBarraclough wrote: post #44

Earlier quoted context omitted.

Mixed-mode execution (interpreter+JIT when justified) isn't the only way to go. .Net doesn't work that way, for instance. Unlike Java's HotSpot, .Net never interprets CIL instructions. They're always compiled down to native code for execution. At least, that was true in 2009, according to Jon Skeet - https://stackoverflow.com/a/1255832/2307853

> .Net never interprets CIL instructions. I wasn't aware of this. To be clear, it's not 100% accurate to say "never", because it depends on the .NET environment, but point taken. I now understand that .NET (on desktop, anyway) compiles all intermediate code (CIL) to native code on every run . This is a way different kind of JIT than Java HotSpot, which executes intermediate code (bytecode) initially, profiles interme…

There's a rats' nest of terminology here.

".Net" and "Common Language Runtime" refer specifically to Microsoft's implementation. The spec is called the "Common Language Infrastructure".

Wikipedia seems to use "Common Language Runtime" in both ways - the Mono article uses it in the generic sense, whereas the CLR's own article explicitly states that it refers to .Net's implementation.

.Net also has ngen.exe which isn't really JIT at all, it's traditional 'offline' ahead-of-time compilation. In a sense that's rather like the traditional Unix compile-and-install model. One sees the oxymoron "pre-JIT" used to describe this.

This is another difference from HotSpot, which historically never cached native code to disk, though other JVMs have always been able to do this (Excelsior Jet, and GCJ, for instance).

Apparently AOT may be coming to Java though - it may even already be here, I'm not quite sure.

https://www.infoq.com/news/2016/10/AOT-HotSpot-OpenJDK-9 , http://openjdk.java.net/jeps/295

I recall once reading that one of the reasons HotSpot didn't cache native code, was the issue of securing that cache. Not sure what they've done to address that, nor what .Net does; it seems a valid concern.

That's an interesting read re. Hejlsberg. Worth noting that Mono takes them up on that and does feature an interpreter.

Re: Writing a basic x86-64 JIT compiler from scratch in stock Python

#50

Post by MaxBarraclough »

defined wrote: post #47
MaxBarraclough wrote: post #44

Earlier quoted context omitted.

Mixed-mode execution (interpreter+JIT when justified) isn't the only way to go. .Net doesn't work that way, for instance. Unlike Java's HotSpot, .Net never interprets CIL instructions. They're always compiled down to native code for execution. At least, that was true in 2009, according to Jon Skeet - https://stackoverflow.com/a/1255832/2307853

> .Net never interprets CIL instructions. I wasn't aware of this. To be clear, it's not 100% accurate to say "never", because it depends on the .NET environment, but point taken. I now understand that .NET (on desktop, anyway) compiles all intermediate code (CIL) to native code on every run . This is a way different kind of JIT than Java HotSpot, which executes intermediate code (bytecode) initially, profiles interme…

Speak of the devil - https://news.ycombinator.com/item?id=15686875

Edit: I'm wrong about the terminology! https://news.ycombinator.com/item?id=15688290

Post reply on HN