#14372: CMM contains a bunch of tail-merging opportunities -------------------------------------+------------------------------------- Reporter: heisenbug | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.2.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by heisenbug): Replying to [comment:10 simonpj]:
It's hard to review code without a spec, or even some input/output examples.
And an overview Note to explain the goal and how the impl works would be useful.
Will do. I am in a very early stage :-)
Based on a very quick look, it seems that you hope to branch from one
`CmmProc` to a label in another. That's a huge thing to do, because it messes up all dataflow analyses, which expect to be able to see all the places that jump to a block (e.g. constant propagation). I'm 99% sure that you can't translate it into LLVM. You are correct. We can only branch to labels that we emit. So this probably will be a NCG-only thing (C-backend is off-limits too), unless we find a way to invoke the "outliner" in LLVM specifically. What I have observed that sometimes previously noted labels disappear due to optimisation. So I'll have to move the commoning to the end of the Cmm pipeline.
Before you invest a lot more effort, can we debate the goal? There are
others on ghc-devs who know about this back-end stuff. Sure. First and above all I am trying to learn and understand the inner workings of Cmm. Let me just get this working with bootstrap and we can discuss. I'll also need to gather some statistics. Thanks for the review, Simon! -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14372#comment:11 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler