5 ms·

JIT-compilation using GCC 5

206 points by Tsiolkovsky 12y ago

arez 12y ago
so did I got it right? If I want to write a script language I can use this library to implement JIT compiling. Because with compiled languages that are not running on an VM like JVM it is impossible to use JIT compiling, because I don't have the source files.
- duaneb 12y ago
  > Because with compiled languages that are not running on an VM like JVM it is impossible to use JIT compiling, because I don't have the source files. Nonsense, you can jit compile either through the JVM (gen bytecode -> bytecode is jit'd) or through normal ways (llvm, poking memory and protecting it appropriately, etc.)
  - Joky 12y ago
    Did you miss the "not" in "not running on a VM"?
- Joky 12y ago
  It is not completely true. See for instance what PNaCl is doing: you can compile C++ (or any similar compiled language) to (LLVM) byte code, and then JIT it at runtime.
userbinator 12y ago
Bellard's TCC had a similar feature - it could be used to JIT scripts written in C, and libtcc was its library for dynamic code generation. Being much simpler it does nowhere near as much optimisation as GCC, but that also means it can compile much faster: http://bellard.org/tcc/tccboot.html http://bellard.org/tcc/tccboot.html TCCBOOT is only 138 KB big (uncompressed code) and it can compile and run a typical Linux kernel in less than 15 seconds on a 2.4 GHz Pentium 4. It'd be interesting to see someone try a similar experiment with libgccjit.
- AceJohnny2 12y ago
  > a typical Linux kernel in less than 15 seconds on a 2.4 GHz Pentium 4. Hah! In 2004 (last news) maybe, and I wonder which featureset. In fact, it's been a long time since I've compiled the kernel myself [1]. I wonder how long a minimal kernel takes to build nowadays. Of course, I guess "minimal" is even more ambiguous today than it was back then... [1] Back in the day, "recompile your kernel" was the Linux joke answer to problems the way "try turning it off and on again" was for Windows...
  - agumonkey 12y ago
    I wanted to reverse tccboot to port it to recent kernels, unfortunately it's far above my skills (far^far), I wonder if anyone attempted such a thing.
  - teraflop 12y ago
    On an i3 machine with plenty of RAM and an SSD, and GCC 4.9, a recent kernel builds in about 3.5 minutes. That's using the "defconfig" target, which is pretty bare-bones (only ext4, very few device drivers or fancy new features). As another data point, Debian's kernel configuration on the same hardware takes 26 minutes to build.
    - userbinator 12y ago
      One of the other benchmarks linked to on the page shows tcc compiling 10x faster than GCC 4.9 -O0 running on the RPi, and 41x faster than GCC -O3: http://lists.nongnu.org/archive/html/tinycc-devel/2013-02/msg00043.html http://lists.nongnu.org/archive/html/tinycc-devel/2013-02/ms... I think 3.5 minutes down to less than 15 seconds for a recent kernel is probably possible. On the other hand, tcc generates code which runs 3.5x slower, so you'd have to be doing a lot of compiling and not much running for it to save any time. JIT'ing the kernel is a fun exercise, nonetheless.
- azinman2 12y ago
  That's amazing!!!
the_mitsuhiko 12y ago
Given that this adds a blob of GPL code into your application I wonder who the intended target audience is.
- DSMan195276 12y ago
  Well, the blob is only added to the compiler portion of your code. Ex. In this article, only the Python script is under the GPL because of libgccjit. The .bf scripts shouldn't/don't have a requirement to be under the GPL.
  - the_mitsuhiko 12y ago
    Is that true? So you can use the JIT compiler in a non GPL project?
    - ploxiln 12y ago
      You've always been able to use gcc to compile non-gpl code into an executable, and distribute it. You've always been able to distribute a distro that includes gcc and also includes gpl-incompatible programs (like flash player, or apache-licensed programs). What has to be GPL is the application that you combine the gcc code with, either when you link in the new libgccjit, or in your python code that imports the python module that links to libgccjit, or if you modify/extend the code for gcc itself (or other GPL codebases).
      - nitrogen 12y ago
        Can a GPL-licensed interpreter run non-GPL scripts? It seems like one should be able to build a GPL JIT+interpreter as a single application, then run whatever scripts they want on top of it.
        drewcrawford 12y ago
        The GPL applies to "derivative works", which is a legal term without a precise technical definition. If you decide to write an interpreter for an existing language, and run existing programs on it, then those existing programs "probably" do not have to be under the GPL, for definitions of "probably" that Legal will not be happy about. If you write an interpreter for a made-up language and then write programs in that language as part of a thinly-veiled scheme to do an end-run around the GPL, then that "probably" violates the GPL. There is currently no bright-line rule separating those cases; this is something IP lawyers argue about themselves and reach wildly different conclusions.
comex 12y ago
Here are the raw API docs: https://gcc.gnu.org/onlinedocs/jit/genindex.html https://gcc.gnu.org/onlinedocs/jit/genindex.html It seems much nicer, if more limited, than the old 'everything is type "tree"' GENERIC API: https://gcc.gnu.org/onlinedocs/gccint/GENERIC.html#GENERIC https://gcc.gnu.org/onlinedocs/gccint/GENERIC.html#GENERIC More competitive with LLVM for projects that don't mind the GPL. I'm looking forward to finding a project to try it out on.
Twirrim 12y ago
From the perspective of using a JIT within python that the blog post raises you can see two different approaches of this being done in Python with the LLVM: 1) The pyston project from Dropbox is an attempt to build a python runtime sitting on top of the LLVM JIT https://github.com/dropbox/pyston https://github.com/dropbox/pyston 2) The numba project for python also uses llvm's jit direct from Python, just requiring you to add a decorator: http://numba.pydata.org/ http://numba.pydata.org/ a quick introduction is here: http://numba.pydata.org/numba-doc/0.18.1/user/jit.html http://numba.pydata.org/numba-doc/0.18.1/user/jit.html
fafner 12y ago
I think the GNU Octave folks are using libgccjit to write a JIT for their Matlab-like language.

206 points by Tsiolkovsky 12y ago

arez 12y ago

so did I got it right? If I want to write a script language I can use this library to implement JIT compiling. Because with compiled languages that are not running on an VM like JVM it is impossible to use JIT compiling, because I don't have the source files.

duaneb 12y ago
> Because with compiled languages that are not running on an VM like JVM it is impossible to use JIT compiling, because I don't have the source files. Nonsense, you can jit compile either through the JVM (gen bytecode -> bytecode is jit'd) or through normal ways (llvm, poking memory and protecting it appropriately, etc.)
- Joky 12y ago
  Did you miss the "not" in "not running on a VM"?
Joky 12y ago
It is not completely true. See for instance what PNaCl is doing: you can compile C++ (or any similar compiled language) to (LLVM) byte code, and then JIT it at runtime.

userbinator 12y ago

Bellard's TCC had a similar feature - it could be used to JIT scripts written in C, and libtcc was its library for dynamic code generation. Being much simpler it does nowhere near as much optimisation as GCC, but that also means it can compile much faster: http://bellard.org/tcc/tccboot.html http://bellard.org/tcc/tccboot.html TCCBOOT is only 138 KB big (uncompressed code) and it can compile and run a typical Linux kernel in less than 15 seconds on a 2.4 GHz Pentium 4. It'd be interesting to see someone try a similar experiment with libgccjit.

AceJohnny2 12y ago
> a typical Linux kernel in less than 15 seconds on a 2.4 GHz Pentium 4. Hah! In 2004 (last news) maybe, and I wonder which featureset. In fact, it's been a long time since I've compiled the kernel myself [1]. I wonder how long a minimal kernel takes to build nowadays. Of course, I guess "minimal" is even more ambiguous today than it was back then... [1] Back in the day, "recompile your kernel" was the Linux joke answer to problems the way "try turning it off and on again" was for Windows...
- agumonkey 12y ago
  I wanted to reverse tccboot to port it to recent kernels, unfortunately it's far above my skills (far^far), I wonder if anyone attempted such a thing.
- teraflop 12y ago
  On an i3 machine with plenty of RAM and an SSD, and GCC 4.9, a recent kernel builds in about 3.5 minutes. That's using the "defconfig" target, which is pretty bare-bones (only ext4, very few device drivers or fancy new features). As another data point, Debian's kernel configuration on the same hardware takes 26 minutes to build.
  - userbinator 12y ago
    One of the other benchmarks linked to on the page shows tcc compiling 10x faster than GCC 4.9 -O0 running on the RPi, and 41x faster than GCC -O3: http://lists.nongnu.org/archive/html/tinycc-devel/2013-02/msg00043.html http://lists.nongnu.org/archive/html/tinycc-devel/2013-02/ms... I think 3.5 minutes down to less than 15 seconds for a recent kernel is probably possible. On the other hand, tcc generates code which runs 3.5x slower, so you'd have to be doing a lot of compiling and not much running for it to save any time. JIT'ing the kernel is a fun exercise, nonetheless.
azinman2 12y ago
That's amazing!!!

the_mitsuhiko 12y ago

Given that this adds a blob of GPL code into your application I wonder who the intended target audience is.

DSMan195276 12y ago
Well, the blob is only added to the compiler portion of your code. Ex. In this article, only the Python script is under the GPL because of libgccjit. The .bf scripts shouldn't/don't have a requirement to be under the GPL.
- the_mitsuhiko 12y ago
  Is that true? So you can use the JIT compiler in a non GPL project?
  - ploxiln 12y ago
    You've always been able to use gcc to compile non-gpl code into an executable, and distribute it. You've always been able to distribute a distro that includes gcc and also includes gpl-incompatible programs (like flash player, or apache-licensed programs). What has to be GPL is the application that you combine the gcc code with, either when you link in the new libgccjit, or in your python code that imports the python module that links to libgccjit, or if you modify/extend the code for gcc itself (or other GPL codebases).
    - nitrogen 12y ago
      Can a GPL-licensed interpreter run non-GPL scripts? It seems like one should be able to build a GPL JIT+interpreter as a single application, then run whatever scripts they want on top of it.
      - drewcrawford 12y ago
        The GPL applies to "derivative works", which is a legal term without a precise technical definition. If you decide to write an interpreter for an existing language, and run existing programs on it, then those existing programs "probably" do not have to be under the GPL, for definitions of "probably" that Legal will not be happy about. If you write an interpreter for a made-up language and then write programs in that language as part of a thinly-veiled scheme to do an end-run around the GPL, then that "probably" violates the GPL. There is currently no bright-line rule separating those cases; this is something IP lawyers argue about themselves and reach wildly different conclusions.

comex 12y ago

Here are the raw API docs: https://gcc.gnu.org/onlinedocs/jit/genindex.html https://gcc.gnu.org/onlinedocs/jit/genindex.html It seems much nicer, if more limited, than the old 'everything is type "tree"' GENERIC API: https://gcc.gnu.org/onlinedocs/gccint/GENERIC.html#GENERIC https://gcc.gnu.org/onlinedocs/gccint/GENERIC.html#GENERIC More competitive with LLVM for projects that don't mind the GPL. I'm looking forward to finding a project to try it out on.

Twirrim 12y ago

From the perspective of using a JIT within python that the blog post raises you can see two different approaches of this being done in Python with the LLVM: 1) The pyston project from Dropbox is an attempt to build a python runtime sitting on top of the LLVM JIT https://github.com/dropbox/pyston https://github.com/dropbox/pyston 2) The numba project for python also uses llvm's jit direct from Python, just requiring you to add a decorator: http://numba.pydata.org/ http://numba.pydata.org/ a quick introduction is here: http://numba.pydata.org/numba-doc/0.18.1/user/jit.html http://numba.pydata.org/numba-doc/0.18.1/user/jit.html

fafner 12y ago

I think the GNU Octave folks are using libgccjit to write a JIT for their Matlab-like language.