#16039: 'GHC.Magic.noinline <var>' should not float out -------------------------------------+------------------------------------- Reporter: heisenbug | Owner: heisenbug Type: bug | Status: new Priority: normal | Milestone: 8.10.1 Component: Compiler | Version: 8.2.1 Resolution: | Keywords: FloatOut Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: #15155 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by heisenbug): Replying to [comment:10 simonpj]:
`exprIsCheapX` helped a lot
What does that mean? How did it help?
There seem to be two distinct mechanisms to float out expressions to toplevel. After fixing `notWorthFloating` and `exprIsTrivial` (which took care of the float-out pass) still a bunch of expressions (e.g `noinline @ (Role -> Type -> MCoercionN -> Coercion) mkGReflCo`) that are destined for class dictionaries got floated out for no reason. After fixing `exprIsCheapX` this is not happening any more.
Rather than doing this deep pattern matching during `exprIsCheapX`
(which will happen repeatedly in an `App`-chain), I think it might be nicer (more efficient) to accumulate a list of (value) arguments instead of the `n` argument to `go`.
Check out exprIsBig too.
Maybe. But concentrate on the places where we know it makes a
difference. Sounds like a plan. I'll wrap up the changes that I have so far and work on more testcases. I think that it'll be okay to land this for 8.10. If you want to look at my changes, you'll find them in https://gitlab.haskell.org/ghc/ghc/commits/wip/T16039 . Then I'll scrutinise the code for similar problems in a low-prio follow-up ticket. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/16039#comment:11 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler