GitLab

Now that flattening doesn't produce flattening variables,
it's not really flattening anything: it's rewriting. This
change also means that the rewriter can no longer be confused
the core flattener (in GHC.Core.Unify), which is sometimes used
during type-checking.

This patch redesigns the flattener to simplify type family applications
directly instead of using flattening meta-variables and skolems. The key new
innovation is the CanEqLHS type and the new CEqCan constraint (Ct). A CanEqLHS
is either a type variable or exactly-saturated type family application; either
can now be rewritten using a CEqCan constraint in the inert set.

Because the flattener no longer reduces all type family applications to
variables, there was some performance degradation if a lengthy type family
application is now flattened over and over (not making progress). To
compensate, this patch contains some extra optimizations in the flattener,
leading to a number of performance improvements.

Close #18875.
Close #18910.

There are many extra parts of the compiler that had to be affected in writing
this patch:

* The family-application cache (formerly the flat-cache) sometimes stores
  coercions built from Given inerts. When these inerts get kicked out, we must
  kick out from the cache as well. (This was, I believe, true previously, but
  somehow never caused trouble.) Kicking out from the cache requires adding a
  filterTM function to TrieMap.

* This patch obviates the need to distinguish "blocking" coercion holes from
  non-blocking ones (which, previously, arose from CFunEqCans). There is thus
  some simplification around coercion holes.

* Extra commentary throughout parts of the code I read through, to preserve
  the knowledge I gained while working.

* A change in the pure unifier around unifying skolems with other types.
  Unifying a skolem now leads to SurelyApart, not MaybeApart, as documented
  in Note [Binding when looking up instances] in GHC.Core.InstEnv.

* Some more use of MCoercion where appropriate.

* Previously, class-instance lookup automatically noticed that e.g. C Int was
  a "unifier" to a target [W] C (F Bool), because the F Bool was flattened to
  a variable. Now, a little more care must be taken around checking for
  unifying instances.

* Previously, tcSplitTyConApp_maybe would split (Eq a => a). This is silly,
  because (=>) is not a tycon in Haskell. Fixed now, but there are some
  knock-on changes in e.g. TrieMap code and in the canonicaliser.

* New function anyFreeVarsOf{Type,Co} to check whether a free variable
  satisfies a certain predicate.

* Type synonyms now remember whether or not they are "forgetful"; a forgetful
  synonym drops at least one argument. This is useful when flattening; see
  flattenView.

* The pattern-match completeness checker invokes the solver. This invocation
  might need to look through newtypes when checking representational equality.
  Thus, the desugarer needs to keep track of the in-scope variables to know
  what newtype constructors are in scope. I bet this bug was around before but
  never noticed.

* Extra-constraints wildcards are no longer simplified before printing.
  See Note [Do not simplify ConstraintHoles] in GHC.Tc.Solver.

* Whether or not there are Given equalities has become slightly subtler.
  See the new HasGivenEqs datatype.

* Note [Type variable cycles in Givens] in GHC.Tc.Solver.Canonical
  explains a significant new wrinkle in the new approach.

* See Note [What might match later?] in GHC.Tc.Solver.Interact, which
  explains the fix to #18910.

* The inert_count field of InertCans wasn't actually used, so I removed
  it.

Though I (Richard) did the implementation, Simon PJ was very involved
in design and review.

This updates the Haddock submodule to avoid #18932 by adding
a type signature.

-------------------------
Metric Decrease:
    T12227
    T5030
    T9872a
    T9872b
    T9872c
Metric Increase:
    T9872d
-------------------------

This sets the stage for a later change, where this
algorithm will be needed from GHC.Core.InstEnv.

This commit also splits GHC.Core.Map into
GHC.Core.Map.Type and GHC.Core.Map.Expr,
in order to avoid module import cycles
with GHC.Core.

This will allow us to back out the allocations per compiler pass from
the eventlog. Note that we dump the allocation counter rather than the
difference since this will allow us to determine how much work is done
*between* `withTiming` blocks.

Harmonize the internal (big sum type) names of the native vs fixed-sized
number primops a bit. (Mainly by renaming the former.)

No user-facing names are changed.

Inside `regsUsedIn` we can avoid some thunks by specializing the
recursion. In particular we avoid the thunk for `(f e z)` in the
MachOp/Load branches, where we know this will evaluate to z.

Reduces allocations for T3294 by ~1%.

Co-authored-by: Sven Tennie <sven.tennie@gmail.com>
Co-authored-by: Matthew Pickering <matthewtpickering@gmail.com>
Co-authored-by: Ben Gamari <bgamari.foss@gmail.com>

Fixes #18994

Co-Author: Benjamin Maurer <maurer.benjamin@gmail.com>

[This is @Ericson2314 writing a commit message for @hsyl20's patch.]

(Progress towards #11953, #17377, #17375)

`Int64Rep` and `Word64Rep` are currently broken on 64-bit systems.  This
is because they should use "native arg rep" but instead use "large arg
rep" as they do on 32-bit systems, which is either a non-concept or a
128-bit rep depending on one's vantage point.

Now, these reps currently aren't used during 64-bit compilation, so the
brokenness isn't observed, but I don't think that constitutes reasons
not to fix it. Firstly, the linked issues there is a clearly expressed
desire to use explicit-bitwidth constructs in more places. Secondly, per
[1], there are other bugs that *do* manifest from not threading
explicit-bitwidth information all the way through the compilation
pipeline. One can therefore view this as one piece of the larger effort
to do that, improve ergnomics, and squash remaining bugs.

Also, this is needed for !3658. I could just merge this as part of that,
but I'm keen on merging fixes "as they are ready" so the fixes that
aren't ready are isolated and easier to debug.

[1]: https://mail.haskell.org/pipermail/ghc-devs/2020-October/019332.html

This replaces all Word<N> = W<N># Word# and Int<N> = I<N># Int#  with
Word<N> = W<N># Word<N># and Int<N> = I<N># Int<N>#, thus providing us
with properly sized primitives in the codegenerator instead of pretending
they are all full machine words.

This came up when implementing darwinpcs for arm64.  The darwinpcs reqires
us to pack function argugments in excess of registers on the stack.  While
most procedure call standards (pcs) assume arguments are just passed in
8 byte slots; and thus the caller does not know the exact signature to make
the call, darwinpcs requires us to adhere to the prototype, and thus have
the correct sizes.  If we specify CInt in the FFI call, it should correspond
to the C int, and not just be Word sized, when it's only half the size.

This does change the expected output of T16402 but the new result is no
less correct as it eliminates the narrowing (instead of the `and` as was
previously done).

Bumps the array, bytestring, text, and binary submodules.
Co-Authored-By: Ben Gamari <ben@well-typed.com>

Metric Increase:
    T13701
    T14697

Otherwise `opt` fails with:

    error: use of undefined value '@memcmp$def'

To dump output of the C backend.

Previously we failed to apply the info table offset to the aranges and
DIEs, meaning that we often failed to unwind in gdb. For some reason
this only seemed to manifest in the RTS's Cmm closures. Nevertheless,
now we can unwind completely up to `main`

Some plugins can be added via TH (cf addCorePlugin). Initialize them in
the driver instead of in the Core2Core pipeline.

Loaded plugins have nothing to do in DynFlags so this patch moves them
into HscEnv (session state).

"DynFlags plugins" become "Driver plugins" to still be able to register
static plugins.

Bump haddock submodule

This introducing a new compiler flag to provide a convenient way to
introduce profiler cost-centers on all occurrences of the named
identifier.

Closes #18566.

As outlined in #18903, interleaving usage and strictness demands not
only means a more compact demand representation, but also allows us to
express demands that we weren't easily able to express before.

Call demands are *relative* in the sense that a call demand `Cn(cd)`
on `g` says "`g` is called `n` times. *Whenever `g` is called*, the
result is used according to `cd`". Example from #18903:

```hs
h :: Int -> Int
h m =
  let g :: Int -> (Int,Int)
      g 1 = (m, 0)
      g n = (2 * n, 2 `div` n)
      {-# NOINLINE g #-}
  in case m of
    1 -> 0
    2 -> snd (g m)
    _ -> uncurry (+) (g m)
```

Without the interleaved representation, we would just get `L` for the
strictness demand on `g`. Now we are able to express that whenever
`g` is called, its second component is used strictly in denoting `g`
by `1C1(P(1P(U),SP(U)))`. This would allow Nested CPR to unbox the
division, for example.

Fixes #18903.
While fixing regressions, I also discovered and fixed #18957.

Metric Decrease:
    T13253-spj

Their strictness signatures said the primops are strict in their first
argument, which is wrong: Handing it a thunk will prefetch the pointer
to the thunk, but not evaluate it. Hence not strict.

The regression test `T8256` actually tests for laziness in the first
argument, so GHC apparently never exploited the strictness signature.

See also #8256 (comment 310867),
where this came up.

In order to avoid confusion as in #18932, we display the type of the
match variables in the non-exhaustiveness warning, e.g.

```
T18932.hs:14:1: warning: [-Wincomplete-patterns]
    Pattern match(es) are non-exhaustive
    In an equation for ‘g’:
        Patterns of type  ‘T a’, ‘T a’, ‘T a’ not matched:
            (MkT2 _) (MkT1 _) (MkT1 _)
            (MkT2 _) (MkT1 _) (MkT2 _)
            (MkT2 _) (MkT2 _) (MkT1 _)
            (MkT2 _) (MkT2 _) (MkT2 _)
            ...
   |
14 | g (MkT1 x) (MkT1 _) (MkT1 _) = x
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
```

It also allows us to omit the type signature on wildcard matches which
we previously showed in only some situations, particularly
`-XEmptyCase`.

Fixes #18932.

This reuses the codegen used for ByteArray#'s atomic primops.

This addes the necessary logic to support aarch64 on elf, as well
as aarch64 on mach-o, which Apple calls arm64.

We change architecture name to AArch64, which is the official arm
naming scheme.

The use of `tcSplitForAllTyVars` in `tcDataFamInstHeader` was the immediate
cause of #18939, and replacing it with a new `tcSplitForAllInvisTyVars`
function (which behaves like `tcSplitForAllTyVars` but only splits invisible
type variables) fixes the issue. However, this led me to realize that _most_
uses of `tcSplitForAllTyVars` in GHC really ought to be
`tcSplitForAllInvisTyVars` instead. While I was in town, I opted to replace
most uses of `tcSplitForAllTys` with `tcSplitForAllTysInvis` to reduce the
likelihood of such bugs in the future.

I say "most uses" above since there is one notable place where we _do_ want
to use `tcSplitForAllTyVars`: in `GHC.Tc.Validity.forAllTyErr`, which produces
the "`Illegal polymorphic type`" error message if you try to use a higher-rank
`forall` without having `RankNTypes` enabled. Here, we really do want to split
all `forall`s, not just invisible ones, or we run the risk of giving an
inaccurate error message in the newly added `T18939_Fail` test case.

I debated at some length whether I wanted to name the new function
`tcSplitForAllInvisTyVars` or `tcSplitForAllTyVarsInvisible`, but in the end,
I decided that I liked the former better. For consistency's sake, I opted to
rename the existing `splitPiTysInvisible` and `splitPiTysInvisibleN` functions
to `splitInvisPiTys` and `splitPiTysInvisN`, respectively, so that they use the
same naming convention. As a consequence, this ended up requiring a `haddock`
submodule bump.

Fixes #18939.

There is a zoo of `splitForAll-` functions in `GHC.Core.Type` (as well as
`tcSplitForAll-` functions in `GHC.Tc.Utils.TcType`) that all do very similar
things, but vary in the particular form of type variable that they return. To
make things worse, the names of these functions are often quite misleading.
Some particularly egregious examples:

* `splitForAllTys` returns `TyCoVar`s, but `splitSomeForAllTys` returns
  `VarBndr`s.
* `splitSomeForAllTys` returns `VarBndr`s, but `tcSplitSomeForAllTys` returns
  `TyVar`s.
* `splitForAllTys` returns `TyCoVar`s, but `splitForAllTysInvis` returns
  `InvisTVBinder`s. (This in particular arose in the context of #18939, and
  this finally motivated me to bite the bullet and improve the status quo
  vis-à-vis how we name these functions.)

In an attempt to bring some sanity to how these functions are named, I have
opted to rename most of these functions en masse to use consistent suffixes
that describe the particular form of type variable that each function returns.
In concrete terms, this amounts to:

* Functions that return a `TyVar` now use the suffix `-TyVar`.
  This caused the following functions to be renamed:
  * `splitTyVarForAllTys` -> `splitForAllTyVars`
  * `splitForAllTy_ty_maybe` -> `splitForAllTyVar_maybe`
  * `tcSplitForAllTys` -> `tcSplitForAllTyVars`
  * `tcSplitSomeForAllTys` -> `tcSplitSomeForAllTyVars`
* Functions that return a `CoVar` now use the suffix `-CoVar`.
  This caused the following functions to be renamed:
  * `splitForAllTy_co_maybe` -> `splitForAllCoVar_maybe`
* Functions that return a `TyCoVar` now use the suffix `-TyCoVar`.
  This caused the following functions to be renamed:
  * `splitForAllTy` -> `splitForAllTyCoVar`
  * `splitForAllTys` -> `splitForAllTyCoVars`
  * `splitForAllTys'` -> `splitForAllTyCoVars'`
  * `splitForAllTy_maybe` -> `splitForAllTyCoVar_maybe`
* Functions that return a `VarBndr` now use the suffix corresponding to the
  most relevant type synonym. This caused the following functions to be renamed:
  * `splitForAllVarBndrs` -> `splitForAllTyCoVarBinders`
  * `splitForAllTysInvis` -> `splitForAllInvisTVBinders`
  * `splitForAllTysReq` -> `splitForAllReqTVBinders`
  * `splitSomeForAllTys` -> `splitSomeForAllTyCoVarBndrs`
  * `tcSplitForAllVarBndrs` -> `tcSplitForAllTyVarBinders`
  * `tcSplitForAllTysInvis` -> `tcSplitForAllInvisTVBinders`
  * `tcSplitForAllTysReq` -> `tcSplitForAllReqTVBinders`
  * `tcSplitForAllTy_maybe` -> `tcSplitForAllTyVarBinder_maybe`

Note that I left the following functions alone:

* Functions that split apart things besides `ForAllTy`s, such as `splitFunTys`
  or `splitPiTys`. Thankfully, there are far fewer of these functions than
  there are functions that split apart `ForAllTy`s, so there isn't much of a
  pressing need to apply the new naming convention elsewhere.
* Functions that split apart `ForAllCo`s in `Coercion`s, such as
  `GHC.Core.Coercion.splitForAllCo_maybe`. We could theoretically apply the new
  naming convention here, but then we'd have to figure out how to disambiguate
  `Type`-splitting functions from `Coercion`-splitting functions. Ultimately,
  the `Coercion`-splitting functions aren't used nearly as much as the
  `Type`-splitting functions, so I decided to leave the former alone.

This is purely refactoring and should cause no change in behavior.

Standard debugging tools don't know how to understand these so let's not
produce them unless asked.

Previously the `.debug_aranges` and `.debug_info` (DIE) DWARF
information would claim that procedures (represented with a
`DW_TAG_subprogram` DIE) would only span the range covered by their entry
block. This omitted all of the continuation blocks (represented by
`DW_TAG_lexical_block` DIEs), confusing `perf`. Fix this by introducing
a end-of-procedure label and using this as the `DW_AT_high_pc` of
procedure `DW_TAG_subprogram` DIEs

Fixes #17605.

See Note [Exciting arity] why we emit the warning at all and why we only
do after the second iteration now.

Fixes #18937.

As we found out in #18870, `andArityType` is not monotone, with
potentially severe consequences for termination of fixed-point
iteration. That showed in an abundance of "Exciting arity" DEBUG
messages that are emitted whenever we do more than one step in
fixed-point iteration.

The solution necessitates also recording `OneShotInfo` info for
`ABot` arity type. Thus we get the following definition for `ArityType`:

```
data ArityType = AT [OneShotInfo] Divergence
```

The majority of changes in this patch are the result of refactoring use
sites of `ArityType` to match the new definition.

The regression test `T18870` asserts that we indeed don't emit any DEBUG
output anymore for a function where we previously would have.
Similarly, there's a regression test `T18937` for #18937, which we
expect to be broken for now.

Fixes #18870.

In ticket #18733 we noticed a rather serious deficiency in the current
fingerprinting logic for recursive groups. I have described the old
fingerprinting story and its problems in Note [Fingerprinting recursive
groups] and have reworked the story accordingly to avoid these issues.

Fixes #18733.

Fixes #16525 by tracking dependencies between object file symbols and
marking symbol liveness during garbage collection

See Note [Object unloading] in CheckUnload.c for details.

This seems like a reasonable default as the object file size increases
by around 5%.

It turns out that some important native debugging/profiling tools (e.g.
perf) rely only on symbol tables for function name resolution (as
opposed to using DWARF DIEs). However, previously GHC would emit
temporary symbols (e.g. `.La42b`) to identify module-internal
entities. Such symbols are dropped during linking and therefore not
visible to runtime tools (in addition to having rather un-helpful unique
names). For instance, `perf report` would often end up attributing all
cost to the libc `frame_dummy` symbol since Haskell code was no covered
by any proper symbol (see #17605).

We now rather follow the model of C compilers and emit
descriptively-named local symbols for module internal things. Since this
will increase object file size this behavior can be disabled with the
`-fno-expose-internal-symbols` flag.

With this `perf record` can finally be used against Haskell executables.
Even more, with `-g3` `perf annotate` provides inline source code.

In various places in the NCG we need the Module currently being
compiled. Let's move this into the environment instead of chewing threw
another register.

It appears this was an oversight as there is no reason the full DynFlags
is necessary.

1) Don't modify DynFlags (too much) for -dynamic-too: now when we
   generate dynamic outputs for "-dynamic-too", we only set "dynamicNow"
   boolean field in DynFlags instead of modifying several other fields.
   These fields now have accessors that take dynamicNow into account.

2) Use DynamicTooState ADT to represent -dynamic-too state. It's much
   clearer than the undocumented "DynamicTooConditional" that was used
   before.

As a result, we can finally remove the hscs_iface_dflags field in
HscRecomp. There was a comment on this field saying:

   "FIXME (osa): I don't understand why this is necessary, but I spent
   almost two days trying to figure this out and I couldn't .. perhaps
   someone who understands this code better will remove this later."

I don't fully understand the details, but it was needed because of the
changes made to the DynFlags for -dynamic-too.

There is still something very dubious in GHC.Iface.Recomp: we have to
disable the "dynamicNow" flag at some point for some Backpack's "heinous
hack" to continue to work. It may be because interfaces for indefinite
units are always non-dynamic, or because we mix and match dynamic and
non-dynamic interfaces (#9176), or something else, who knows?

This refactors the GHC AST to remove `HsImplicitBndrs` and replace it with
`HsOuterTyVarBndrs`, a type which records whether the outermost quantification
in a type is explicit (i.e., with an outermost, invisible `forall`) or
implicit. As a result of this refactoring, it is now evident in the AST where
the `forall`-or-nothing rule applies: it's all the places that use
`HsOuterTyVarBndrs`. See the revamped `Note [forall-or-nothing rule]` in
`GHC.Hs.Type` (previously in `GHC.Rename.HsType`).

Moreover, the places where `ScopedTypeVariables` brings lexically scoped type
variables into scope are a subset of the places that adhere to the
`forall`-or-nothing rule, so this also makes places that interact with
`ScopedTypeVariables` easier to find. See the revamped
`Note [Lexically scoped type variables]` in `GHC.Hs.Type` (previously in
`GHC.Tc.Gen.Sig`).

`HsOuterTyVarBndrs` are used in type signatures (see `HsOuterSigTyVarBndrs`)
and type family equations (see `HsOuterFamEqnTyVarBndrs`). The main difference
between the former and the latter is that the former cares about specificity
but the latter does not.

There are a number of knock-on consequences:

* There is now a dedicated `HsSigType` type, which is the combination of
  `HsOuterSigTyVarBndrs` and `HsType`. `LHsSigType` is now an alias for an
  `XRec` of `HsSigType`.
* Working out the details led us to a substantial refactoring of
  the handling of explicit (user-written) and implicit type-variable
  bindings in `GHC.Tc.Gen.HsType`.

  Instead of a confusing family of higher order functions, we now
  have a local data type, `SkolemInfo`, that controls how these
  binders are kind-checked.

  It remains very fiddly, not fully satisfying. But it's better
  than it was.

Fixes #16762. Bumps the Haddock submodule.
Co-authored-by: Simon Peyton Jones <simonpj@microsoft.com>
Co-authored-by: Richard Eisenberg <rae@richarde.dev>
Co-authored-by: Zubin Duggal <zubin@cmi.ac.in>