-
ce616f49
by Simon Peyton Jones at 2025-04-27T21:10:25+01:00
Fix infelicities in the Specialiser
On the way to #23109 (unary classes) I discovered some infelicities
(or maybe tiny bugs, I forget) in the type-class specialiser.
I also tripped over #25965, an outright bug in the rule matcher
Specifically:
* Refactor: I enhanced `wantCallsFor`, whih previously always said
`True`, to discard calls of class-ops, data constructors etc. This is
a bit more efficient; and it means we don't need to worry about
filtering them out later.
* Fix: I tidied up some tricky logic that eliminated redundant
specialisations. It wasn't working correctly. See the expanded
Note [Specialisations already covered], and
(MP3) in Note [Specialising polymorphic dictionaries].
See also the new top-level `alreadyCovered`
function, which now goes via `GHC.Core.Rules.ruleLhsIsMoreSpecific`
I also added a useful Note [The (CI-KEY) invariant]
* Fix #25965: fixed a tricky bug in the `go_fam_fam` in
`GHC.Core.Unify.uVarOrFam`, which allows matching to succeed
without binding all type varibles.
I enhanced Note [Apartness and type families] some more
* #25703. This ticket "just works" with -fpolymorphic-specialisation;
but I was surprised that it worked! In this MR I added documentation
to Note [Interesting dictionary arguments] to explain; and tests to
ensure it stays fixed.
-
22d11fa8
by Simon Peyton Jones at 2025-04-28T18:05:19-04:00
Track rewriter sets more accurately in constraint solving
The key change, which fixed #25440, is to call `recordRewriter` in
GHC.Tc.Solver.Rewrite.rewrite_exact_fam_app. This missing call meant
that we were secretly rewriting a Wanted with a Wanted, but not really
noticing; and that led to a very bad error message, as you can see
in the ticket.
But of course that led me into rabbit hole of other refactoring around
the RewriteSet code:
* Improve Notes [Wanteds rewrite Wanteds]
* Zonk the RewriterSet in `zonkCtEvidence` rather than only in GHC.Tc.Errors.
This is tidier anyway (e.g. de-clutters debug output), and helps with the
next point.
* In GHC.Tc.Solver.Equality.inertsCanDischarge, don't replace a constraint
with no rewriters with an equal constraint that has many. See
See (CE4) in Note [Combining equalities]
* Move zonkRewriterSet and friends from GHC.Tc.Zonk.Type into
GHC.Tc.Zonk.TcType, where they properly belong.
A handful of tests get better error messages.
For some reason T24984 gets 12% less compiler allocation -- good
Metric Decrease:
T24984
-
6467d61e
by Brandon Chinn at 2025-04-29T18:36:03-04:00
Fix lexing "\^\" (#25937)
This broke in the refactor in !13128, where the old code parsed escape
codes and collapsed string gaps at the same time, but the new code
collapsed gaps first, then resolved escape codes. The new code used a
naive heuristic to skip escaped backslashes, but didn't account for
"\^\".
-
99868a86
by Jens Petersen at 2025-04-29T18:36:44-04:00
hadrian: default selftest to disabled
-
aba2a4a5
by Zubin Duggal at 2025-04-30T06:35:59-04:00
get-win32-tarballs.py: List tarball files to be downloaded if we cannot find them
Fixes #25929
-
d99a617b
by Ben Gamari at 2025-04-30T06:36:40-04:00
Move Data ModuleName instance to Language.Haskell.Syntax.Module.Name
Fixes #25968.
-
9995c2b7
by Serge S. Gulin at 2025-05-04T17:13:36+03:00
Support for ARM64 Windows (LLVM-enabled) (fixes #24603)
1. Add Windows AArch64 cross-compilation support via CI jobs
Introduce new CI configurations for cross-compiling to Windows ARM64 using Debian12Wine, FEX, and MSYS2.
Configure toolchain variables for LLVM MinGW and Wine emulation in CI pipelines.
2. Adjust compiler and RTS for AArch64 Windows compatibility
Reserve register `x18` on Windows and Darwin platforms in AArch64 codegen.
Handle Windows-specific relocations and data sections in AArch64 assembler.
Update PEi386 linker to recognize ARM64 binaries and support exception handling.
Adjust LLVM target definitions and data layouts for new architectures.
Update `ghc-toolchain` and build scripts to handle `TablesNextToCode` on Windows ARM64.
3. Enhance CI scripts and stability
Modify `ci.sh` to handle mingw cross-targets, fixing GHC executable paths and test execution.
Use `diff -w` in tests to ignore whitespace differences, improving cross-platform consistency.
4. Refactor and clean up code
Remove redundant imports in hello.hs test.
Improve error messages and checks for unsupported configurations in the driver.
Add `EXDEV` error code to `errno.js`.
Add async/sync flags to IO logs at `base.js`.
Improve POSIX compatibility for file close at `base.js`: decrease indeterminism for mixed cases of async and sync code.
5. Update dependencies: `Cabal`, `Win32`, `directory`, `process`, `haskeline`, and `unix`.
submodule
Co-authored-by: Cheng Shao <terrorjack@type.dance>
Co-authored-by: Dmitrii Egorov <egorov.d.i@icloud.com>
Co-authored-by: Andrei Borzenkov <root@sandwitch.dev>
-
50fa8165
by Javran Cheng at 2025-05-05T05:55:39-04:00
Suppress unused do-binding if discarded variable is Any or ZonkAny.
Consider example (#25895):
> do { forever (return ()); blah }
where `forever :: forall a b. IO a -> IO b`.
Nothing constrains `b`, so it will be instantiates with `Any` or
`ZonkAny`.
But we certainly don't want to complain about a discarded do-binding.
Fixes #25895
-
2229b15b
by Andreas Klebinger at 2025-05-06T13:59:44+02:00
Interpreter: Add limited support for direct primop evaluation.
This commit adds support for a number of primops directly
to the interpreter. This avoids the indirection of going
through the primop wrapper for those primops speeding interpretation
of optimized code up massively.
Code involving IntSet runs about 25% faster with optimized core and these
changes. For core without breakpoints it's even more pronouced and I
saw reductions in runtime by up to 50%.
Running GHC itself in the interpreter was sped up by ~15% through this
change.
Additionally this comment does a few other related changes:
testsuite:
* Run foundation test in ghci and ghci-opt ways to test these
primops.
* Vastly expand the foundation test to cover all basic primops
by comparing result with the result of calling the wrapper.
Interpreter:
* When pushing arguments for interpreted primops extend each argument to
at least word with when pushing. This avoids some issues with big
endian. We can revisit this if it causes performance issues.
* Restructure the stack chunk check logic. There are now macros for
read accesses which might cross stack chunk boundries and macros which
omit the checks which are used when we statically know we access an
address in the current stack chunk.