GitLab

Before this patch, GHC always printed the * kind unparenthesized.

This led to two issues:

1. Sometimes GHC printed invalid or incorrect code.
   For example, GHC would print:  type F @*   x = x
         when it meant to print:  type F @(*) x = x
   In the former case, instead of a kind application we were getting a
   type operator (@*).

2. Sometimes GHC printed kinds that were correct but hard to read.
   Should  Either * Int  be read as  Either (*) Int
                              or as  (*) Either Int  ?
   This depends on whether -XStarIsType is enabled, but it would be
   easier if we didn't have to check for the flag when reading the code.

We can solve both problems by assigning (*) a different precedence. Note
that Haskell98 kinds are not affected:

  ((* -> *) -> *) -> *  does NOT become  (((*) -> (*)) -> (*)) -> (*)

The parentheses are added when (*) is used in a function argument
position:

   F * * *   becomes  F (*) (*) (*)
   F A * B   becomes  F A (*) B
   Proxy *   becomes  Proxy (*)
   a * -> *  becomes  a (*) -> *

Along with some refactoring.

The LLVM backend does not guarantee any particular semantics for
division by zero, making this test unreliable across platforms.

Previously we would clear the bitmaps of segments which we are going to
sweep during the preparatory pause. However, this is unnecessary: the
existence of the mark epoch ensures that the sweep will correctly
identify non-reachable objects, even if we do not clear the bitmap.

We now defer clearing the bitmap to sweep, which happens concurrently
with mutation.

We use two tricks, as suggested here [1]:

 * Use microtype to try to reduce the incidence of underfull boxes
 * Bump up \hbadness to eliminate the warnings

This patch implements GHC Proposal #176:
  https://github.com/ghc-proposals/ghc-proposals/blob/master/proposals/0176-scc-parsing.rst

Before the change:

  1 /                    2 / 2 = 0.25
  1 / {-# SCC "name" #-} 2 / 2 = 1.0

After the change:

  1 /                    2 / 2 = 0.25
  1 / {-# SCC "name" #-} 2 / 2 = parse error

Previously we would use the `.int` assembler directive to generate
32-bit words in the note section. However, `.int` is note guaranteed to
produce 4-bytes; in fact, on some platforms (e.g. AArch64) it produces
8-bytes. Use the `.4bytes` directive to avoid this.

Moreover, we used the `.align` directive, which is quite platform
dependent. On AArch64 it appears to not even be idempotent (despite what
the documentation claims). `.balign` is consequentially preferred as it
offers consistent behavior across platforms.

This does two things:

 * Eliminate all uses of Unique.deriveUnique, which was quite easy to
   mis-use and extremely subtle.

 * Rename the previous "derived unique" notion to "local unique". This
   is possible because the only places where `uniqAway` can be safely
   used are those where local uniqueness (with respect to some
   InScopeSet) is sufficient.

 * Rework the implementation of VarEnv.uniqAway, as discussed in #17462.
   This should make the operation significantly more efficient than its
   previous iterative implementation..

Metric Decrease:
    T9872c
    T12227
    T9233
    T14683
    T5030
    T12545
    hie002

Metric Increase:
    T9961

This relied on deriveUnique, which was far too subtle to be safely
applied. Thankfully the instance doesn't appear to be used so let's just
drop it.

    narrowN (x .&. m)
    m .&. (2^N-1) = 2^N-1
    ==> narrowN x

e.g.  narrow16 (x .&. 0x12FFFF) ==> narrow16 x

In #17424 Simon PJ noted that there is a potentially unsafe occurrence
of unsafeCoerce#, coercing from an unlifted to lifted type. However,
nowhere in the compiler do we assume that a BCO# is not a thunk.
Moreover, in the case of a CAF the result returned by `createBCO` *will*
be a thunk (as noted in [Updatable CAF BCOs]).  Consequently it seems
better to rather make BCO# a lifted type and rename it to BCO.

Be more precise at detecting cross-compilation case.

Before the change configuration
  $ ./configure --host=x86_64-pc-linux-gnu --target=x86_64-gentoo-linux-musl
was not considered a cross-target. Even though libcs are different (`glibc` vs. `musl`).

Without this patch build fails as:

```
"inplace/bin/ghc-cabal" check libraries/integer-gmp
"inplace/bin/ghc-cabal" configure libraries/integer-gmp dist-install \
    --with-ghc="/home/slyfox/dev/git/ghc/inplace/bin/ghc-stage1" \
    --with-ghc-pkg="/home/slyfox/dev/git/ghc/inplace/bin/ghc-pkg" \
    --disable-library-for-ghci --enable-library-vanilla --enable-library-for-ghci \
    --enable-library-profiling --enable-shared --with-hscolour="/usr/bin/HsColour" \
    --configure-option=CFLAGS="-Wall \
    -Werror=unused-but-set-variable -Wno-error=inline \
    -iquote /home/slyfox/dev/git/ghc/libraries/integer-gmp" \
    --configure-option=LDFLAGS="  " --configure-option=CPPFLAGS="   \
    " --gcc-options="-Wall     -Werror=unused-but-set-variable -Wno-error=inline -iquote /home/slyfox/dev/git/ghc/libraries/integer-gmp   \
    " --with-gcc="x86_64-gentoo-linux-musl-gcc" --with-ld="x86_64-gentoo-linux-musl-ld.gold" --with-ar="x86_64-gentoo-linux-musl-ar" \
    --with-alex="/usr/bin/alex" --with-happy="/usr/bin/happy"
Configuring integer-gmp-1.0.2.0...
configure: WARNING: unrecognized options: --with-compiler
checking build system type... x86_64-pc-linux-gnu
checking host system type... x86_64-pc-linux-gnu
checking target system type... x86_64-pc-linux-gnu
checking for gcc... /usr/lib/ccache/bin/x86_64-gentoo-linux-musl-gcc
checking whether the C compiler works... yes
checking for C compiler default output file name... a.out
checking for suffix of executables...
checking whether we are cross compiling... configure: error: in `/home/slyfox/dev/git/ghc/libraries/integer-gmp/dist-install/build':
configure: error: cannot run C compiled programs.
If you meant to cross compile, use `--host'.
See `config.log' for more details
make[1]: *** [libraries/integer-gmp/ghc.mk:5: libraries/integer-gmp/dist-install/package-data.mk] Error 1
make: *** [Makefile:126: all] Error 2
```

Note: here `ghc-stage1` is assumed to target `musl` target but is passed
`glibc` toolchain. It happens because initial ./configure phase did not
detect host/target as different.
Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>

(cherry picked from commit 67b5de88ef923971f1980335137e3c7193213abd)

(cherry picked from commit cbedb3c4a90649f474cb716842ba53afc5a642ca)

(cherry picked from commit d092d859)

Apparently variable interpolation in the `artifacts.paths` key of
`gitlab-ci.yml` doesn't work on Windows as it does on WIndows.

(cherry picked from commit 100cc756)

It otherwise seems to default to ascii

Fall back to xz

Installs pxz on Centos7

(cherry picked from commit 86960e69)

Previously we were including <sys/unistd.h> which is available on glibc
but not musl.

(cherry picked from commit e44b695c)

The docutils version available appears to be too old to support the
`table` directive's `:widths:` options.

(cherry picked from commit 75764487)

The `process` changes have now been merged into `hsc2hs`.

(cherry picked from commit fa029f53)

A colleague recently hit the panic in Simplify.addEvals and I noticed
that the message is quite unreadable due to incorrect pretty-printing.
Fix this.

As suggested by RyanGlScott in !2163.

The code fragment

  type family Proxy2' ∷ ∀ k → k → Type where
    Proxy2' = Proxy'

Generates AnnRarrow instead of AnnRarrowU for the first →.

Fixes #17519

Since the invariants always hold in the testsuite, we can convert
them to asserts.