On 28 December 2010 19:22, Eliot Miranda eliot.miranda@gmail.com wrote:

Hi Yoshiki,

On Tue, Dec 28, 2010 at 2:19 AM, Yoshiki Ohshima yoshiki@vpri.org wrote:

...

Hi,

We've been playing with John's MicroSqueak and it occured to me that having a bytecode compiler that is implemented outside of Squeak opens some possibilities, such as generate a growable image file from all text files, or make deep changes to the system without shooting yourself.

I wrote a longer explanation so if you are interested, please go to:

https://github.com/yoshikiohshima/SqueakBootstrapper

and check it out.

I simply don't see the benefit of putting energy into other languages.  I see the benefit of a textual bootstrap.  But why is it worth-while implementing that in C instead of Smalltalk?  If Smalltalk is more productive (which it is) then writing such a bootstrap in C is a waste of effort, reinvents several wheels at considerable expense and produces an artifact that is less flexible, less extensible, less useful than implementing the same functionality in Smalltalk. On the other hand, as Andreas suggests, trying to implement something using the simulator looks to be really powerful.  Recent;y I've been playing tangentally in this area.  In recent days I've produced a new code generator for Cog that has some useful speedups (Compiler recompileAll ~ 9% faster, benchFib 2x).  To test the code generator I needed to check stack depths at various points in JIT compilation and execution of the JITted code.  I have a Smalltalk class StackDepthFinder that answers the stack depths for each bytecode of a method.  By adding two classes VMObjectProxy and VMCompiledMethodProxy I could apply StackDepthFinder to methods in the simulator's heap and hence derive stack depths for any method in the simulators image.  To test the JIT it was also convenient to be able to JIT methods in my work image, synthesised test cases etc, not just methods in the simulated image.  Again a facade class allows the simulator to JIT any method in my work image.  This worked well and was easy to implement.  Extending in this direction seems straight-forward. One starts up wit a simulator and an empty heap and bootstraps objects into that heap, using whatever bytecode set and object format one chooses.  One can test the image using the simulator which should be quite fast enough if the image is a small kernel.  All the implementation is useful and adds to the simulator/VMMaker ecosystem.  All the code is Squeak and can reuse substantial parts of the system.  Seems like a win to me.  I think I'll take this approach in implementing the new object format.  It could be a new backend to MicroSqueak.

Likewise, in NativeBoost, to debug a generated native code, i simply instructing assembler to generate int3 instruction at point where i need, and then after accepting a method and running the doit, which using this method, i pop into debugger (like gdb) and can step by step see what it does..

I don't remember being able to do that from any other language, because code-compile-run cycle makes it simply impossible. Ah yes.. there is 'compile and continue' feature for C.. but i never used that after couple failures. And besides often compile & continue takes same time to compile things than simply build everything from scratch.. so at the moment when it done compiling, you can forget what you doing there :)

cheers Eliot

...

Thank you!

-- Yoshiki

At Tue, 28 Dec 2010 10:22:35 -0800, Eliot Miranda wrote:

I simply don't see the benefit of putting energy into other languages. I see the benefit of a textual bootstrap. But why is it worth-while implementing that in C instead of Smalltalk? If Smalltalk is more productive (which it is) then writing such a bootstrap in C is a waste of effort, reinvents several wheels at considerable expense and produces an artifact that is less flexible, less extensible, less useful than implementing the same functionality in Smalltalk.

Like I wrote in another reply, it is more about the a meta-language (in this case, a PEG or PEG-like generator). In the last S3 conference, Ian and Takashi and myself showed that the same S-expression language can be targeted to x86 code and Adobe Bytecode. Similar to this, a good bootstrapping strategy would be to write the major part of compiler in something like PEG and provide different backend to produce different executables. For my Bootstrap Compiler, it is not C as much as Leg; C is convenient and have a workable parser generator, so it was an okay step, I thought. (Arguably, Leg does not support structual matching of trees; One of the phase in the reference implementation is written directly in C, even though the pattern in the code just resembles to the PEG implementation.

On the other hand, as Andreas suggests, trying to implement something using the simulator looks to be really powerful. Recent;y I've been playing tangentally in this area. In recent days I've produced a new code generator for Cog that has some useful speedups (Compiler recompileAll ~ 9% faster, benchFib 2x). To test the code generator I needed to check stack depths at various points in JIT compilation and execution of the JITted code. I have a Smalltalk class StackDepthFinder that answers the stack depths for each bytecode of a method. By adding two classes VMObjectProxy and VMCompiledMethodProxy I could apply StackDepthFinder to methods in the simulator's heap and hence derive stack depths for any method in the simulators image. To test the JIT it was also convenient to be able to JIT methods in my work image, synthesised test cases etc, not just methods in the simulated image. Again a facade class allows the simulator to JIT any method in my work image. This worked well and was easy to implement. Extending in this direction seems straight-forward.

Ah, okay. Just an analogy, in the way Slang works, the meta-language here happen to be Slang/Smalltalk and different executables are produced with different backend. As you said, however, Slang is not a language implementation but a practical vehicle to get to a place; with a nicer meta-language, the picture could look prettier.

One starts up wit a simulator and an empty heap and bootstraps objects into that heap, using whatever bytecode set and object format one chooses. One can test the image using the simulator which should be quite fast enough if the image is a small kernel. All the implementation is useful and adds to the simulator/VMMaker ecosystem. All the code is Squeak and can reuse substantial parts of the system. Seems like a win to me. I think I'll take this approach in implementing the new object format. It could be a new backend to MicroSqueak. cheers

Eliot

Thank you!