GitLab

Alan Zimmerman pushed to branch wip/az/exactprint-annotation-rationalisation at Glasgow Haskell Compiler / GHC

Commits:

93a2b20f

by Andreas Klebinger at 2026-08-18T04:31:04-04:00

Fix a number of incorrect module references:

Fix module reference in Note [DataCon wrappers are conlike].
Fix module reference in Note [Detailed InertCans Invariants].
Fix module reference in Note [GHC's data format representations].
Fix module reference in Note [Grand plan for static forms].
Fix module reference in Note [How tuples work].
Fix module reference in Note [Solved dictionaries].
Fix module reference in Note [Stacktraces from Info Table Provenance Entries (IPE based stack unwinding)].
Fix module reference in Note [The VarBndr type and its uses].

Found the incorrect references with a llm.

eb0dfb01

by Simon Jakobi at 2026-08-18T04:31:44-04:00

ci: Run stack-hadrian-build only in full-ci pipelines

The job exists to catch changes that break hadrian/build-stack (#18726),
but nothing in the pipeline depends on it, and it can only break when
hadrian's dependencies change. Restricting it to full-ci (like
hadrian-multi) still covers marge-bot merge batches, so such breakage
cannot reach master unnoticed, while ordinary validate pipelines skip
the job.

Assisted-by: Claude Fable 5

b9160962

by Simon Jakobi at 2026-08-20T14:57:52-04:00

testsuite: Show baseline sample count and range in perf failures

A perf baseline is the mean of all samples recorded for a commit, and
it prints as a single number, hiding how far the samples spread. When
the spread is wide, this can indicate an unstable metric that isn't
actually useful as a signal for the perf tests.
For example, in #27602, T27336's peak_megabytes_allocated baseline
showed as 757 when the underlying samples were 605 and 909.

When the baseline is averaged from more than one sample, say so in the
failure output: the one-line stat-failure reason shows the sample
range, and the detail block lists the raw samples. Single-sample
baselines print exactly as before.

Context: #27602

Assisted-by: Claude Fable 5

a4979877

by Simon Jakobi at 2026-08-20T14:57:52-04:00

testsuite: Fold Baseline into CommitMetric

A Baseline was just a CommitMetric plus the commit it came from, built
by copying fields across. Since get_commit_metric already knows that
commit, record it on CommitMetric itself and drop Baseline. This also
collapses both branches of find_baseline into plain returns.

Assisted-by: Claude Fable 5

99fb8d68

by Simon Jakobi at 2026-08-20T14:57:52-04:00

ci: Clarify comment on pushing perf notes after failures

Context: #27602

Assisted-by: Claude Fable 5

2ca87972

by Alan Zimmerman at 2026-08-20T14:58:36-04:00

EPA: Remove LocatedBC / SrcSpanBF

The custom annotations are now in the BooleanFormula TTG extension
points, so LBooleanFormula can now use the standard LocatedA.

d2bc32aa

by Simon Peyton Jones at 2026-08-21T12:59:26-04:00

Better handling of serialisation of wired-in names

Fixes #27501

d2795ffc

by Alan Zimmerman at 2026-08-21T13:00:05-04:00

EPA: Remove NoEpTok/NoEpUniTok, using an unhelpful SrcSpan instead

Also introduce helper functions noEpTok and noEpUniTok to serve
as simple replacements in code inserting an token annotation without
location information.

b5d29ab8

by Brandon Chinn at 2026-08-25T18:42:08-04:00

Add Data.RealFloat and Infinity/NegInfinity/NaN pattern synonyms (#26961)

e60eb3bc

by Andreas Klebinger at 2026-08-25T18:42:59-04:00

rts linker: Fix pointer arithmetic issue in flushInstructionCacheRISCV64

We accidentally operated over `uint64_t*` when we should use `uint8_t`.

Fixes #27569

e9bbe8f9

by Andreas Klebinger at 2026-08-26T15:09:23-04:00

cmm dumps: Add machop width info with -dppr-debug for infix ops.

86e3a9d8

by Andreas Klebinger at 2026-08-26T15:09:24-04:00

CmmLint: Check for unsupported MachOp widths

machOpArgReps now maps MachOp + Width to a list of supported
argument widths or Nothing if the given operation is not supported
at the given width.

This allows us to check for nonsensical combinations like FloatToInt
at Word16.

Similarly we now check that every address is actually wordwidth.

13781cca

by Andreas Klebinger at 2026-08-26T15:09:24-04:00

arm64 ncg: The big subword truncation fix.

A set of slightly related fixes to arm subword handling:

Bitmask immediates:

Don't produce overflowing assembly literals.
There is still another bug here that causes us to miss some valid
literals but we will fix that later.

Improve subword truncation handling:

We now use a small set of helpers to truncate `Register` values rather
than truncating immediate `Reg` values which greatly simplifies the code
structure. This fixes a great many bugs to do with sign/zero extending subwords
or the lack thereof.

We now establish the invariant that subword values are zero-extended at
every site at which they come into "scope" of the ncg, and rely on the
invariant throughout rather than pessimistically inserting redundant
extensions in a hodgepodge manner at the use sites of these values.

This fixes at least the bugs described in issues #27533, #27430
 #27537, #27538, #27539, and #27550. But likely more bugs yet not
found.

Subword ffi results:
Apply truncations when calling functions returning
subword values.

genCondJump:
Don't sign extend signed values in the input register as
it might map to a local variable, corrupting the value stored within.

Fix subword store/load instructions.:

We used to read those at 32bit width even for smaller values possibly
resulting in invalid memory access. Now we construct the suffix for
subword variants based on the instruction format for these.

d8fa5d7c

by Andreas Klebinger at 2026-08-26T15:09:24-04:00

arm64 ncg: Fix MO_V_Broadcast for non-literals.

We now use OpReg instead of OpScalarAsVec as required since we broadcast a gp register.

Also adds a test. Fixes #27565.

94822c95

by Andreas Klebinger at 2026-08-26T15:09:24-04:00

Add some test cases covering bugs in the arm ncg.

* Test for #27430 (subword ffi results)
* #27537 - subword conversions
* #27538 - subwords used in conditional
* #27533 - single byte read

dd1ba88a
by Andreas Klebinger at 2026-08-26T15:09:24-04:00
```
cmmLint: Lint against MO_FS_Truncate subword use.
```

fd22f71e

by Zubin Duggal at 2026-08-26T15:10:20-04:00

ghc-internal: annotateSTM should use catchSTM# rather than catch#

A catch# frame inside a transaction breaks retry and async exception
delivery.

Fixes #27657

bb324171

by Rodrigo Mesquita at 2026-08-26T15:10:59-04:00

rts: refactor to reduce THREADED_RTS in MSG_UPD_TSO_FLAGS

- No behavior change in this commit (well, a small optimization here
  makes us do less work if the target TSO owned by the curr. capability)
- Move all THREADED_RTS CPP needed into `updThreadFlag`
- Merge MSG_SET_TSO_FLAGS and MSG_UNSET_TSO_FLAGS into MSG_UPD_TSO_FLAGS
  plus a `set` bool field in the MessageUpdTSOFlag struct

Towards #27729

ed99b7b7

by Rodrigo Mesquita at 2026-08-26T15:10:59-04:00

rts: Fix race condition in MSG_UPD_TSO_FLAGS execution

The code for processing the MSG_UPD_TSO_FLAGS message was not taking
into consideration that the TSO's owner might have moved in between that
capability receiving the message (since it was its previous owner) and
starting to process its inbox (a point at which it was no longer the
owner)

Added Note [TSO owner may change in between Msg being sent and received]
to explain this race and the pattern used to fix this, where we just
forward the message to the new owner.

Fixes #27729

cd653714

by Alan Zimmerman at 2026-08-26T15:11:49-04:00

EPA: Uses Parsers.parseModule for exactprint tests

Parsers.parseModule is the advertised way to parse for use for exact
printing in the ghc-exactprint library. This commit updates the GHC
exact print testing to use it.

This requires moving the comment balancing that was occurring
only in the test path into the advertising parser path, so it moves
from Transforms.hs to Utils.hs.

Also update the comment adding to honour trailing annotations

b00b30d8

by Alan Zimmerman at 2026-08-26T20:18:38+01:00

EPA: Some Haddock processing tweaks

These changes to the Haddock postprocessing should not change
behaviour, but just bring it more closely in line with the
original, changed at 44309cd377f

And add some haddock exactprint tests to show they work.

3eecccf1