Skip to content

Interconditional signature - #32

Closed
adienes wants to merge 108 commits into
masterfrom
interconditional-signature
Closed

adienes wants to merge 108 commits into
masterfrom
interconditional-signature

Conversation

@adienes

@adienes adienes commented Oct 30, 2026

Copy link
Copy Markdown
Owner

No description provided.

KristofferC and others added 30 commits October 14, 2026 14:23
Fixes JuliaLang#19280.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#23985.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#25675.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#26828.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ex number

Fixes JuliaLang#34571.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#40460.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#41215.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#41844.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#45591.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#45776.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#46092.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#46285.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#47006.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#48613.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…cstrings

Fixes JuliaLang#48762.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#53043.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#53436.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#53645.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#53734.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ialize

Fixes JuliaLang#54169.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#54194.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#31943.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes JuliaLang#59454.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…docs

Fixes JuliaLang#59611.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
read_linux_id_map_max parsed the last field of /proc/self/uid_map as
Int. Outside a restricted user namespace that field is 4294967295,
which overflows Int32, so on 32-bit Linux the OverflowError aborted
test/file.jl at top level and its remaining tests never ran. CI's
i686 jobs run in sandboxes with smaller ranges and do not hit this.

Parse as Int64 before the existing reduction to Cint; on 64-bit the
expression is unchanged.

Assisted-by: Generative AI
…aLang#64363)

* tuple: throw `ArgumentError` from `split_rest` for negative `n`

Co-authored-by: Miles Cranmer <miles.cranmer@gmail.com>

* tuple: use `LazyString` for `split_rest` errors

Co-authored-by: Miles Cranmer <miles.cranmer@gmail.com>

---------

Co-authored-by: Miles Cranmer <miles.cranmer@gmail.com>
Co-authored-by: Dilum Aluthge <5619885+DilumAluthge@users.noreply.github.com>
…uliaLang#64362)

Co-authored-by: Miles Cranmer <miles.cranmer@gmail.com>
Co-authored-by: Dilum Aluthge <5619885+DilumAluthge@users.noreply.github.com>
…uliaLang#64301)

* mpfr: throw `ArgumentError` for invalid base in `tryparse(BigFloat)`

Co-authored-by: Miles Cranmer <miles.cranmer@gmail.com>

* mpfr: defer invalid base error formatting

Co-authored-by: Miles Cranmer <miles.cranmer@gmail.com>

---------

Co-authored-by: Miles Cranmer <miles.cranmer@gmail.com>
Co-authored-by: Dilum Aluthge <5619885+DilumAluthge@users.noreply.github.com>
…4415)

Like many other method that supports do-style syntax, the method
`lock(f, ::Channel)` only forwards `f`. Julia by default does not specialize
on forwarded arguments. This makes this function call uninferrible, unless
the callee happens to be inlined.

Co-authored-by: Dilum Aluthge <5619885+DilumAluthge@users.noreply.github.com>
…#64389)

since JuliaLang#59025 the `Array` method indexes into `I`, which `LogicalIndex` doesn't support, so `A[only(to_indices(A, (mask,)))] = X` threw a `CanonicalIndexError`; send it through the generic path, which iterates

Fixes JuliaLang#64386

Assisted-by: Claude Code (Opus 5.5)
IanButterworth and others added 27 commits October 31, 2026 10:47
…ground runs

Cover reporting of circular dependencies, verbose timing output, process stats, merging a request into a running session, stopping a background run, waiting on a held cache file lock, and the missing-dependency error that lists manifests from other Julia versions.

Assisted-by: Claude Code (Opus 5.5)
Stopping a run before any package had finished printed no summary and
stored no result, so neither the caller nor a later status request
showed that the in-flight packages were canceled. The canceled or
interrupted summary is now produced whenever the run was stopped.

Assisted-by: Claude Code (Opus 5.5)
A graceful stop lets running packages finish, so its summary read
"N dependencies precompiled, 0 interrupted". The background task then
caught the resulting interrupt and replaced that summary with a generic
"Background precompilation was interrupted", so a later status request
lost what had been done. The summary now leaves out a zero count, and
the generic message is only stored when the run did not report.

Assisted-by: Claude Code (Opus 5.5)
TOML reads `deps = []` as a `Vector{Any}`, which failed the
`Vector{String}` type assertions where Base and the precompile driver
read a manifest entry's `deps` and `weakdeps`, so loading or
precompiling such an environment threw a `TypeError`. An empty list is
now read as an empty list of names.

Assisted-by: Claude Code (Opus 5.5)
When a package failed to load a name that is not its dependency, the
hint listing manifests resolved by another julia printed "(vnothing)"
for a manifest without a `julia_version` field. Such a manifest was
most likely written before Julia 1.7, so it stays in the list, now
marked as having no recorded julia version.

Assisted-by: Claude Code (Opus 5.5)
Co-authored-by: IanButterworth <1694067+IanButterworth@users.noreply.github.com>
…err` (JuliaLang#64508)

* JuliaLowering tests: don't leave an `err` global in Main

The vendored tests run in `Main`, so `err = try ... end` in quoting.jl
created `Main.err`, shadowing the REPL's `err`. A REPL test later on the
same worker then read the wrong output and hung. Rename the variable, and
check that `ans` and `err` in `Main` are left alone.

* REPL tests: close the test's IO when the timeout expires

`kill_timer` threw an error into the test task, but an enclosing `@test`
caught it and the task blocked on its next read, so a desynchronized test
hung until the CI job timed out. Close the test's pipes instead, so reads
return EOF, and fail the test if the timer fired.
…agree with `tmerge` (JuliaLang#64441)

`tmerge` embeds the const as a conditional on the same slot but `⊑` only accepted the degenerate one, so `stupdate!` never converged

Assisted-by: Claude Code (Opus 5.5)
…64503)

JuliaLang#61915 made `Type{X}` its own kind and the new branch stopped counting it as a level, so recursion that moves types within a `Type{}` argument widens where 1.13 didn't

Assisted-by: Claude Code (Claude Opus 5.5)
…t`/`searchsortedlast` (JuliaLang#64468)

* ranges: Fix overflow in `_findin` for narrow and mixed integer ranges

`_findin(r, span)`, which backs `findall(in(span), r)` for integer ranges,
computed endpoint differences in the ranges' element type. For narrow,
unsigned or mixed-signedness element types these differences could wrap,
returning out-of-bounds indices (e.g. `-105:11` for
`findall(in(Int8(50):Int8(100)), Int8(-100):Int8(-90))`), throwing a
`DivideError`, or missing matches for negative steps.

Return an empty result up front when either range is empty or the two do
not overlap. In the remaining cases every difference is bounded by
`abs(last(r) - first(r))`, so it is computed exactly as an unsigned value
and divided by `uabs(step(r))`.

Fixes JuliaLang#64453

Assisted-by: Nocturne (Gemini 4 Argon)
Assisted-by: Claude Code (Opus 5.5)

* ranges: Use `searchsortedfirst`/`searchsortedlast` in `_findin`

Move the unsigned endpoint-difference computation to the integer range
methods of `searchsortedfirst` and `searchsortedlast` in `base/sort.jl`
(fixing the same overflow/underflow bug there for both forward and
reverse orderings), and delegate `_findin` in `base/range.jl` to them.

Assisted-by: Nocturne (Gemini 4 Argon)
Assisted-by: Claude Code (Opus 5.5)

* ranges: Keep `_findin` overflow tests within `typemax(Int)` indices

Two regression tests used ranges whose indices do not fit in `Int` on
32-bit platforms, so `searchsortedfirst` threw `InexactError` there.
Use ranges that still exercise the overflow in the endpoint difference
while keeping the resulting indices representable on 32-bit.

Assisted-by: Claude Code (Opus 5.5)

* ranges: Handle disjoint and non-fixed-width ranges in `_findin`

Return an empty result up front when either range is empty or the two
do not overlap. Otherwise `searchsortedfirst` can overflow
`length(r) + 1` when `length(r) == typemax(Int)`, e.g. for
`findall(in(UInt(2)^63:UInt(2)^63+5), 1:typemax(Int))`.

Only use `searchsortedfirst`/`searchsortedlast` for fixed-width integer
element types. Their results are converted to `Int`, which throws for
`BigInt` ranges with indices beyond `typemax(Int)`, so other integer
types keep the original arithmetic, which is exact for them.

Assisted-by: Nocturne (Gemini 4 Argon)
Assisted-by: Claude Code (Opus 5.5)

* ranges: Use `searchsortedfirst`/`searchsortedlast` for all integer types in `_findin`

Remove the separate implementation for non-fixed-width integers. `BigInt`
ranges whose indices exceed `typemax(Int)` now throw `InexactError`, matching
`searchsortedfirst`.

Assisted-by: Nocturne (Gemini 4 Argon)
Assisted-by: Claude Code (Opus 5.5)
…iaLang#60913)

filterfirst(f, itr) returns the first element of itr for which f
returns true, equivalent to first(Iterators.filter(f, itr)). It throws
an ArgumentError if no element matches, because a matching element may
itself be nothing, so nothing cannot signal "not found".

Also note in the first docstring that only an AbstractRange returns a
value when empty.

Fixes JuliaLang#26287
Fixes JuliaLang#44996

Assisted-by: Generative AI
`filesize` returns an `Int64` on every platform, but the verbose timing
formatter only accepted an `Int`. On 32-bit platforms every package then
failed to precompile with a `MethodError` when `verbose=true`. The
formatter now accepts any `Integer`, and the cache size is summed as an
`Int64`.

Seen in https://buildkite.com/julialang/julia-pr/builds/5076

Assisted-by: Claude Code (Opus 5.5)
Bumps [julia-actions/setup-julia](https://github.com/julia-actions/setup-julia) from 3.0.2 to 3.0.3.
- [Release notes](https://github.com/julia-actions/setup-julia/releases)
- [Commits](julia-actions/setup-julia@fa02766...da90847)

---
updated-dependencies:
- dependency-name: julia-actions/setup-julia
  dependency-version: 3.0.3
  dependency-type: direct:production
  update-type: version-update:semver-patch
...

Signed-off-by: dependabot[bot] <support@github.com>
…parseArrays-b8476d0-master

🤖 Bump SparseArrays stdlib 66a7b5c → b8476d0
…ctions/master/julia-actions/setup-julia-3.0.3

build(deps): bump julia-actions/setup-julia from 3.0.2 to 3.0.3
Co-authored-by: DilumAluthge <5619885+DilumAluthge@users.noreply.github.com>
Co-authored-by: DilumAluthge <5619885+DilumAluthge@users.noreply.github.com>
* test: only count Base methods in the `core` comparison-method check

The check that `>` and `>=` have a single fallback method counts every
method in the process, so it fails when an earlier test on the same
worker defined comparisons for its own types, as the Random tests do for
`FakeFloat64`. It only passed because `core` used to run before Random.
Count the methods defined in Base, which is what the check is about.

* test: only check Julia's own methods in numbers' and missing's inference checks

These checks look at every method in the process, so they failed when an
earlier test on the same worker had added some: the constructors for
ranges' `RealWrapper` and math's `EvenInteger` infer as `Any`, and the
Random tests load SparseArrays, whose `isequal` methods don't infer to
`Bool`. Check only Core and Base constructors, and Base's `isequal`.

* test: start the longest tests first

The parallel tests run in the order of `TESTNAMES`, with the stdlib
tests appended, so several of the longest tests (cmdlineargs, compileall,
jit, JuliaLowering, Random) only start 10 to 17 minutes into a CI run and
then run alone while the other workers sit idle. On the 12-worker Linux
jobs that leaves 27 to 33% of the worker time unused.

Replace the hand-picked list of tests moved to the front by a table of
the longest tests, ordered by their duration in recent CI runs on Linux,
macOS and Windows, and sort the tests by it. Replaying recent CI runs,
this keeps the workers of the 12-worker Linux suite about 0.84 busy (about
5 minutes less per job) and takes about 4.5 minutes off the 6-worker
Windows suite; the 4-worker Macs already pack well and gain little.

`ambiguous` now always goes first. It has to be the first test on its
worker, but more tests were moved in front of it, so with fewer than six
workers it could run on a worker that had already run other tests.
…4485)

The tests that must run on node 1 run one after another once all workers
are done, which on CI takes 3.5 to 4.5 minutes with every other worker
idle. On 64-bit, only the tests that add worker processes still need
node 1:

- `ccall` (JuliaLang#50350) and `gc` (JuliaLang#49579) were moved there for failures of
  32-bit workers that had used a lot of memory; keep them on node 1 on
  32-bit only.
- `stress` was moved there so that exhausting file descriptors and pipe
  memory would not affect other tests (JuliaLang#28270), but the descriptor test
  has run in its own subprocess with a low ulimit since JuliaLang#35040.

Replaying recent CI runs, together with longest-first ordering this keeps
the 12 workers of a Linux CI job about 0.93 busy, against 0.67 to 0.73
before.
…ature

`conditional_argtype` already did this for `Const` and `Bool` returns, so an equivalent `InterConditional` lost the signature

Assisted-by: Claude Code (Opus 5.5)
@adienes
adienes force-pushed the interconditional-signature branch from fd3beb8 to ace3618 Compare November 1, 2026 19:15
@adienes adienes closed this Nov 1, 2026
@adienes
adienes deleted the interconditional-signature branch November 2, 2026 11:46
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.