[GHC] #14003: Allow more worker arguments in SpecConstr
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature | Status: new request | Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Keywords: JoinPoints, | Operating System: Unknown/Multiple Fusion | Architecture: | Type of failure: Runtime Unknown/Multiple | performance bug Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- Starting with GHC 8.2 (rc1 -- head) I noticed that the SpecConstr pass does not always optimize completely with SpecConstr-heavy code. Setting ```-fmax-worker-args=100``` leads to complete specialization again. However, given that code annotated with ```SPEC``` should be optimized until no more ```SPEC``` arguments are alive, shouldn't ```callToNewPats``` in ```compiler/specialise/SpecConstr.hs``` specialize *irrespective* of the size of the worker argument list? Code that actually fails to specialize is fairly large, hence no test case -- though I have some files with core output showing insufficient specialization. (I'd be willing to write a patch for this) -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by bgamari): In general your code will perform pretty poorly with or without worker/wrapper if you have >100 arguments as this is well beyond what we can accommodate in physical registers. Of course, I'm happy to be proven wrong: do you have a concrete example of a program with this many arguments which you think should optimize well? -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:1 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by choenerzs): I'm using a lot of intermediate data constructors for control flow which are all eliminated if SpecConstr runs to completion. The ```100``` for W/W was only to provide a ceiling I'm not likely to bump against. However, this is part of the problem, in that the user of the library now needs to think about which value to give W/W. The function wstream_Strng2_V_1 becomes more than x10 faster if specconstr runs to completion and allocates *a lot* less (in my allocation benchmark it is 48 bytes versus ~ 600 Kilo-bytes . It has arity 18 after full SpecConstr. In GHC 8.0.2 and earlier, I did not have to change the maximal number of W/W arguments for SpecConstr. The core output for v1-slow (W/W = 10) and v1-fast (W/W = 100) is given as attachments. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:2 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Changes (by choenerzs): * Attachment "v1-slow.dump-simpl" added. core output with default W/W 10 -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Changes (by choenerzs): * Attachment "v1-fast.dump-simpl" added. core output with W/W 100 -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by simonpj): I'm confused. * I think `-fmax-worker-args` applies to the demand analyser, not `SpecConstr`. So I'm not sure how it is affecting things. * Are you getting the same code with 8.0.2 and 8.2 after `SpecConstr`? Overall it's going to be hard to help without a test case. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:3 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by choenerzs): SpecConstr checks via ```isWorkerSmallEnough``` in ```callsToNewPats``` if the pattern to be specialised is small enough. This is controlled by ```-fmax-worker-args```. Line 1913 in SpecConstr.hs. And no, 8.0.2 fully specialises irrespective of the number of worker arguments. Specialisation in 8.2 depends on the number of arguments the pattern to be specialised has. I'll try to construct a test case. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:4 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: #11565 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Changes (by bgamari): * related: => #11565 Comment:
The function `wstream_Strng2_V_1` becomes more than x10 faster if specconstr runs to completion and allocates *a lot* less (in my allocation benchmark it is 48 bytes versus ~ 600 Kilo-bytes . It has arity 18 after full SpecConstr.
And no, 8.0.2 fully specialises irrespective of the number of worker arguments. Specialisation in 8.2 depends on the number of arguments the
Where is this `wstream_Strng2_V_1` function that you refer to? It would be helpful to see a concrete example. pattern to be specialised has. This change is due to two causes: * a48de37dcca98e7d477040b0ed298bcd1b3ab303, which fixes #11565 where `-fmax-worker-args` was simply ignored * f93c363fab1ac8ce6f0b474f5967b0b097995827, which extended the reach of `-fmax-worker-args` to include SpecConstr in order to improve compiler performance. Perhaps you want the SpecConstr limit to be independent from the worker/wrapper limit? -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:5 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: #11565 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by choenerzs): The core output of the ```wstream_Strng2_V_1``` function is in the two attached files. I still need to prepare a test case that doesn't depend on a lot of dependencies. What I want is SpecConstr to continue specialising functions that have a ```SPEC``` argument irrespective of the number of arguments. That is, those where sc_force is set, I think? -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:6 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: #11565 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by simonpj): That seems plausible. We could not call `isWorkerSmallEnough` if `sc_force` is true. Would you like to try that? If so, we should update this comnent {{{ ForceSpecConstr arguments are spotted in scExpr' and scTopBinds which then set sc_force to True when calling specLoop. This flag does four things: * Ignore specConstrThreshold, to specialise functions of arbitrary size (see scTopBind) * Ignore specConstrCount, to make arbitrary numbers of specialisations (see specialise) * Specialise even for arguments that are not scrutinised in the loop (see argToPat; Trac #4488) * Only specialise on recursive types a finite number of times (see is_too_recursive; Trac #5550; Note [Limit recursive specialisation]) }}} Now there'd be five things! -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:7 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: 8.2.2 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: #11565 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by choenerzs): Ok! I've HEAD prepared, and will start with simplified example code. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:8 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: choenerzs Type: feature request | Status: new Priority: normal | Milestone: 8.4.1 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: #11565 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Changes (by bgamari): * owner: (none) => choenerzs * milestone: 8.2.2 => 8.4.1 Comment: Any progress on this, choenerzs? Bumping off to 8.4.1 as there is a workaround for the regression (e.g. increase `-fmax-worker-args`). -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:9 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14003: Allow more worker arguments in SpecConstr -------------------------------------+------------------------------------- Reporter: choenerzs | Owner: choenerzs Type: feature request | Status: new Priority: normal | Milestone: 8.4.1 Component: Compiler | Version: 8.2.1-rc3 Resolution: | Keywords: JoinPoints, | Fusion Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: #11565 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by choenerzs): Hi Ben, sorry for the delay. Work caught up with me. I think, this month should work out. Best, Christian -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14003#comment:10 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
participants (1)
-
GHC