[IronPython] What are plans concerning pre-runtime compilation?
jvm_cop at spamcop.net
Tue Aug 22 04:51:54 CEST 2006
In CPython, the runtime will look for a .pyc file corresponding to any .py file it's asked to access, and its contents are used if the .pyc file's timestamp is later than that of the .py file. The .pyc file contents aren't native code, but in theory things are (at least a little bit) faster when the .pyc exists as the syntax analysis (and some other work) has already been done.
As I understand IP's internals, the process of compiling is explicitly in two steps -- syntax analysis (resulting in an AST being built) and IL code generation. Code generation is sometimes (often?) delayed until a method is actually called, with the dual advantages of (a) being able to generate code that can optimize appropriately for the actual parameters used, and (b) avoiding the work of generating code for methods that are never called.
Are there any plans to do something so that frequently-used (imported) code can be at least syntax-checked once, even if code isn't to be generated until runtime? A mechanism that produced a file holding a representation of the AST would be the equivalent of the .pyc file mechanism; would an optimization like that be useful for IP, and if so is it planned?
I can see producing an assembly from the AST file, where calls to every method would result in the code generation step being executed for that method. (That perhaps could be the default for the format of the file holding the AST, with the AST a resource in the assembly.) An additional step could invoke the code generator for particular methods with particular argument type specs and build an assembly that already has type-specific IL implementations of some methods.
I'm thinking that a "plain AST" assembly would have the equivalent of the C# method
object MethodName(params object args)
for each method. The IL for that method would call a routine in the runtime to generate code if it hadn't been generated before, then call it. With work on the part of the developer (possibly directed by the contents of a file generated by tracing the execution of the code), there would be other methods with specific relevant signatures, and type-specific code generated for those.
Is this in line with what's being discussed? Where is the discussion taking place?
Thanks for listening.
J. Merrill / Analytical Software Corp
More information about the Ironpython-users