I'm about to merge into the HEAD a branch that Michael Adams has been working
on, consisting of about 50 or so patches affecting mostly files in compiler/cmm.
I've tested a complete bootstrap and testsuite, and as far as I can tell, the
changes should have absolutely no effect on anything at all.
This is the first stage in a refactoring of the backend. Some details are here:
http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/CPS
The main idea is that we introduce a higher-level version of Cmm that includes
full function calls and parameter passing, and no explicit stack manipulation.
This language is generated by the Stg->Cmm code generator, and transformed into
Cmm with explicit stack manipulation by a separate CPS-like phase. This has
some nice advantages:
- it completely separates stack-manipulation from the code generator
- it means we can make use of function calls and parameter passing in
the RTS .cmm code, which simplifies lots of things. Our special
primop calling conventions go away, for example. PrimopWrappers
goes away. The RET_DYN and RET_FUN frames go away, and stack
walking gets much simpler. PAPs will also have a simpler layout,
although this isn't a forced change.
- ultimately, we'll get better code this way. When we drop the via-C
route, we'll be able to use more registers for passing parameters.
So far, the existing code generation route is unaffected, but the changes have
been made to the Cmm datatype and some of the code to do the CPS pass is there.
We wanted to get the branch merged in now to avoid future conflicts.
Comments welcome!
Cheers,
Simon
_______________________________________________
Cvs-ghc mailing list
[email protected]
http://www.haskell.org/mailman/listinfo/cvs-ghc
