Terrific. Would you like to do a nofib run before-and-after, and dump
#15103: Speed optimizations for elimCommonBlocks -------------------------------------+------------------------------------- Reporter: AndreasK | Owner: AndreasK Type: task | Status: patch Priority: normal | Milestone: 8.6.1 Component: Compiler | Version: 8.5 Resolution: | Keywords: CodeGen Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Phab:D4597 Wiki Page: | -------------------------------------+------------------------------------- Comment (by AndreasK): Replying to [comment:9 simonpj]: the `nofib-analyse` summary in here? It's very helpful just to record the benefits of particular changes, for posterity. {{{ -------------------------------------------------------------------------------- Program Size Allocs Runtime Elapsed TotalMem -------------------------------------------------------------------------------- maillist 0.0% 0.0% 0.024 0.025 +6.5% mate -0.0% 0.0% +4.6% +4.6% 0.0% cryptarithm1 0.0% 0.0% -1.7% -1.8% 0.0% -------------------------------------------------------------------------------- Min -0.0% -0.0% -1.7% -1.8% 0.0% Max +0.0% +0.1% +4.6% +4.6% +6.5% Geometric Mean -0.0% +0.0% +0.3% +0.3% +0.1% Compile Times ------------------------------------------------------------------------------- Program logNoOpt logWithOpt ------------------------------------------------------------------------------- -1 s.d. ----- -1.9% +1 s.d. ----- +1.5% Average ----- -0.2% Compile Allocations ------------------------------------------------------------------------------- Program logNoOpt logWithOpt ------------------------------------------------------------------------------- -1 s.d. ----- -0.3% +1 s.d. ----- -0.1% Average ----- -0.2% }}} These runtime results took me by surprise! It turns out code layout depends on uniques. And the final uniques change with this patch. So we get slightly different codelayout. I verified that the code for mate is otherwise the same. Using different starting or increment values for uniques also lead to different results with up to the same performance spread. So I don't think the runtime change is representative. Compiler performance impact is as expected. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/15103#comment:10 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler