#13629: sqrt should use machine instruction on x86_64 -------------------------------------+------------------------------------- Reporter: bgamari | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: 8.4.1 Component: Compiler (NCG) | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: Runtime | Test Case: performance bug | numeric/num009 Blocked By: | Blocking: Related Tickets: #13570 | Differential Rev(s): Phab:D3508 Wiki Page: | -------------------------------------+------------------------------------- Comment (by kavon): The verbose lowering of sqrt that was there before this patch must have stumbled upon some sort of lucky interaction in the processor (arithmetic vs memory unit ports) in that particular hot patch of code in n-body, which is not something that can be relied on in the grand scheme of things. What we ''can'' tell is that this patch produces better code, more closely matching what LLVM produces as well. bgamari has also kindly quantified how much better it is. Thus, I think this ~4% drop in n-body is not something to worry about, as it seems even non-functional changes can knock a few nofib runtime numbers by that amount [1]. Ticket #13570 is also a separate issue: I noticed the poor sqrt lowering when I was investigating that ticket. [1] https://perf.haskell.org/ghc/#revision/945c45ad50ed31e3acb96fdaafb21640c4669... -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/13629#comment:20 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler