#7258: Compiling DynFlags is jolly slow -------------------------------------+------------------------------------- Reporter: simonpj | Owner: simonpj Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.6.1 Resolution: | Keywords: deriving-perf Operating System: Unknown/Multiple | Architecture: Type of failure: Compile-time | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by tdammers): Progress made so far: - Reduced algorithmic complexity in register spilling ([comment:33]) - Minor improvements in the pretty-based code writer ([comment:34]) - Refactored generated `Read` code (comment [comment:37] / ticket [ticket:14364]) - Observed that `Read` instances generate long chains of monadic binds over a monad that introduces nested closures when binding (`ReadP`, specifically); and that deeply nested closures in general cause the register allocator to repeatedly unpack and re-pack free variables, which leads to quadratic behavior. The fix for this has been split off into [ticket:14461]. Things left to do: - Replace `pretty` with something that doesn't share its nonlinearities ([ticket:14005]) - Once [ticket:14461] is done, it may be worth digging further; right now, the nested-closures nonlinearity makes it difficult to isolate other bottlenecks though. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/7258#comment:92 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler