#14672: Make likelyhood of branches/conditions available throughout the compiler. -------------------------------------+------------------------------------- Reporter: AndreasK | Owner: (none) Type: task | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.2.2 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Phab:D4316 Wiki Page: | Phab:D4324 Phab:D4327 -------------------------------------+------------------------------------- Comment (by svenpanne): Just a general remark: [https://phabricator.haskell.org/D4327 D4327] talks about annotations as a source of likelihood values, which is fine, but having actual data from previous runs, i.e. using profile-guided optimization, would probably have much more potential. Humans are notoriously bad at guessing what actually eats performace, so manual annotations can only get you so far. As an example: The performance of the CPython interpreter itself increases by roughly 10% if it is compiled with PGO and LTO (link-time optimization), see e.g. * https://bugs.python.org/issue24915 * https://www.activestate.com/blog/2014/06/python-performance-boost- using-profile-guided-optimization Of course the effect of PGO heavily depends on the actual program and the programming language, but I think it is something which should be considered in the long run. Perhaps our LLVM backend already has some (relatively) easy way to make profile data available to the LLVM pipeline, but I might be wrong here. https://msdn.microsoft.com/en-us/library/e7k32f4k.aspx lists a few PGOs, and some are definitely worthwhile for GHC, e.g. using profile data in the register allocator to decide which registers to spill etc. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14672#comment:8 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler