Latest commit f094a3c
May 28, 2011
|Failed to load latest commit information.|
=============================================================================== SqueakBootstraper by Yoshiki Ohshima =============================================================================== SqueakBootstrapper is the first step to create a Squeak image from "text files." It is an MIT-licensed project. The idea consists of a few parts: - A very small image (~120k), derived from MicroSqueak by John Maloney with minimum capability of loading pre-compiled class definitions and methods is serialized to an array embedded in a C file. - The BootstrapCompiler, a Squeak Bytecode Compiler written in C (and leg originally by Ian Piumarata), which can generate the pre-compiled files from regular Squeak Smalltalk file outs. - A "reference implementation" of above compiler, written in OMeta2/Squeak. This produces the bit-identical binary file with the above compiler. - A mildly shrunk-down but otherwise regular Squeak Compiler. - A host Squeak image that can create the MicroSqueak image above and development. The bootstrap process goes like this: 0. The MicroSqueak image is created from the development image, and the result is rendered to a C file that contains the content of entire file as an "unsigned char" array. We pretend that this file is given. This image is equipped with "BootStrapReader", which can read an external file format called ".sto" from file. 1. The BootstrapCompiler in C reads the definition of the Squeak Compiler written in Squeak, and generates the .sto (called "compiler.sto") file. 2. The MicroSqueak reads compiler.sto and save itself. Now, this image can further read more code and compile them. 3. As the first step, compile the same Squeak Compiler code by itself. 4. It now can read more files, including doits. 5. Theoretically, it can go up to the full image. Currently, only 0. through 3 are working. I/ Installation --------------- First install my version of "greg" from https://github.com/yoshikiohshima/greg . This version of greg (which is a minor variation of leg) is equipped with the memoization feature, which is essential to parse non-trivial amount of code. Place the greg directory by this repository. Type "make test" in this repository. The "image.c" is generated from the MicroSqueak image but put in the repo. The command created from image.c creates a .image file, the compiler source code is turned to .sto file and it is loaded into the image. II/ What is it good for? ------------------------- - By separating the compiler implementation and class hierarchy, people can do a deep change to Squeak without shooting their own foot. - A long-standing problem of not being able to create the image from text files may be now solvable. - A clean starting point that can grow may give you a fresh opportunity to redesign the entire system. - A C-based compiler is blazingly fast. Compiling the 150k of code takes less than half-second. Loading the .sto file is also quite fast. This format alone may have some other use (i.e., a lightweight persistent/code migration). III/ Limitations ----------------- - The Bootstrap Compiler and the reference implementation compiler generate a less efficient bytecode sequence than the regular compiler. They do not produce storeintopop instructions (other than the initialization of block temps). They also have a limitation on the number of literal variables in a method. (This is why the initialize methods of VariableNode and ParseNode are split into a few.) The generated methods always use big contexts. - The compilers are not closure-aware. - The class builder in the image does not support class mutation - only the addition of new classes. - The entry point #start method is intended to be "clean" over image saving; i.e., when the image is saved, it schedules the new process that should be launched freshly in the new session but the process that saved the image is supposed to cleanly go away. In other words, if you keep loading the same methods and save-and-quit the image over and over again, the image size should stay exactly the same. But there is a context object or two accumulating every cycle. - The development image is not a standard version. It is 3.8-derived, with some changes from trunk. Also, the dictionary that is used by the other parts of the dev-image and MicroSqueak hierarchy are different; this tends to cause troubles. III/ Development ----------------- You (still) need the development image: http://tinlizzie.org/~ohshima/SqueakBootstrapper.zip With this, you would add more classes to the MObject hierarchy and try to test things there. There is MTestCase class so it is a good idea to write some tests, as the ability of interactive debugging is often limited. You can generate a new MicroSqueak image with your change, but also file out your code to .st and create a .sto file, and load it by giving the file name to the command line argument of Squeak invocation. (The third argument of the command line specifies whether the image should be saved at the end or not.) III/ Plans ---------------------- - Adapt closure compiler. - Add more features for supporting self-development. - Clean up the MicroSqueak hiearachy again to make the image smaller and cleaner. - Do something with the big C array idea. We could have a even smaller image that can load binary (it could be somewhat like Spoon), or we could capture the dynamic behavior of MicroSqueakImageBuilder and store the sequence of image writing commands in, say, C. Then we could "replay" the commands to create the image file.