summaryrefslogtreecommitdiff
path: root/compiler/stranal
Commit message (Collapse)AuthorAgeFilesLines
* Update Trac ticket URLs to point to GitLabRyan Scott2019-03-153-22/+22
| | | | | This moves all URL references to Trac tickets to their corresponding GitLab counterparts.
* Always do the worker/wrapper split for NOINLINEsSebastian Graf2019-03-071-0/+31
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Trac #10069 revealed that small NOINLINE functions didn't get split into worker and wrapper. This was due to `certainlyWillInline` saying that any unfoldings with a guidance of `UnfWhen` inline unconditionally. That isn't the case for NOINLINE functions, so we catch this case earlier now. Nofib results: -------------------------------------------------------------------------------- Program Allocs Instrs -------------------------------------------------------------------------------- fannkuch-redux -0.3% 0.0% gg +0.0% +0.1% maillist -0.2% -0.2% minimax 0.0% -0.8% -------------------------------------------------------------------------------- Min -0.3% -0.8% Max +0.0% +0.1% Geometric Mean -0.0% -0.0% Fixes #10069. ------------------------- Metric Increase: T9233 -------------------------
* Remove ExnStr and ThrowsExn businessSebastian Graf2019-02-011-1/+1
|
* Stomp a few typos and grammarosGabor Greif2018-12-171-2/+2
| | | | Also use 'id'
* Remove another unused import from WwLibSimon Peyton Jones2018-12-131-1/+1
|
* Remove dead codeSimon Peyton Jones2018-12-121-14/+0
| | | | | | | This is a follow-up to d77501cd5b Improvements to demand analysis I forgot to remove some now-dead code
* Improvements to demand analysisSimon Peyton Jones2018-12-122-66/+175
| | | | | | | | | | | | | | | | | | | | | | | | This patch collects a few improvements triggered by Trac #15696, and fixing Trac #16029 * Stop making toCleanDmd behave specially for unlifted types. This special case was the cause of stupid behaviour in Trac #16029. And to my joy I discovered the let/app invariant rendered it unnecessary. (Maybe the special case pre-dated the let/app invariant.) Result: less special-case handling in the compiler, and better perf for the compiled code. * In WwLib.mkWWstr_one, treat seqDmd like U(AAA). It was not being so treated before, which again led to stupid code. * Update and improve Notes There are .stderr test wibbles because we get slightly different strictness signatures for an argumment of unlifted type: <L,U> rather than <S,U> for Int# <S,U> rather than <S(S),U(U)> for Int
* dmdAnal: Move handling of datacon strictness to mkWWstr_oneBen Gamari2018-12-112-59/+81
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Previously datacon strictness was accounted for when we demand analysed a case analysis. However, this results in pessimistic demands in some cases. For instance, consider the program (from T10482) data family Bar a data instance Bar (a, b) = BarPair !(Bar a) !(Bar b) newtype instance Bar Int = Bar Int foo :: Bar ((Int, Int), Int) -> Int -> Int foo f k = case f of BarPair x y -> case burble of True -> case x of BarPair p q -> ... False -> ... We really should be able to assume that `p` is already evaluated since it came from a strict field of BarPair. However, as written the demand analyser can not conclude this since we may end up in the False branch of the case on `burble` (which places no demand on `x`). By accounting for the data con strictness later, applied to the demand of the RHS, we get the strict demand signature we want. See Note [Add demands for strict constructors] for a more comprehensive discussion. Test Plan: Validate Reviewers: simonpj, osa1, goldfire Subscribers: rwbarton, carter GHC Trac Issues: #15696 Differential Revision: https://phabricator.haskell.org/D5226
* Rename literal constructorsSylvain Henry2018-11-221-1/+1
| | | | | | | | | | | | | | | | | | | | | | | | In a previous patch we replaced some built-in literal constructors (MachInt, MachWord, etc.) with a single LitNumber constructor. In this patch we replace the `Mach` prefix of the remaining constructors with `Lit` for consistency (e.g., LitChar, LitLabel, etc.). Sadly the name `LitString` was already taken for a kind of FastString and it would become misleading to have both `LitStr` (literal constructor renamed after `MachStr`) and `LitString` (FastString variant). Hence this patch renames the FastString variant `PtrString` (which is more accurate) and the literal string constructor now uses the least surprising `LitString` name. Both `Literal` and `LitString/PtrString` have recently seen breaking changes so doing this kind of renaming now shouldn't harm much. Reviewers: hvr, goldfire, bgamari, simonmar, jrtc27, tdammers Subscribers: tdammers, rwbarton, thomie, carter Differential Revision: https://phabricator.haskell.org/D4881
* Fix cardinality change of fields in addDataConStrictnessÖmer Sinan Ağacan2018-10-151-1/+1
| | | | | | | | | | | | Test Plan: This validates Reviewers: simonpj, bgamari Reviewed By: bgamari Subscribers: rwbarton, carter Differential Revision: https://phabricator.haskell.org/D5225
* Add RubbishLit for absent bindings of UnliftedRepSebastian Graf2018-10-141-10/+26
| | | | | | | | | | | | | | | | | | | | | | | | | Summary: Trac #9279 reminded us that the worker wrapper transformation copes really badly with absent unlifted boxed bindings. As `Note [Absent errors]` in WwLib.hs points out, we can't just use `absentError` for unlifted bindings because there is no bottom to hide the error in. So instead, we synthesise a new `RubbishLit` of type `forall (a :: TYPE 'UnliftedRep). a`, which code-gen may subsitute for any boxed value. We choose `()`, so that there is a good chance that the program crashes instead instead of leading to corrupt data, should absence analysis have been too optimistic (#11126). Reviewers: simonpj, hvr, goldfire, bgamari, simonmar Reviewed By: simonpj Subscribers: osa1, rwbarton, carter GHC Trac Issues: #15627, #9279, #4306, #11126 Differential Revision: https://phabricator.haskell.org/D5153
* Define activeAfterInitial, activeDuringFinalSimon Peyton Jones2018-09-051-2/+2
| | | | | | | This is pure refactoring, just adding a couple of definitions to BasicTypes, and using them. Plus some whitespace stuff.
* Replace most occurences of foldl with foldl'.klebinger.andreas@gmx.at2018-08-211-2/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This patch adds foldl' to GhcPrelude and changes must occurences of foldl to foldl'. This leads to better performance especially for quick builds where GHC does not perform strictness analysis. It does change strictness behaviour when we use foldl' to turn a argument list into function applications. But this is only a drawback if code looks ONLY at the last argument but not at the first. And as the benchmarks show leads to fewer allocations in practice at O2. Compiler performance for Nofib: O2 Allocations: -1 s.d. ----- -0.0% +1 s.d. ----- -0.0% Average ----- -0.0% O2 Compile Time: -1 s.d. ----- -2.8% +1 s.d. ----- +1.3% Average ----- -0.8% O0 Allocations: -1 s.d. ----- -0.2% +1 s.d. ----- -0.1% Average ----- -0.2% Test Plan: ci Reviewers: goldfire, bgamari, simonmar, tdammers, monoidal Reviewed By: bgamari, monoidal Subscribers: tdammers, rwbarton, thomie, carter Differential Revision: https://phabricator.haskell.org/D4929
* Revert "Don't inline functions with RULES too early"Ben Gamari2018-08-011-2/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This commit causes significant performance regressions: ``` bytes allocated value is too high: Expected T9872d(normal) bytes allocated: 578498120 +/-5% Lower bound T9872d(normal) bytes allocated: 549573214 Upper bound T9872d(normal) bytes allocated: 607423026 Actual T9872d(normal) bytes allocated: 677179968 Deviation T9872d(normal) bytes allocated: 17.1 % bytes allocated value is too high: Expected T9872c(normal) bytes allocated: 3096670112 +/-5% Lower bound T9872c(normal) bytes allocated: 2941836606 Upper bound T9872c(normal) bytes allocated: 3251503618 Actual T9872c(normal) bytes allocated: 3601872536 Deviation T9872c(normal) bytes allocated: 16.3 % bytes allocated value is too high: Expected T9872b(normal) bytes allocated: 3730686224 +/-5% Lower bound T9872b(normal) bytes allocated: 3544151912 Upper bound T9872b(normal) bytes allocated: 3917220536 Actual T9872b(normal) bytes allocated: 4374298272 Deviation T9872b(normal) bytes allocated: 17.3 % bytes allocated value is too high: Expected T9872a(normal) bytes allocated: 2729927408 +/-5% Lower bound T9872a(normal) bytes allocated: 2593431037 Upper bound T9872a(normal) bytes allocated: 2866423779 Actual T9872a(normal) bytes allocated: 3225788896 Deviation T9872a(normal) bytes allocated: 18.2 % ``` It's not clear that this was intentional so I'm going to revert for now. This reverts commit 2110738b280543698407924a16ac92b6d804dc36.
* Don't inline functions with RULES too earlySimon Peyton Jones2018-07-311-2/+2
| | | | | | | | | Trac #15445 showed that a function with an automatically generated specialisation RULE coudl be inlined before the RULE had a chance to fire. This patch attaches a NOINLINE[2] activation to the Id, to stop this happening.
* WorkWrap: Rip out unsafeGlobalDynFlags usage in mkWwInlineRuleBen Gamari2018-06-071-1/+1
| | | | | | Subscribers: rwbarton, thomie, carter Differential Revision: https://phabricator.haskell.org/D4775
* Conservatively estimate levity in worker/wrapperBen Gamari2018-06-021-1/+11
| | | | | | | | | | | | | | | | | | | | | | | | The worker/wrapper transform needs to determine the levity of the result to determine whether it needs to introduce a lambda to preserve laziness of the result. For this is previously used isUnliftedType. However, this may fail in the presence of levity polymorphism. We now instead use isLiftedType_maybe, assuming that a lambda is needed if the levity of the result cannot be determined. Fixes #15186. Test Plan: make test=T15186 Reviewers: simonpj, goldfire, tdammers Reviewed By: simonpj Subscribers: rwbarton, thomie, carter GHC Trac Issues: #15186 Differential Revision: https://phabricator.haskell.org/D4755
* Do not unpack class dictionaries with INLINABLESimon Peyton Jones2018-04-262-28/+46
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Matthew Pickering uncovered a bad performance hole in the way that single-method dictionaries work, described in Trac #14955. See Note [Do not unpack class dictionaries] in WwLib. I tried to fix this 6 years ago, but got it slightly wrong. This patch fixes it, which makes a dramatic improvement in the test case. Nofib highlights: not much happening: Program Size Allocs Runtime Elapsed TotalMem ----------------------------------------------------------------- VSM -0.3% +2.7% -7.4% -7.4% 0.0% cacheprof -0.0% +0.1% +0.3% +0.7% 0.0% integer -0.0% +1.1% +7.5% +7.5% 0.0% tak -0.1% -0.2% 0.024 0.024 0.0% ----------------------------------------------------------------- Min -4.4% -0.2% -7.4% -7.4% -8.0% Max +0.6% +2.7% +7.5% +7.5% 0.0% Geom Mean -0.1% +0.0% +0.1% +0.1% -0.2% I investigated VSM. The patch unpacks class dictionaries a bit more than before (i.e. does so if there is no INLINABLE pragma). And that gives better code in VSM (less dictionary selection etc), but one closure gets one word bigger. I'll accept these changes in exchange for more robust performance. Some ghci.debugger output wobbled around (order of bindings being displayed). I have no idea why; but I accepted the changes.
* Inline wrappers earlierSimon Peyton Jones2018-04-201-53/+84
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | This patch has a single significant change: strictness wrapper functions are inlined earlier, in phase 2 rather than phase 0. As shown by Trac #15056, this gives a better chance for RULEs to fire. Before this change, a function that would have inlined early without strictness analyss was instead inlining late. Result: applying "optimisation" made the program worse. This does not make too much difference in nofib, but I've stumbled over the problem more than once, so even a "no-change" result would be quite acceptable. Here are the headlines: -------------------------------------------------------------------------------- Program Size Allocs Runtime Elapsed TotalMem -------------------------------------------------------------------------------- cacheprof -0.5% -0.5% +2.5% +2.5% 0.0% fulsom -1.0% +2.6% -0.1% -0.1% 0.0% mate -0.6% +2.4% -0.9% -0.9% 0.0% veritas -0.7% -23.2% 0.002 0.002 0.0% -------------------------------------------------------------------------------- Min -1.4% -23.2% -12.5% -15.3% 0.0% Max +0.6% +2.6% +4.4% +4.3% +19.0% Geometric Mean -0.7% -0.2% -1.4% -1.7% +0.2% * A worthwhile reduction in binary size. * Runtimes are not to be trusted much but look as if they are moving the right way. * A really big win in veritas, described in comment:1 of Trac #15056; more fusion rules fired. * I investigated the losses in 'mate' and 'fulsom'; see #15056.
* Fix grammaros in commentsGabor Greif2017-10-171-1/+1
|
* Fix nasty bug in w/w for absence analysisSimon Peyton Jones2017-10-031-14/+23
| | | | | | | | | | | | | | | | | | | | This dark corner was exposed by Trac #14285. It involves the interaction between absence analysis and INLINABLE pragmas. There is a full explanation in Note [aBSENT_ERROR_ID] in MkCore, which you can read there. The changes in this patch are * Make exprIsHNF return True for absentError, treating absentError like an honorary data constructor. * Make absentError /not/ be diverging, unlike other error Ids. This is all a bit horrible. * While doing this I found that exprOkForSpeculation didn't have a case for value lambdas so I added one. It's not really called on lifted types much, but it seems like the right thing
* Comments onlySimon Peyton Jones2017-10-031-1/+1
|
* A bunch of typofixesGabor Greif2017-09-261-1/+1
|
* compiler: introduce custom "GhcPrelude" PreludeHerbert Valerio Riedel2017-09-193-0/+6
| | | | | | | | | | | | | | | | | | This switches the compiler/ component to get compiled with -XNoImplicitPrelude and a `import GhcPrelude` is inserted in all modules. This is motivated by the upcoming "Prelude" re-export of `Semigroup((<>))` which would cause lots of name clashes in every modulewhich imports also `Outputable` Reviewers: austin, goldfire, bgamari, alanz, simonmar Reviewed By: bgamari Subscribers: goldfire, rwbarton, thomie, mpickering, bgamari Differential Revision: https://phabricator.haskell.org/D3989
* Allow CSE'ing of work-wrapped bindings (#14186)Joachim Breitner2017-09-121-1/+5
| | | | | | | | the worker/wrapper creates an artificial INLINE pragma, which caused CSE to not do its work. We now recognize such artificial pragmas by using `NoUserInline` instead of `Inline` as the `InlineSpec`. Differential Revision: https://phabricator.haskell.org/D3939
* Zap stable unfoldings in worker/wrapperSimon Peyton Jones2017-06-281-2/+5
| | | | This patch fixes the buglet described in Trac #13890.
* Typos in comments [ci skip]Gabor Greif2017-05-051-8/+8
|
* Improve demand analysis for join pointsSimon Peyton Jones2017-04-121-20/+62
| | | | | | | | | I realised (Trac #13543) that we can improve demand analysis for join point quite straightforwardly. The idea is explained in Note [Demand analysis for join points] in DmdAnal
* Add comments on DmdAnal space leak fixSimon Peyton Jones2017-04-061-0/+3
|
* Stamp out space leaks from demand analysisReid Barton2017-04-011-1/+21
| | | | | | | | | | | | | | | | This reduces peak memory usage by ~30% on my test case (DynFlags), and (probably as a result of reduced GC work) decreases compilation time by a few percent as well. Also fix a bug in seqStrDmd so that demeand info is fully evaluated. Reviewers: simonpj, austin, bgamari Reviewed By: bgamari Subscribers: dfeuer, thomie Differential Revision: https://phabricator.haskell.org/D3400
* Broaden demand analysis IO hack notesDavid Feuer2017-03-131-4/+8
| | | | | | | | | | | | | | | The I/O hack for demand analysis has broader and arguably more important implications than the note expressed. Broaden it. [skip ci] Reviewers: austin, bgamari Reviewed By: bgamari Subscribers: carter, rwbarton, thomie Differential Revision: https://phabricator.haskell.org/D3307
* Upgrade UniqSet to a newtypeDavid Feuer2017-03-011-1/+2
| | | | | | | | | | | | | | | | | | | | | The fundamental problem with `type UniqSet = UniqFM` is that `UniqSet` has a key invariant `UniqFM` does not. For example, `fmap` over `UniqSet` will generally produce nonsense. * Upgrade `UniqSet` from a type synonym to a newtype. * Remove unused and shady `extendVarSet_C` and `addOneToUniqSet_C`. * Use cached unique in `tyConsOfType` by replacing `unitNameEnv (tyConName tc) tc` with `unitUniqSet tc`. Reviewers: austin, hvr, goldfire, simonmar, niteria, bgamari Reviewed By: niteria Subscribers: thomie Differential Revision: https://phabricator.haskell.org/D3146
* DmdAnal: Clarify reference to Cardinality Analysis paperBen Gamari2017-02-281-1/+1
|
* Tweaks and typos in manual, note refs, commentsGabor Greif2017-02-091-1/+1
|
* Do Worker/Wrapper for NOINLINE thingsEric Seidel2017-02-051-12/+74
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Disabling worker/wrapper for NOINLINE things can cause unnecessary reboxing of values. Consider {-# NOINLINE f #-} f :: Int -> a f x = error (show x) g :: Bool -> Bool -> Int -> Int g True True p = f p g False True p = p + 1 g b False p = g b True p the strictness analysis will discover f and g are strict, but because f has no wrapper, the worker for g will rebox p. So we get $wg x y p# = let p = I# p# in -- Yikes! Reboxing! case x of False -> case y of False -> $wg False True p# True -> +# p# 1# True -> case y of False -> $wg True True p# True -> case f p of { } g x y p = case p of (I# p#) -> $wg x y p# Now, in this case the reboxing will float into the True branch, an so the allocation will only happen on the error path. But it won't float inwards if there are multiple branches that call (f p), so the reboxing will happen on every call of g. Disaster. Solution: do worker/wrapper even on NOINLINE things; but move the NOINLINE pragma to the worker. Test Plan: make test TEST="13143" Reviewers: simonpj, bgamari, dfeuer, austin Reviewed By: simonpj, bgamari Subscribers: dfeuer, thomie Differential Revision: https://phabricator.haskell.org/D3046
* Spelling fixesGabor Greif2017-02-021-1/+1
|
* Join pointsLuke Maurer2017-02-013-23/+163
| | | | | | | | | | | | | | | | | | | This major patch implements Join Points, as described in https://ghc.haskell.org/trac/ghc/wiki/SequentCore. You have to read that page, and especially the paper it links to, to understand what's going on; but it is very cool. It's Luke Maurer's work, but done in close collaboration with Simon PJ. This Phab is a squash-merge of wip/join-points branch of http://github.com/lukemaurer/ghc. There are many, many interdependent changes. Reviewers: goldfire, mpickering, bgamari, simonmar, dfeuer, austin Subscribers: simonpj, dfeuer, mpickering, Mikolaj, thomie Differential Revision: https://phabricator.haskell.org/D2853
* Record evaluated-ness on workers and wrappersSimon Peyton Jones2017-01-231-24/+84
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Summary: This patch is a refinement of the original commit (which was reverted): commit 6b976eb89fe72827f226506d16d3721ba4e28bab Date: Fri Jan 13 08:56:53 2017 +0000 Record evaluated-ness on workers and wrappers In Trac #13027, comment:20, I noticed that wrappers created after demand analysis weren't recording the evaluated-ness of strict constructor arguments. In the ticket that led to a (debatable) Lint error but in general the more we know about evaluated-ness the better we can optimise. This commit adds that info * both in the worker (on args) * and in the wrapper (on CPR result patterns). See Note [Record evaluated-ness in worker/wrapper] in WwLib On the way I defined Id.setCaseBndrEvald, and used it to shorten the code in a few other places Then I added test T13077a to test the CPR aspect of this patch, but I found that Lint failed! Reason: simpleOptExpr was discarding evaluated-ness info on lambda binders because zapFragileIdInfo was discarding an Unfolding of (OtherCon _). But actually that's a robust unfolding; there is no need to discard it. To fix this: * zapFragileIdInfo only zaps fragile unfoldings * Replace isClosedUnfolding with isFragileUnfolding (the latter is just the negation of the former, but the nomenclature is more consistent). Better documentation too Note [Fragile unfoldings] * And Simplify.simplLamBndr can now look at isFragileUnfolding to decide whether to use the longer route of simplUnfolding. For some reason perf/compiler/T9233 improves in compile-time allocation by 10%. Hooray Nofib: essentially no change: -------------------------------------------------------------------------------- Program Size Allocs Runtime Elapsed TotalMem -------------------------------------------------------------------------------- cacheprof +0.0% -0.3% +0.9% +0.4% +0.0% -------------------------------------------------------------------------------- Min +0.0% -0.3% -2.4% -2.4% +0.0% Max +0.0% +0.0% +9.8% +11.4% +2.4% Geometric Mean +0.0% -0.0% +1.1% +1.0% +0.0%
* Typos and grammar in manual/commentsGabor Greif2017-01-232-2/+2
|
* Spelling fixes in comments [ci skip]Gabor Greif2017-01-181-2/+2
|
* Typos in comments only [ci skip]Gabor Greif2017-01-181-1/+1
|
* Revert "Record evaluated-ness on workers and wrappers"Matthew Pickering2017-01-151-83/+24
| | | | | | | This reverts commit 6b976eb89fe72827f226506d16d3721ba4e28bab. Ben, Ryan and I decided to revert this for now due to T12234 failing and causing all harbormaster builds to fail.
* Record evaluated-ness on workers and wrappersSimon Peyton Jones2017-01-131-24/+83
| | | | | | | | | | | | | | | | In Trac #13027, comment:20, I noticed that wrappers created after demand analysis weren't recording the evaluated-ness of strict constructor arguments. In the ticket that led to a (debatable) Lint error but in general the more we know about evaluated-ness the better we can optimise. This commit adds that info both in the worker (on args) and in the wrapper (on CPR result patterns). See Note [Record evaluated-ness in worker/wrapper] in WwLib On the way I defined Id.setCaseBndrEvald, and used it to shorten the code in a few other places
* Never apply worker/wrapper to DFunsSimon Peyton Jones2016-12-211-7/+3
| | | | | | | | | While fixing Trac #12444 I found an occasion on which we applied worker/wrapper to a DFunId. This is bad: it destroys the magic DFunUnfolding. This patch is a minor refactoring that stops this corner case happening, and tidies up the code a bit too.
* Add HsSyn prettyprinter testsAlan Zimmerman2016-12-071-3/+3
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Summary: Add prettyprinter tests, which take a file, parse it, pretty print it, re-parse the pretty printed version and then compare the original and new ASTs (ignoring locations) Updates haddock submodule to match the AST changes. There are three issues outstanding 1. Extra parens around a context are not reproduced. This will require an AST change and will be done in a separate patch. 2. Currently if an `HsTickPragma` is found, this is not pretty-printed, to prevent noise in the output. I am not sure what the desired behaviour in this case is, so have left it as before. Test Ppr047 is marked as expected fail for this. 3. Apart from in a context, the ParsedSource AST keeps all the parens from the original source. Something is happening in the renamer to remove the parens around visible type application, causing T12530 to fail, as the dumped splice decl is after the renamer. This needs to be fixed by keeping the parens, but I do not know where they are being removed. I have amended the test to pass, by removing the parens in the expected output. Test Plan: ./validate Reviewers: goldfire, mpickering, simonpj, bgamari, austin Reviewed By: simonpj, bgamari Subscribers: simonpj, goldfire, thomie, mpickering Differential Revision: https://phabricator.haskell.org/D2752 GHC Trac Issues: #3384
* Fix shadowing in mkWwBodiesSimon Peyton Jones2016-10-172-27/+55
| | | | | | | | | | | | | | | | | | This bug, exposed by Trac #12562 was very obscure, and has been lurking for a long time. What happened was that, in the worker/wrapper split a tyvar binder for a worker function accidentally shadowed an in-scope term variable that was mentioned in the body of the function It's jolly hard to provoke, so I have not even attempted to make a test case. There's a Note [Freshen WW arguments] to explain. Interestingly, fixing the bug (which meant fresher type variables) revealed a second lurking bug: I'd failed to apply the substitution to the coercion in the second last case of mkWWArgs, which introduces a Cast.
* extend '-fmax-worker-args' limit to specialiser (Trac #11565)Sergei Trofimovich2016-09-021-5/+9
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | It's a complementary change to a48de37dcca98e7d477040b0ed298bcd1b3ab303 restore -fmax-worker-args handling (Trac #11565) I don't have a small example but I've noticed another discrepancy when was profiling GHC for performance cmmExprNative :: ReferenceKind -> CmmExpr -> CmmOptM CmmExpr was specialised by 'spec_one' down to a function with arity 159. As a result 'perf record' pointed at it as at slowest function in whole ghc library. I've extended -fmax-worker-args effect to 'spec_one' as it does the same worker/wrapper split to push arguments to the heap. The change decreases heap usage on a synth.bash benchmark (Trac #9221) from 67G down to 64G (-4%). Benchmark runtime decreased from 14.5 s down to 14.s (-7%). Signed-off-by: Sergei Trofimovich <siarheit@google.com> Reviewers: ezyang, simonpj, austin, goldfire, bgamari Subscribers: thomie Differential Revision: https://phabricator.haskell.org/D2507 GHC Trac Issues: #11565
* restore -fmax-worker-args handling (Trac #11565)Sergei Trofimovich2016-09-011-1/+29
| | | | | | | | | | | | | | | | | | | | | | | | | | maxWorkerArgs handling was accidentally lost 3 years ago in a major update of demand analysis commit 0831a12ea2fc73c33652eeec1adc79fa19700578 Old regression is noticeable as: - code bloat (requires stack reshuffling) - compilation slowdown (more code to optimise/generate) - and increased heap usage (DynFlags unboxing/reboxing?) On a simple compile benchmark this change causes heap allocation drop from 70G don to 67G (ghc perf build). Signed-off-by: Sergei Trofimovich <siarheit@google.com> Reviewers: simonpj, ezyang, goldfire, austin, bgamari Reviewed By: simonpj, ezyang Subscribers: thomie Differential Revision: https://phabricator.haskell.org/D2503 GHC Trac Issues: #11565
* WwLib: Add strictness signature to "let x = absentError …"Joachim Breitner2016-08-251-5/+9
| | | | | indicating that it is bottom. This should help making the "empty cases" lint error give less false alarms.
* DmdAnal: Add a final, safe iterationJoachim Breitner2016-08-251-78/+110
| | | | | | | | | | | | | | | | | | | this fixes #12368. It also refactors dmdFix a bit, removes some redundancies (such as passing around an strictness signature right next to an id, when that id is guaranteed to have been annotated with that strictness signature). Note that when fixed-point iteration does not terminate, we conservatively delete their strictness signatures (set them to nopSig). But this loses the information on how its strict free variables are used! Lazily used variables already escape via lazy_fvs. We ensure that in the case of an aborted fixed-point iteration, also the strict variables are put there (with a conservative demand of topDmd). Differential Revision: https://phabricator.haskell.org/D2392
View Lorry Controller Status