738300 - Experiment: make wordcode the *only* interpreter

Status: RESOLVED WONTFIX

(Tamarin Graveyard :: Interpreter, defect)

Description

[Edwin Smith]

There are a number of potential advantages to having just one interpreter configuration (wordcode, or whatever variant it becomes)
* less builds to test
* less hair in the source code
* always translating makes evolving abc simpler
* always translating makes evolving interpreter code format simpler
* always translating simplifies annotating code with xlation time data

Comment 1

[Edwin Smith]

Attachment: WIP: have only one interpreter

This is obviously a hacky work-in-progress.

Prior to march 2012, this patch was impeded by the JIT and wordcode interpreter's different use of MethodEnv::lookupCache.  However, finddef was redesigned and the lookup cache is gone, so wordcode and jitcode can coexist.

Evicting wordcode that goes cold should be easy - free it and reset trampolines to recompile if necessary.

Comment 2

[Edwin Smith]

Attachment: WIP - make wordcode interpreter and JIT play nice together

Patch to try to make wordcode + JIT play nice together

* enable MS::_abc_body_pos in all configs
* remove AVMFEATURE_COMPILATION_POLICY restrictions
* remove at-most-one JIT + WORDCODE_INTERP
* remove JIT requires ABC_INTERP

problems found so far:
* OSR's apis are arranged around uint8_t* abc_pc, but wordcode typedefs
the code pointer as uintptr_t*.  Either the true abc_pc needs to be a
canonical "position" value everywhere (probably) or OSR should be fixed up
to work with wordcode pointers (not likely since CodegenLIR generating code from the original abc, not wordcode).