#14490: TTG Snags -------------------------------------+------------------------------------- Reporter: alanz | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.3 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: #14482 Differential Rev(s): | Wiki Page: | ImplementingTreesThatGrow -------------------------------------+------------------------------------- == Context ImplementingTreesThatGrow The intention is to get the first-pass implementation of Trees that Grow into GHC 8.4.1. So far the following commits are on master - https://phabricator.haskell.org/rGHC438dd1cbba13d35f3452b4dcef3f94ce9a216905 - https://phabricator.haskell.org/rGHCe3ec2e7ae94524ebd111963faf34b84d942265b4 - https://phabricator.haskell.org/rGHC47ad6578ea460999b53eb4293c3a3b3017a56d65 Unfortunately we have picked up compile-time performance regressions in compiling GHC itself, with these patches. The perf.haskell.org results are - https://perf.haskell.org/ghc/#revision/438dd1cbba13d35f3452b4dcef3f94ce9a216... - https://perf.haskell.org/ghc/#revision/e3ec2e7ae94524ebd111963faf34b84d94226... - https://perf.haskell.org/ghc/#revision/47ad6578ea460999b53eb4293c3a3b3017a56... Summary: each of the three patches caused GHC validate time to worsen by another 5% or so, cumulatively pushing compile time from 2011 to 2400 seconds. It seems the major culprit is deriving the `Data` instances, probably related to the complex constraint sets required to ensure that all the extension points have `Data` instances too. Some discussion around this is captured in ImplementingTreesThatGrow/Instances == Things tried so far PLAN E (ImplementingTreesThatGrow/Instances#PLANE), in https://github.com/ghc/ghc/tree/wip/ttg5-data-one-file-2017-11-17. This puts all the `Data` derivations for the hsSyn AST into a single file, `hsSyn/HsInstances.hs`. This causes compile-time memory usage to spike up to around 7GB when compiling this file. Gipedia reports a compilation time of 2468 secs for this version, which perhaps indicates that the instance derivation is not the main culprit. The module itself only takes in the order of 45secs to compile, on `home .smart-cactus.org` (http://lpaste.net/7779809210564345856) Splitting the `HsInstances.hs` file in 2, as per https://github.com/ghc/ghc/tree/wip/ttg5-data-2017-11-17 reduces the memory requirement for each, but shows up #14482, where there is a linker failure due to incorrectly processing boot files. == Next Steps I guess the first thing to do is to try to isolate what precisely is causing the slowdown. It may not be instance generation at all. Also, it is probably better to revert the three patches from master, when cutting the 8.4 initial freeze. Open to suggestions ... -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14490 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler