- Oct 06, 2021
-
-
David Feuer authored
It's generally best practice to limit unsafe equalities as much as possible, to avoid mistakes. A good way to do that is to use `unsafeEqualityProof` to produce fake evidence of the smallest equality needed. For example, if you know that `A` and `P` are sufficiently similar and you need to coerce `[A] -> B -> C` to `[P] -> B -> C`, it's much nicer to explicitly introduce fake evidence that `A ~ P` than to use `unsafeCoerce` and hope you aren't accidentally coercing some unrelated things elsewhere in the type. Unfortunately, actually using the limited fake evidence can be tricky. You might think you could write something like ```haskell cabc :: ([A] -> B -> C) -> [P] -> B -> C cabc f = case unsafeEqualityProof @A @P of UnsafeRefl -> f ``` But this won't work! When type checking the case branch, GHC rejects `A ~ P` without even noticing that there's evidence for it available. This means that more explicit utilities are needed to manipulate unsafe equalities. One simple option would be to offer just a single function ```haskell unsafeEqualityProofToEquality :: UnsafeEquality a b -> a :~: b unsafeEqualityProofToEquality UnsafeRefl = Refl ``` Then people could use the utilities in `Data.Type.Equality` to manipulate the resulting equality. But it seems to me that it's much nicer to help users keep equality proofs *obviously* unsafe right up to the point of use, to avoid a sort of "equality blindness". This commit adds versions of the various equality proof functions in `Data.Type.Equality` for `UnsafeEquality`, along with a few more for convenience inspired by the Agda standard library.
-
- Oct 05, 2021
-
-
-
-
Ticket #20200 (the Agda failure) showed another case in which lookupIdSubst would fail to find a local Id in the InScopeSet. This time it was because SetLevels was given a program in which the top-level bindings were not in dependency order. The Simplifier (see Note [Glomming] in GHC.Core.Opt.Occuranal) and the specialiser (see Note [Top level scope] in GHC.Core.Opt.Specialise) may both produce top-level bindings where an early binding refers to a later one. One solution would be to run the occurrence analyser again to put them all in the right order. But a simpler one is to make SetLevels OK with this input by bringing all top-level binders into scope at the start. That's what this patch does.
-
This (big) commit finishes porting the GHC.Tc.Deriv module to support the new diagnostic infrastructure (#18516) by getting rid of the legacy calls to `TcRnUnknownMessage`. This work ended up being quite pervasive and touched not only the Tc.Deriv module but also the Tc.Deriv.Utils and Tc.Deriv.Generics module, which needed to be adapted to use the new infrastructure. This also required generalising `Validity`. More specifically, this is a breakdown of the work done: * Add and use the TcRnUselessTypeable data constructor * Add and use TcRnDerivingDefaults data constructor * Add and use the TcRnNonUnaryTypeclassConstraint data constructor * Add and use TcRnPartialTypeSignatures * Add T13324_compile2 test to test another part of the TcRnPartialTypeSignatures diagnostic * Add and use TcRnCannotDeriveInstance data constructor, which introduces a new data constructor to TcRnMessage called TcRnCannotDeriveInstance, which is further sub-divided to carry a `DeriveInstanceErrReason` which explains the reason why we couldn't derive a typeclass instance. * Add DerivErrSafeHaskellGenericInst data constructor to DeriveInstanceErrReason * Add DerivErrDerivingViaWrongKind and DerivErrNoEtaReduce * Introduce the SuggestExtensionInOrderTo Hint, which adds (and use) a new constructor to the hint type `LanguageExtensionHint` called `SuggestExtensionInOrderTo`, which can be used to give a bit more "firm" recommendations when it's obvious what the required extension is, like in the case for the `DerivingStrategies`, which automatically follows from having enabled both `DeriveAnyClass` and `GeneralizedNewtypeDeriving`. * Wildcard-free pattern matching in mk_eqn_stock, which removes `_` in favour of pattern matching explicitly on `CanDeriveAnyClass` and `NonDerivableClass`, because that determine whether or not we can suggest to the user `DeriveAnyClass` or not.
-
This commit makes the `Validity` type polymorphic: ``` data Validity' a = IsValid -- ^ Everything is fine | NotValid a -- ^ A problem, and some indication of why -- | Monomorphic version of @Validity'@ specialised for 'SDoc's. type Validity = Validity' SDoc ``` The type has been (provisionally) renamed to Validity' to not break existing code, as the monomorphic `Validity` type is quite pervasive in a lot of signatures in GHC. Why having a polymorphic Validity? Because it carries the evidence of "what went wrong", but the old type carried an `SDoc`, which clashed with the new GHC diagnostic infrastructure (#18516). Having it polymorphic it means we can carry an arbitrary, richer diagnostic type, and this is very important for things like the `checkOriginativeSideConditions` function, which needs to report the actual diagnostic error back to `GHC.Tc.Deriv`. It also generalises Validity-related functions to be polymorphic in @a@.
-
-
-
-
-
-
-
-
We've had Sum CPR (#5075) for top-level bindings for a couple of years now. That begs the question why we didn't also activate it for local bindings, and the reasons for that are described in `Note [CPR for sum types]`. Only that it didn't make sense! The Note said that Sum CPR would destroy let-no-escapes, but that should be a non-issue since we have syntactic join points in Core now and we don't WW for them (`Note [Don't w/w join points for CPR]`). So I simply activated CPR for all bindings of sum type, thus fixing #5075 and \#16570. NoFib approves: ``` -------------------------------------------------------------------------------- Program Allocs Instrs -------------------------------------------------------------------------------- comp_lab_zift -0.0% +0.7% fluid +1.7% +0.7% reptile +0.1% +0.1% -------------------------------------------------------------------------------- Min -0.0% -0.2% Max +1.7% +0.7% Geometric Mean +0.0% +0.0% ``` There were quite a few metric decreases on the order of 1-4%, but T6048 seems to regress significantly, by 26.1%. WW'ing for a `Just` constructor and the nested data type meant additional Simplifier iterations and a 30% increase in term sizes as well as a 200-300% in type sizes due to unboxed 9-tuples. There's not much we can do about it, I'm afraid: We're just doing much more work there. Metric Decrease: T12425 T18698a T18698b T20049 T9020 WWRec Metric Increase: T6048
-
The examples in the Note were inaccurate (`$s$dm` has arity 1 and that seems OK) and the code didn't actually nuke the demand *signature* anyway. Specialise has to nuke it, but it starts from a clean IdInfo anyway (in `newSpecIdM`). So I just deleted the code. Fixes #20450.
-
In #18824 we saw that the Simplifier didn't nuke a CPR signature of a join point when it pushed a continuation into it when it better should have. But join points are local, mostly non-exported bindings. We don't use their CPR signature anyway and would discard it at the end of the Core pipeline. Their main purpose is to propagate CPR info during CPR analysis and by the time worker/wrapper runs the signature will have served its purpose. So we zap it! Fixes #18824.
-
We should reject "type family Foo where Bar = ()". This check was done in kcTyFamInstEqn but not in tcTyFamInstEqn. I factored out arity checking, which was duplicated.
-
By adding an early abort flag in `TcSEnv`, we can fail fast in the presence of insoluble constraints. This helps us avoid a lot of work in valid hole-fits, and we geta massive speed-up by avoiding a lot of useless work solving constraints that never come into play. Additionally, we add a simple check for degenerate hole types, such as when the type of the hole is an immutable type variable (as is the case when the hole is completely unconstrained). Then the only valid fits are the locals, so we can ignore the global candidates. This fixes #16875
-
-
-
The EpaAnnCO we were using contained an Anchor instead of EpaLocation, making it harder to work with. At the same time, using EpaLocation by itself isn't possible either, as we may have tokens without location information. Hence the new data type: data TokenLocation = NoTokenLoc | TokenLoc !EpaLocation
-
Previously registration of ticky entry counters was racy, performing a read-modify-write to add the new counter to the ticky_entry_ctrs list. This could result in the list becoming cyclic if multiple threads entered the same closure simultaneously. Fixes #20451.
-
-
This reduces the output from the testsuite to a more manageable level. Fixes #20432
-
- Oct 04, 2021
-
-
-
Sylvain Henry authored
Only for small integral types for now.
-
Modern clang versions are quite picky when it comes to reporting redundant arguments. In particular, they will warn when -no-pie is passed when no linking is necessary. Previously the configure script used a `$CC -Werror -no-pie -E` invocation to test whether `-no-pie` is necessary. Unfortunately, this meant that clang would throw a redundant argument warning, causing configure to conclude that `-no-pie` was not supported. We now rather use `$CC -Werror -no-pie`, ensuring that linking is necessary and avoiding this failure mode. Fixes #20463.
-
Closes #20404.
-
- Oct 03, 2021
-
-
Ben Gamari authored
-
Ben Gamari authored
Previously we would push perf notes using a standard user and SSH key-based authentication. However, configuring SSH is unnecessarily fiddling. We now rather use HTTPS and a project access token.
-
- Oct 02, 2021
-
-
-
When we derive the Show instance of the big record in #16577, I get the following compilation times (with -O): Before: 0.91s After: 0.77s Metric Decrease: T19695 -
This fixes an inconsistency where one dirtiness check would not mask out the STACK_DIRTY flag, meaning it may also be affected by the STACK_SANE flag.
-
Previously PerformPut failed to respect the non-moving collector's snapshot invariant, hiding references to an MVar and its new value by overwriting a stack frame without dirtying the stack. Fix this. PerformTake exhibited a similar bug, failing to dirty (and therefore mark) the blocked stack before mutating it. Closes #20399.
-
Close #20356. See addendum to Note [coreView vs tcView] in GHC.Core.Type for the details. Also killed old Note about metaTyVarUpdateOK, which has been gone for some time. test case: typecheck/should_fail/T20356
-
Like S390 and PPC64, RISC-V requires parameters for foreign calls to be extended to full words.
-
Closes #20307.
-
To install libncurses-dev on Debian targets.
-
- Oct 01, 2021
-
-
The goal here is to somewhat sanitize the environment so that performance tests don't fluctuate as much as they have been doing. In particular the length of the commit message was causing benchmarks to increase because gitlab stored the whole commit message twice in environment variables. Therefore when we used `getEnvironment` it would cause more allocation because more string would be created. See #20431 ------------------------- Metric Decrease: T10421 T13035 T18140 T18923 T9198 T12234 T12425 -------------------------
-
Old strategy: For each variable linearly scan through all the blocks and check to see if the variable is any of the block register mappings. This is very slow when you have a lot of blocks. New strategy: Maintain a map from virtual registers to the first real register the virtual register was assigned to. Consult this map in findPrefRealReg. The map is updated when the register mapping is updated and is hidden behind the BlockAssigment abstraction. On the mmark package this reduces compilation time from about 44s to 32s. Ticket: #19471
-
