#14137: Do more inlining into non-recursive join points -------------------------------------+------------------------------------- Reporter: simonpj | Owner: nomeata Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.2.1 Resolution: | Keywords: JoinPoints 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 nomeata): Replying to [comment:11 nomeata]:
Now lvl6 will be inlined at its single call site.
Indeed the change to `preInlineUnconditionally` has the prognosed effect. I passed it to `perf.haskell.org` for performance evaluation.
Two changes observed: Runtime of `scs` improved by 3% (could be noise). But compiler performance regressed slightly (around +1% allocations in a few cases). Biggest loser is the parsing compiler performance benchmark `tests/alloc/parsing001`, 3% increase of allocations. I am not inclined to hunt down that regression (I’d require building GHC twice with identical settings and staring at lots of ticky profiles, and then staring at core code), I hope that’s ok. My completely unfounded gut feeling is that a join-point is inlined into a recursive function, and then floated out again in a form that is no longer a join point. Given these results, do you want me to write a Note and push it to master, or not do that just yet and maybe first follow the breadcrumb “we'd like the FloatOut pass not to float out tail-calls from a joinrec”? (But see above why it seems to me that floating out tail-calls from a joinrec is very desirable…) -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14137#comment:13 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler