# 2.1. Version 9.4.1¶

The significant changes to the various parts of the compiler are listed in the following sections.

The LLVM backend of this release is to be used with LLVM 10, 11, 12, or 13.

## 2.1.1. Breaking Changes¶

In this section we list changes that may require changes in user programs while upgrading to GHC 9.4:

• There were previously cases around functional dependencies and injective type families where the result of type inference would depend on the order of constraints, as written in a source file. These cases are fundamentally ambiguous. While GHC previously made an arbitrary decision, it now notices the ambiguity and rejects the program. This means that some previously accepted programs are now rejected. The solution is to add a type annotation or type application to resolve the ambiguity.

This is the fix for #18851.

• Unboxed sums now require the UnboxedSums extension to be enabled.

• ArrayArray# and its operations are no longer exported from GHC.Prim and are deprecated, having been superceded by the now levity-polymorphic Array# type.

• The type equality operator, (~), is not considered to be a type operator (exported from Prelude) and therefore requires the enabling of the TypeOperators extension rather than GADTs or TypeFamilies as was sufficient previously.

• GHC is now a bit better at detecting redundant hs-boot files and may report warnings where it previously did not. In such cases the correct solution is generally to remove the hs-boot file in question.

• The boxed Word64 and Int64 types are now internally represented by the unboxed Word64# and Int64# primitive types, in contrast previous releases where they were represented by Word# and Int# on 64-bit platforms.

• Due to various changes in the typechecker’s constraint solver, some programs may need additional constraints to be explicitly provided. See the Migration guide for details.

• When LambdaCase is enabled, GHC will now parse the sequence \ cases as the herald of a multi-pattern lambda-case expression. Consequently, programs of the form \ cases -> ... will not parse; the solution is likely to rename the cases binder.

## 2.1.2. Language¶

• GHC Proposal #511 has been implemented, introducing a new language extension, :lang-ext:DeepSubsumption. This extension allows the user to opt-in to the deep type subsumption-checking behavior implemented by GHC 8.10 and earlier.

• A small change has been made to the way GHC infers types for definitions with no type signature: GHC will no longer generalize a function over a type variable determined by a functional dependency. For example:

class C a b | a -> b where
op :: a -> b -> ()
f x = op True x


Previously, GHC inferred f :: C Bool b => b -> (). However, the functional dependency says that only one type could ever be used for b: this function is hardly valid “for all” bs. With the change, GHC will reject, looking for the (non-existent) instance for C Bool b.

If you want to retain the old behavior, add a (backward-compatible) type signature, explicitly requesting this unusual quantification.

• GHC Proposal #371 has been implemented. This means:

• The use of equality constraints no longer requires -XGADTs or -XTypeFamilies.

• The use of equality constraint syntax a ~ b requires -XTypeOperators, otherwise results in a warning (-Wtype-equality-requires-operators).

• (~) is now a legal name for a user-defined type operator:

class a ~ b where
...


This used to be rejected with “Illegal binding of built-in syntax”.

• The built-in type equality is now exported from Data.Type.Equality and re-exported from Prelude. When (~) is not in scope, its use results in a warning (-Wtype-equality-out-of-scope).

• GHC Proposal #302 has been implemented. This means under -XLambdaCase, a new expression heralded by \cases is available, which works like \case but can match on multiple patterns. This means constructor patterns with arguments have to parenthesized here, just like in lambda expressions.

• The parsing of implicit parameters is slightly more permissive, as GHC now allows

foo :: (?ip :: forall a. a -> a)


without requiring parentheses around forall a. a -> a. Note that implicit parameters with such kinds are unlikely to be very useful, due to #18759.

• Changes to the treatment of UnboxedSums:

## 2.1.3. Compiler¶

• The compiler now accepts arguments via GNU-style response files (#16476).

• New -Wredundant-strictness-flags that checks for strictness flags (!) applied to unlifted types, which are always strict.

• New -Wforall-identifier (enabled by default) that warns against using the name forall as an identifer on the term level.

• New -fprof-late that adds automatic CCS annotations to all top level functions after core optimisation have been run.

• New -fprof-manual which allows surpression of profiling cost centre annotations. It can be disabled as well which can be useful to surpress cost centres originating in library code.

• Typechecking plugins now support type-family rewriting. The TcPlugin datatype now contains an extra field, tcPluginRewrite, which allows typechecking plugin authors to specify which type families should be rewritten by the plugin, returning for each type family application a TcPluginRewriteResult. In addition, typechecking plugins now have the ability to emit new constraints at the same time as contradictions. To account for these changes, the TcPluginResult datatype has been renamed to TcPluginSolveResult, which bundles pattern synonyms TcPluginOk and TcPluginContradiction to recover the old interface. Typechecking plugins now have access to irreducible Given constraints, e.g., to enable the plugins to reduce the constraints.

• A new type of plugin: defaulting plugins. These plugins can propose defaults for ambiguous variables that would otherwise cause errors just like the built-in defaulting mechanism.

• GHC.Plugins.parsedResultAction now takes and returns a value of type ParsedResult, containing the HsParsedModule as well as PsMessages, which contains warnings and errors encountered by the parser before they’re shown to the user, as long as none of the errors prevented the AST from being built. This means plugins can remove, modify, or add any of these, or simply pass them through unchanged.

• The way GHC checks for representation polymorphism has been overhauled: all the checks are now done during typechecking. The error messages now contain more detailed information about the specific check that was performed.

• A new pragma, OPAQUE, that ensures that every call of a named function annotated with an OPAQUE pragma remains a call of that named function, not some name-mangled variant. This implements GHC Proposal #415.

• Constructed Product Result analysis (c.f. -fcpr-anal) has been overhauled and will now unbox nestedly, if termination properties of the function permit. This allows unboxing of constructed results returned by IO actions. E.g.:

sumIO :: [Int] -> IO Int
sumIO []     = return 0
sumIO (x:xs) = do
r <- sumIO xs
return $! x + r  Note the use of $!: Without it, GHC would be unable to see that evaluation of r and x terminates (and rapidly, at that). An alternative would be to evaluate both with a bang pattern or a seq, but the return $! <res> idiom should work more reliably and needs less thinking. • Demand analysis (cf. -fstrictness) now integrates a Boxity Analysis that tracks whether a function needs a parameter boxed. If that is the case, the worker/wrapper transformation (cf. -fworker-wrapper) will not unbox that parameter, leading to less reboxing in many cases. For reasons of backwards-compatible performance, you may find that the new mechanism is too aggressive in a few cases (e.g., still unboxing a parameter that is used boxed in a hot path). Do post a bug report with your example! Then wrap the uses of the parameter in GHC.Exts.lazy for a short-term fix. • Tag inference has been implemented. It’s a new backend optimization pass aimed at avoiding redundant evaluatedness checks. The basic pass is always enabled and not optional. When using -fworker-wrapper-cbv it additionally will generate workers for functions with strict arguments, pushing the evaluation+tagging of the arguments into the wrapper and allowing the worker to simply assume all arguments are fully evaluated and properly tagged. Usually the wrapper will then inline, and if the argument is known to be properly tagged at the call site the wrapper will become a no-op. Giving us a more efficient worker without adding any overhead. If the argument isn’t known to be evaluated we perform the same amount of work, but do it at call sites instead of inside the called function. In general -fworker-wrapper-cbv is very beneficial and can be safely enabled. However sadly there are two exceptions. It can break rules for code which made assumptions about which functions get a W/W split which now no longer hold. See #20364 for the details. For this reason it isn’t enabled by default. For code which has the proper INLINABLE (INLINABLE pragma) and INLINE (INLINE pragma) or that doesn’t define any rule-relevant functions this shouldn’t happen. The longterm fix here is to apply the proper pragmas. There is also a known issue where a function taking multiple unlifted arguments can cause excessive spilling (#20334). This seems to be an edge case. But if you think you are hitting this case please comment on the ticket so that we can prioritize it accordingly. • Support for Sun SPARC architecture has been dropped (#16883). • A fix for GHC’s handling of the XDG Base Directory Specification (#6077, #20684, #20669, #20660): • For the package database previously in ~/.ghc/<arch-ver>, we will continue to use the old path if it exists. For example, if the ~/.ghc/x86_64-linux-9.4.1 directory exists, GHC will use that for its user package database. If this directory does not exist, we will use $XDG_DATA_HOME/ghc/x86_64-linux-9.4.1. This is in order to give tooling like cabal time to migrate
• For GHCi configuration files previously located in ~/.ghc/ like ghci.conf and ghci_history, we will first check if they exist in ~/.ghc and use those if they do. However, we will create new files like ghci_history only in $XDG_DATA_HOME/ghc. So if you don’t have a previous GHC installation which created ~/.ghc/ghci_history, the history file will be written to $XDG_DATA_HOME/ghc. If you already have an older GHC installation which wrote ~/.ghc/ghci_history, then GHC will continue to write the history to that file.
• The -Wunticked-promoted-constructors warning is no longer enabled with -Wall (#20531), as a part of long-term push towards Dependent Haskell.

• In GHCi, the :type command no longer instantiates quantified type variables when given a polymorphic type. (It used to instantiate inferred type variables.)

## 2.1.4. Packaging¶

• GHC’s package database now comes with a virtual system-cxx-std-lib package which captures the compiler configuration necessary to link aginst the C++ standard library.

## 2.1.5. Runtime system¶

• Support for GHC’s eventlog is now enabled in all runtime system configurations, eliminating the need to pass the -eventlog flag to use the eventlog. This flag has been deprecated (#18948).
• Summary statistics, i.e. the output of -s [⟨file⟩], now correctly accounts for bytes copied during sequential collections.

## 2.1.6. base library¶

• GHC.Generics now provides a set of newtypes, Generically and Generically1, for deriving generic instances via :lang-ext:DerivingVia. Generically instances include Semigroup and Monoid.

• There’s a new special function withDict in GHC.Exts:

withDict :: forall {rr :: RuntimeRep} cls meth (r :: TYPE rr). WithDict cls meth => meth -> (cls => r) -> r


where cls must be a class containing exactly one method, whose type must be meth. This requirement is enforced by the constraint WithDict cls meth.

This function converts meth to a type class dictionary. It removes the need for unsafeCoerce in implementation of reflection libraries. It should be used with care, because it can introduce incoherent instances.

• See the base library’s changelog.md for a full accounting.

## 2.1.7. ghc-prim library¶

• Primitive types and functions which handle boxed values are now levity-polymorphic, meaning that they now also work with unlifted boxed values (i.e. values whose type has kind TYPE (BoxedRep Unlifted)).

The following type constructors are now levity-polymorphic:

 Array# SmallArray# Weak# StablePtr# StableName# MutableArray# SmallMutableArray# MutVar# TVar# MVar# IOPort#

For example, Array# used to have kind:

Type -> UnliftedType


but it now has kind:

forall {l :: Levity}. TYPE (BoxedRep l) -> UnliftedType


Similarly, MutVar# used to have kind:

Type -> Type -> UnliftedType


but it now has kind:

forall {l :: Levity}. Type -> TYPE (BoxedRep l) -> UnliftedType


This means that in Array# a, MutableArray# s a, MutVar# s a, …, the element type a, must always be boxed, but it can now either be lifted or unlifted. In particular, arrays and mutable variables can now be used to store other arrays and mutable variables.

All functions which use these updated primitive types are also levity-polymorphic:

• all array operations (reading/writing/copying/…), for both arrays and small arrays, mutable and immutable:
• newArray#, readArray#, writeArray#, sizeofArray#, sizeofMutableArray#, indexArray#, unsafeFreezeArray#, unsafeThawArray#, copyArray#, copyMutableArray#, cloneArray#, cloneMutableArray#, freezeArray#, thawArray#, casArray#,
• newSmallArray#, shrinkSmallMutableArray#, readSmallArray#, writeSmallArray#, sizeofSmallArray#, getSizeofSmallMutableArray#, indexSmallArray#, unsafeFreezeSmallArray#, unsafeThawSmallArray#, copySmallArray#, copySmallMutableArray#, cloneSmallArray#, cloneSmallMutableArray#, freezeSmallArray#, thawSmallArray#, casSmallArray#,
• newMutVar#, readMutVar#, writeMutVar#, casMutVar#,
• operations on MVar# and TVar#:
• newTVar#, readTVar#, readTVarIO#, writeTVar#,
• newMVar#, takeMVar#, tryTakeMVar#, putMVar#, tryPutMVar#, readMVar#, tryReadMVar#,
• STM operations atomically#, retry#, catchRetry# and catchSTM#.
• newIOPort#, readIOPort#, writeIOPort#,
• mkWeak#, mkWeakNoFinalizer#, addCFinalizerToWeak#, deRefWeak#, finalizeWeak#,
• makeStablePtr#, deRefStablePtr#, eqStablePtr#, makeStableName#, stableNameToInt#,

For example, the full type of newMutVar# is now:

newMutVar#
:: forall {l :: Levity} s (a :: TYPE (BoxedRep l)).
a -> State# s -> (# State# s, MVar# s a #)


and the full type of writeSmallArray# is:

writeSmallArray#
:: forall {l :: Levity} s (a :: TYPE (BoxedRep l)).
SmallMutableArray# s a -> Int# -> a -> State# s -> State# s

• ArrayArray# and MutableArrayArray# have been moved from GHC.Prim to GHC.Exts. They are deprecated, because their functionality is now subsumed by Array# and MutableArray#.

• mkWeak#, mkWeakNoFinalizer#, touch# and keepAlive# are now levity-polymorphic instead of representation-polymorphic. For instance:

mkWeakNoFinalizer#
:: forall {l :: Levity} {k :: Levity}
(a :: TYPE (BoxedRep l))
(b :: TYPE (BoxedRep k)).
a -> b -> State# RealWorld -> (# State# RealWorld, Weak# b #)


That is, the type signature now quantifies over the GHC.Exts.Levity of a instead of its GHC.Exts.RuntimeRep. In addition, this variable is now inferred, instead of specified, meaning that it is no longer eligible for visible type application. Note that b is now also levity-polymorphic, due to the change outlined in the previous point.

• Primitive functions for throwing and catching exceptions are now more polymorphic than before. For example, catch# now has type:

catch#
:: forall {r :: RuntimeRep} {l :: Levity}
(a :: TYPE r)
(b :: TYPE (BoxedRep l)).
( State# RealWorld -> (# State# RealWorld, a #) )
-> ( b -> State# RealWorld -> (# State# RealWorld, a #) )
-> State# RealWorld -> (# State# RealWorld, a #)


The following functions have been generalised in this way:

• catch#,
• raise#, raiseIO#,
• maskAsyncExceptions#, maskUninterruptible#, unmaskAsyncExceptions#.

Note in particular that raise# is now both representation-polymorphic (with an inferred RuntimeRep argument) and levity-polymorphic, with type:

raise# :: forall {l :: Levity} {r :: RuntimeRep}
(a :: TYPE (BoxedRep l))
(b :: TYPE r).
a -> b

• fork# and forkOn# are now representation-polymorphic. For example, fork# now has type:

fork# :: forall {r :: RuntimeRep} (a :: TYPE r).
(State# RealWorld -> (# State# RealWorld, a #))
-> (State# RealWorld -> (# State# RealWorld, a #))

• GHC.Exts.reallyUnsafePtrEquality# has been made more general, as it is now both levity-polymorphic and heterogeneous:

reallyUnsafePtrEquality#
:: forall {l :: Levity} {k :: Levity}
(a :: TYPE (BoxedRep l))
(b :: TYPE (BoxedRep k))
. a -> b -> Int#


This means that GHC.Exts.reallyUnsafePtrEquality# can be used on primitive arrays such as GHC.Exts.Array# and GHC.Exts.ByteArray#. It can also be used on values of different types, without needing to call GHC.Exts.unsafeCoerce#.

• Added GHC.Exts.reallyUnsafePtrEquality which recovers the previous behaviour of GHC.Exts.reallyUnsafePtrEquality#:

reallyUnsafePtrEquality :: forall (a :: Type). a -> a -> Int#

• Added GHC.Exts.sameArray#, GHC.Exts.sameSmallArray#, GHC.Exts.sameByteArray# and GHC.Exts.sameArrayArray#:

sameArray# :: Array# a -> Array# a -> Int#
sameSmallArray# :: SmallArray# a -> SmallArray# a -> Int#
sameByteArray# :: ByteArray# -> ByteArray# -> Int#
sameArrayArray# :: ArrayArray# -> ArrayArray# -> Int#


## 2.1.8. ghc library¶

• The load function no longer automatically caches interface files in memory between calls. If you want to use a cache then you can supply one explicitly using the loadWithCache function, with your own implementation or a simple cache created by newIfaceCache.
• A new GHC.Hs.Syn.Type module has been introduced which defines functions for computing the Type of an HsExpr GhcTc in a pure fashion. The hsLitType and hsPatType functions that previously lived in GHC.Tc.Utils.Zonk have been moved to this module.
• A Typeable constraint has been added to fromStaticPtr in the class GHC.StaticPtr.IsStatic. GHC automatically wraps each use of the static keyword with fromStaticPtr. Because static requires its argument to be an instance of Typeable, fromStaticPtr can safely carry this constraint as well.
• The newWanted function exported by GHC.Tc.Plugin now passes on the full CtLoc instead of reconstituting it from the type-checking environment. This makes newWanted consistent with newGiven. For authors of type-checking plugins, this means you don’t need to wrap a call to newWanted in setCtLocM to create a new Wanted constraint with the provided CtLoc.
• GHC no longer carries Derived constraints. Accordingly, several functions in the plugin architecture that previously passed or received three sets of constraints (givens, deriveds, and wanteds) now work with two such sets.
• A new argument has been added to the HsOpTy constructor of the HsType datatype, to track the presence of a promotion tick. Plugins which manipulate the Haskell AST will need to take this change into account.
• Removed lookupOrigIO in favor of lookupNameCache.
• Added a new thNameToGhcNameIO function that plugins can use outside the CoreM monad.

## 2.1.9. ghc-heap library¶

• The link field of GHC.Exts.Heap.WeakClosure has been replaced with a weakLink field which is Nothing if and only if link would have been NULL.