I finally resurrected this work I started back in September. I think I've included most of the old feedback, and done a bunch more cleaning.
This series changes ir_call from a value (which can be nested inside expression trees) to a statement which can only appear in the instruction stream. This makes expression trees pure (free from side-effects), which ought to make optimization passes easier to write and less error-prone. It also resolves the ugliness where ir_call was the _only_ ir_rvalue that could appear in an instruction stream, which was bizarre. It also includes some AST->IR clean-ups, because...well. It was a mess. The one outstanding concern I have is IR->Mesa and IR->TGSI: both seem to support function calls, but I'm not sure how that all works. IMHO Mesa IR really shouldn't support them---we should just rely on the GLSL compiler's inliner. I'm incredibly skeptical that the code in IR->Mesa actually works, since it has a comment that functions other than main() should never been seen. Maybe that's just old. Can we delete this? I suspect IR->TGSI needs to support real function calls. The code there still compiles, but it probably breaks. I probably need help adapting it. No piglit regressions on i965. Nothing else was tested. _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
