This commit bumps ci-image revision to use updated wasm toolchain, and
use host_fully_static instead of fully_static for wasm jobs so to
ensure wasm shared libraries can be properly built.

This commit adds the host_fully_static flavour transformer to hadrian,
which ensures stage0 is fully statically linked while still permitting
stage1 libdir to contain shared libraries. This is intended to be used
by the wasm backend to build portable linux bindists that contain wasm
shared libraries.

This commit re-enables --with-pic=yes configuration option of gmp when
building for wasm, given we're about to include support for shared
libraries, TH and ghci.

This commit adds -fvisibility=default to CFLAGS of gmp when building
for wasm. This is required to generate the ghc-bignum shared library
without linking errors. Clang defaults to -fvisibility=hidden for wasm
targets, which will cause issues when a symbol is expected to be
exported in a shared library but without explicit visibility attribute
annotation.

This commit enables internal-interpreter flag for ghc-bin even when
compiling stage0, as long as target supports ghci. It enables ghci
functionality for cross targets that support ghci, since cross ghc-bin
is really stage0.

This commit disable the internal-interpreter flag of ghc library when
cross compiling, only external interpreter works in such cases.

This commit changes the hostSupportsRPaths predicate to
targetSupportsRPaths and use that to decide whether to pass
RPATH-related link-time options. It's not applied to stage0, we should
just use the default link-time options of stageBoot ghc.

This commit fixes conflicting StgRun definition when building dynamic
ways of RTS for wasm in unregisterised mode.

This commit wraps the predefined GlobalRegs in Wasm.S under a CPP
guard to prevent building for PIC mode. When building dynamic ways of
RTS, the wasm globals that represent STG GlobalRegs will be created
and supplied by dyld.mjs. The current wasm dylink convention doesn't
properly support exporting relocatable wasm globals at all, any wasm
global exported by a .so is assumed to be a GOT.mem entry.

This commit drops interpretBCO support from non dynamic rts ways on
wasm. The bytecode interpreter is only useful when the RTS linker also
works, and on wasm it only works for dynamic ways anyway. An
additional benefit of dropping interpretBCO is reduction in code size
of linked wasm modules, especially since interpretBCO references
ffi_call which is an auto-generated large function in libffi-wasm and
unused by most user applications.

This commit wraps imported freeJSVal in a __wrapped_freeJSVal C
function for wasm backend RTS. In general, wasm imports are only
supposed to be directly called by C; they shouldn't be used as
function pointers, which confuses wasm-ld at link-time when generating
shared libraries.

This commit renames prelude.js to prelude.mjs for wasm backend rts
jsbits, and slightly adjusts the jsbits contents. This is for
preparing the implementation of dyld.mjs that contains wasm dynamic
linker logic, which needs to import prelude.mjs as a proper ESM
module.

This patch adds some missing CPP guards to ghc-internal, given those
functions are non existent on wasm and would cause linking issues.

This patch avoids using mmap() to allocate executable memory for
mkConInfoTable on platforms without tables-next-to-code, see added
comment for explanation.

This patch mitigates a severe host/target word size mismatch issue in
BCOByteArray serialization logic introduced since !12142, see added
note for detailed explanation.

Alan Zimmerman authored 6 months ago and Marge Bot committed 6 months ago

EPA: Remove [AddEpAnn] from HsDocTy

EPA: Remove [AddEpAnn] from HsBangTy

EPA: Remove [AddEpAnn] from HsExplicitListTy

EPA: Remove [AddEpAnn] from HsExplicitTupleTy

EPA: Remove [AddEpAnn] from HsTypedBracket

EPA: Remove [AddEpAnn] from HsUntypedBracket

EPA: Remove [AddEpAnn] from PatBuilderOpApp

EPA: break out 'EpToken "|"' from ClassDecl anns

EPA: Remove [AddEpAnn] from ClassDecl

EPA: Remove [AddEpAnn] from SynDecl

ur4t authored 1 year ago and lang committed 6 months ago

Fixes #24266

Progresses #21592

For some reason we still imported GHC.Types.Fixity when the definitino of Fixity and LexicalFixity have already been moved to Language.Haskell.Syntax.Basic. This fixes that for

ARATA Mizuki authored 6 months ago and Marge Bot committed 6 months ago

Fixes #22033

Simon Peyton Jones authored 6 months ago and Marge Bot committed 6 months ago

In the main MR, -Wall now includes -Wincomplete-record-selectors.
However `hadrian-multi` has many, many warnings about incomplete
record selectors, so this patch stops those warnings being treated
as errors.  (See discussion on !13308.)

A better fix would be to remove the use of incomplete record
selectors, since each of them represents a potential crash.

Simon Peyton Jones authored 6 months ago and Marge Bot committed 6 months ago

This patch is a pure refactor of GHC's source code, to avoid the use
of partial record selectors.  It was provoked by adding
-Wincomplete-record-selectors to -Wall (as the GHC Proposal specified),
which in turn showed up lots of places where GHC was using incomplete
record selectors.

This patch does mostly-simple refactoring to make it clear to the pattern
match checker that there is in fact no partiality.

There is one externally-visible change: I changed the data type HoleFit
to split out the two cases

  data HoleFit = TcHoleFit  TcHoleFit | RawHoleFit SDoc
  data TcHoleFit = HoleFit { ...lots of fields }

There are large swathes of code that just deal with `TcHoleFit`, and
having it as a separate data types makes it apparent that `RawHoleFit`
can't occur.

This makes it much better -- but the change is visible in the
HolePlugin interface.  I decided that there are so few clients of this
API that it's worth the change.

I moved several functions from Language.Haskell.Syntax to GHC.Hs.
Reason, when instantiated at (GhcPass _), the extension data construtcor
is guaranteed unused, and that justifies omitted patterns in these
functions.  By putting them in GHC.Hs.X I can specialise the type for
(GhcPass _) and thereby make the function total.

An interesting side-light is that there were a few local function
definitions without a type signature, like this one in GHC.Parser.Header
     convImport (L _ i) = (ideclPkgQual i, reLoc $ ideclName i)
This is fully closed, and so is generalised; but that generalises
it to any old pass, not (GhcPass _), so GHC rightly complains about the
use of the selector `ideclPkgQual`.  I added a type signature to `i`, thus
     convImport (L _ (i::ImportDecl GhcPs))
         = (ideclPkgQual i, reLoc $ ideclName i)
which specialised the function enough to make the record selector complete.
Quite a surprising consequence of local let-generalisation!

Sebastian Graf authored 11 months ago and Marge Bot committed 6 months ago

This commit does several related things:

* Major refactor of the handling of applications in the desugarer.
  Now all applications are handled in `dsApp`, `ds_app` and related
  functions.  This dramatically simplifies the code and removes
  complicated cruft that had accumulated.  Hooray.

  Fixes #25281.

* Improve the handling of -Wincomplete-record-selectors.

  We now incorporate the result type of unsaturated record selector
  applications as well as consider long-distance information in
  getField applications.

  Plus, the implmentation now builds the improved `dsApp` stuff
  above, so it is much easier to understand.

  Plus, incorporates improved error message wording suggested
  by Adam Gundry in !12685.

  Fixes #24824, #24891

  See the long Note [Detecting incomplete record selectors]

* Add -Wincomplete-record-selectors to -Wall, as specified in
  GHC Proposal 516.

  To do this, I also had to add -Wno-incomplete-record-selectors
  to the build flags for Cabal in GHC's CI.  See
  hadrian/src/Settings/Warnings.hs.  We can remove this when
  Cabal is updated so that it doesn't trigger the warning:
  https://github.com/haskell/cabal/issues/10402

2.6% decrease in compile time allocation in RecordUpPerf

Metric Decrease:
    RecordUpdPerf

Alan Zimmerman authored 6 months ago and Marge Bot committed 6 months ago

EPA: Remove [AddEpAnn] from LazyPat

EPA: Remove [AddEpAnn] from RecordCon/RecordUpd/ConPat

EPA: Remove [AddEpAnn] from HsFieldBind

EPA: Remove [AddEpAnn] from PatSynBind

EPA: Remove [AddEpAnn] from IPBind

EPA: Remove [AddEpAnn] from FixSig

EPA: Remove [AddEpAnn] from activation rules

EPA: Remove [AddEpann] from SpecInstSig

EPA: Remove [AddEpAnn] from MinimalSig

EPA: Remove [AddEpAnn] from SCCFunSig

EPA: Remove [AddEpAnn] from CompleteMatchSig

EPA: Remove [AddEpAnn] from AnnSig, as used in PatSynSig, ClassOpSig, TypeSig

EPA: Remove [AddEpAnn] from IEThingAbs

EPA: Remove [AddEpAnn] from IEThingAll / IEThingWith

EPA: Remove [AddEpAnn] from IEModuleContents

EPA: Remove [AddEpAnn] from HsOpTy

EPA: Remove [AddEpAnn] for various binders

EPA: Remove [AddEpAnn] for HsIParamTy

Ben Gamari authored 6 months ago and Marge Bot committed 6 months ago

Resolve #24509.

Zubin authored 6 months ago and Zubin committed 6 months ago

Metric Decrease:
    MultiLayerModulesTH_OneShot
Metric Increase:
    haddock.Cabal

Zubin authored 6 months ago and Zubin committed 6 months ago

Cabal 3.14 introduced a broken symlink in its testsuite. Unfortunately,
this broke our source distribution as we use use `tar --dereference` to avoid
issues with symlink compatibility on windows, and `tar --dereference` chokes
when it encounters any broken symlinks.

We can't get rid of `--dereference` because symlinks are generally broken on
windows, so the only option is to exclude this file from source archives.

see also https://github.com/haskell/cabal/issues/10442

Zubin authored 6 months ago and Zubin committed 6 months ago

If we have a broken symlink in the repository, don't try to `need` the symlink
or the target of the symlink. Attempting to do so has `shake` attempt to read the
target to compute its hash, which fails because the target doesn't exist.

Sven Tennie authored 6 months ago and Marge Bot committed 6 months ago

There are macros in MachRegs.h to figure those out.

Sylvain Henry authored 6 months ago and Marge Bot committed 6 months ago

Tuples with prefix/suffix strings weren't always properly encoded with
their shortcut notations. Fix this.

Ben Gamari authored 6 months ago and Marge Bot committed 6 months ago

Closes #25296.

Artem Pelenitsyn authored 6 months ago and Marge Bot committed 6 months ago

Also, fix a bug in RST with missing newline before a listing.

Co-authored-by: Arnaud Spiwack <arnaud@spiwack.net>

Hassan Al-Awwadi authored 6 months ago and Marge Bot committed 6 months ago

These constructors were only used during the TC stage,
or during template haskell. It seemed clear that it was
independent of the source syntax represented in L.H.S,
and thus we removed it according to #21592.

Ben Gamari authored 7 months ago and Marge Bot committed 6 months ago

As noted in #25066, the exception backtrace proposal introduced a rather
subtle performance regression due to simplification producing Core which
the demand analyser concludes may diverge with a precise exception. The
nature of the problem is more completely described in the new Note
[Hiding precise exception signature in throw].

The (rather hacky) solution we use here hides the problematic
optimisation through judicious use of `noinline`. Ultimately however we
will want a more principled solution (e.g. #23847).

Fixes #255066

CLC proposal: https://github.com/haskell/core-libraries-committee/issues/290

Metric Decrease:
    T9872d