R O A D M A P Version numbers are tentative, and I have overestimated the time each stage will take. Please provide me a similar roadmap if you are doing any work that should be merged into this stream. Version Time Description ------- ---- ----------- 110.12 done 1/13 Included: new array representation, literal lifting, and inplace code object updating. 110.13 done 1/18 MLRISC integration: Introduces 'mega optimizations' into the MLRISC back end. This is mostly Allen's work. The idea is to have these optimizations loaded at top level with an internal hook bound to the new optimizer. Thus these modules will not be part of the base heap image. 110.14 done 2/22 PPC port: Most of it is there already but needs debugging. 110.15 done 3/17 GC API: The interface to GC will proceed through a fixed set of registers. This means that the code generation must bundle up anything extra in a record and unbundle it on return. This will simplify, enormously, the interface to GC and will also make fiddling around with register assignments trivial in comparison to what one has to do now. That is to say, changes in register assignment may need little or no runtime system changes. A performance improvement is?? expected in returning from C calls, and going into and out of the garbage collector. It turns out that this is essential for the new x86 code generator. 110.16 3 weeks x86 code generator: Depends On: 110.13, 110.15 A robust MLRISC x86 code generator exists already, which produces substantially better floating point performance and comparable integer performance. 110.16 is intended to get this version with reasonable compile time performance. I am (almost) certain I understand how to get good integer performance. The missing piece of the puzzle is keeping things in registers, but all the techniques I am aware of are heavy weight. Good integer performance must (at least for now) be a loadable 'mega-optimization'. 110.17 4 weeks New runtime: Depends On: 110.15 110.xx - Polling Support: MacOS (in particular) requires polling support to handle signals. 110.xx - FLINT optimizer: Since an earlier attempt at integrating the FLINT optimizer at Bell Labs (v 110.11) resulted in an unstable version, it is assumed that this integration is best done at Yale. The FLINT optimizer can in principle be incorporated at any point in this road map. The earlier the better. Additional (not scheduled yet): - MIPS code generator - representation analysis for new arrays -