Allow thunk closures to be evaluated

[mpickering]

mpickering> bgamari: Should it be possible to evaluate a thunk using ghc-debug?
5:13 PM <bgamari> Nope
5:13 PM <mpickering> I mean if you unpause
5:13 PM <bgamari> there's no safe way to do this
5:13 PM <mpickering> and repause
5:13 PM <bgamari> ahh
5:13 PM <bgamari> hmm
5:13 PM <bgamari> a good question
5:13 PM <bgamari> I suppose you could spawn a thread to run `evaluate x >> pause`
5:13 PM <bgamari> that would be pretty neat
5:13 PM <mpickering> it seems like it should be possible, and you can implement a "force" request which increases the "pause window" by one and returns the address of the newly evaluated thing
5:14 PM <mpickering> How does the ghci forcing command work?
5:14 PM <bgamari> I would just implement it with an MVar
5:14 PM <mpickering> I want to integrate my rewrite of ghc-vis into ghc-debug somehow you see
5:15 PM <bgamari> forceAndPause :: MVar a -> a -> IO ()
5:15 PM <bgamari> provided in the stub's haskell implementation
5:15 PM <mpickering> Well I'm imagining that you have a GenClosure which happens to be a thunk, and the address of that closure
5:15 PM <mpickering> then you send a message to the RTS to unpause, evaluate the closure, pause again
5:15 PM <bgamari> right
5:16 PM <bgamari> so the stub C implementation would provide a message
5:16 PM <bgamari> which takes a ClosurePtr
5:16 PM <mpickering> because ghc-vis supports this where you click on a thunk and it evaluates
5:16 PM <bgamari> calls forceAndPause on it
5:16 PM <bgamari> and then returns the result
5:16 PM <mpickering> so `forceAndPause` should take a `ClosurePtr` as an argument rather than `a`?
5:17 PM <bgamari> the Message payload should be a ClosurePtr
5:18 PM <bgamari> forceAndPause provided by the Haskell bit should be polymorphic
5:18 PM <bgamari> it would be implemented as
5:19 PM → mdunnio joined ([email protected])
5:19 PM <bgamari> forceAndPause resultVar x = do handleAny (pure . Left) (evaluate x) >>= putMVar resultVar
5:19 PM <bgamari> forceAndPause resultVar x = do handleAny (pure . Left) (fmap Right $ evaluate x) >>= putMVar resultVar
5:19 PM <mpickering> Another thing to think about is that you don't want to actually unpause? because then other computation will happen as well
5:19 PM <bgamari> forceAndPause :: MVar (Either SomeException a) -> a -> IO ()
5:20 PM <bgamari> mpickering, I don't think there is much you can do about that
5:20 PM <bgamari> mpickering, the RTS doesn't really provide any control over scheduling
5:20 PM <mpickering> ok
5:20 PM <bgamari> and indeed it can't
5:20 PM <bgamari> since what if the thunk was unsafePerformIO $ takeMVar someVar
5:20 PM <bgamari> if you forbid any other threads to run then you would deadlock
5:21 PM <bgamari> but the GC would have no way of knowing since someVar is still alive
5:21 PM <mpickering> I suppose it's fine if you pause at a point where the main thread is idle, say, pause >> threadDelay 1000000000
5:21 PM <bgamari> right
5:21 PM <bgamari> mpickering, this is how gdb works as well
5:21 PM <bgamari> if you `print f(x)`
5:22 PM <bgamari> all OS threads other than the one you are looking at will run for a bit while the evaluation occursa
5:22 PM <bgamari> due to limitations in how ptrace works IIRC
5:22 PM <mpickering> ok
5:22 PM <mpickering> seems possible to implement this then
5:22 PM <bgamari> indeed
5:23 PM <mpickering> if you can do this you can answer questions about space usage before/after forcing
5:23 PM <bgamari> sounds like a neat feature
5:23 PM <bgamari> true
5:23 PM <mpickering> which sounds like something you would have found useful recentl
5:23 PM <bgamari> you might also expose a command to force a major GC
5:24 PM <mpickering> as a request?
5:24 PM <mpickering> right now I just do a major gc before pausing
5:24 PM → dennisb joined ([email protected])
5:27 PM <isovector1> i'm trying to solve a stuck tyfam. emitting a wanted `CNonCanonical` of the form `MyTyFam a b c ~ result` works on its own, but won't reduce other families that are stuck on the result of `MyTyFam a b c`
5:31 PM ⇐ mdunnio quit  ↔ luke-jr nipped out  
5:38 PM <bgamari> mpickering, that also works