Otherwise using callcc fails with a segfault for instance in this test case:
TestEnsureAndCallcc#test_bug20655_extension_using_rb_ensure
(cherry picked from commit 142001cd64fcb7ec6896f04b0c837857590e26ce)
When setjmp/longjmp/exceptions are used on Windows it's necessary to store+restore additional information from the TEB.
I didn't find any official documentation about the values to be saved, but found the corresponding boost/context implemenataion:
abf8e04e23
This is similar to the special TIB handling on x86/x86_64 on Windows.
Without this fix an exception in a fiber segfaults without any output:
ruby -e "Fiber.new{ raise 'test' }.resume"
(cherry picked from commit 187b8fdb6990145210ef6394ca8d866c2f885854)
The conflicts in coroutine/arm64/Context.S and Context.h are resolved by keeping the ruby_3_3 context, which has neither the PAC b-key check nor the uintptr_t cast of the start function.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two minor fixes to arm64 instruction decoding when looking for __pioinfo:
1. add_mask was shifted by one bit, it was intended to be 0x7f800000. However, since the mask was already excluding matching the 'sh' bit, and since the purpose of the add following the adrp is to add in the lower 12 bits, I opted to set the mask to 0x7fc00000 and simply remove the handling for the 12 bit shift option which is now required to be disabled in order to match.
2. adrp's immediate was supposed to be sign extended. So far, I have not seen cases where the global variable ends up before the code in memory, but it's a possibility, so handle the sign extension.
(cherry picked from commit 4745338a3f336d7bedb02895534721661551ebb9)
`_umul128` is specific to x86_64 platform, see higher words by
`__umulh` on arm64.
(cherry picked from commit 5a7a1a4a13ae11e4cef274c98e0c5bbcb7b1d192)
Use `_WIN64` for word-size, `_M_AMD64` for CPU-specific feature.
(cherry picked from commit c77f736bc1e6b73d62e9ba3dcf3dcbeff0657b38)
The static assertions added to the fallbacks of rbimpl_atomic_size_inc and rbimpl_atomic_size_dec are left out, as the arm64 mswin build does not need them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With a subexp call, the result of unset_addr_list_fix overwrote the error, so a pattern over the limit compiled without raising RegexpError.
Follow-up to cf35145fc46.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 16d8a554b2ad279e08128b5e1dc4b342fccd33d7)
Concatenated large repeats reach a program past INT_MAX from a few hundred
kilobytes of source, and the offsets emitted into it wrap. Cap the size
while the program is emitted, which does not depend on the length pass that
overflows along with it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
(cherry picked from commit cf35145fc46bc278e0ab0003478fc4afe6564819)
Nested bounded repeats let the int product tlen * qn->lower overflow, so
the check that keeps the compiler from unrolling a large repeat could pass
for an enormous one, hanging the compile or emitting wrapped jump offsets.
Compare against the limit by division instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 7cfe8dc82f9f1b3f9c37c78e362fcbdf3c0b84f4)
Add explicit errors for excessive range-repeat and null-check IDs;
map new errors in regerror; add regression tests for oversized generated
patterns.
(cherry picked from commit 493d26dcdea8b2f0ac5d28e057b401bc809e7fae)
Prevent double free for too big repetition quantifiers
The previous implementation calls `free(node)` twice (on parsing and compiling a
regexp) when it has an error, so it leads to a double-free issue. This
commit enforces `free(node)` once by introducing a temporal pointer to hold
parsing nodes.
(cherry picked from commit 35000ac2ed3a1829f8d193ffb23da0e44cc6fe5f)
CI: mingw: Use the official actions for msys2
Because ruby/setup-ruby is affected to test result.
Also includes the prerequisite changes from e4b386e090, 90f8199412,
a443cd012a and 4adb6f6969 (Remove Strawberry Perl pkg-config, misc
system & package info, `cmd` as the default shell, set up msys2
environment variables).
(cherry picked from commit 98ec35899b, adapted to the mingw workflow on ruby_3_3)
setup-ruby v1.236.0 raises "Cannot find Git's cert file" on Windows
runners because Git for Windows moved from mingw64 to ucrt64.
This was fixed in ruby/setup-ruby#947 (v1.328.0).
(cherry picked from commit 89856b96b1, adapted to the workflows on ruby_3_3)
cast memchr result (or make return const void *) to resolve -Werror incompatible-pointer-types"
(cherry picked from commit 2d4992fc9df22fe581df95879ffedf7470f0503c)
C23 has qualifier-preserving standard library functions, so calling
strchr with a `const char *` will return a `const char *`. We can change
the type of the local variables because we don't mutate the strings.
(cherry picked from commit e068856616922f117b6de74c58a63d386c7496c9)
The result jobs in macOS workflows only check if other jobs failed
(just runs `exit 1`). They don't need macOS runners. Using
ubuntu-latest instead frees up macOS runner slots and starts faster
due to shorter queue times.
To account for cases where binaries for older OS versions are built
using the new SDK, check at runtime whether the API functions are
available actually.
[Bug #21703]
RUBY_CRASH_REPORT does not work in some cases when shelling out on Linux.
For example, given the following shell script dump.sh:
#!/usr/bin/env bash
cat > /tmp/crash
And we see it fails like this:
$ RUBY_CRASH_REPORT="|dump.sh" ruby -rfiddle -e "Fiddle::Pointer.new(1, 10)[0]"
cat: -: Resource temporarily unavailable
MSYS2 dropped perl from its base install and the windows-2022 runner
image picked that up in 20260818.277.1, so autogen.sh now dies with
`/usr/bin/perl: bad interpreter`. master and ruby_4_0 set up their own
MSYS2 with autoconf, which pulls perl in as a dependency, so only the
branches that reuse the image MSYS2 need this.
https://github.com/actions/runner-images/issues/14562
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The full macOS fan-out (10 jobs, about 117 runner-minutes) repeats
test tasks that macos.yml already runs from the git tree on every
event; what is tarball-specific is only that the packaged tarball
configures, builds and installs. Run a single check on macos-26 for
pull requests and pushes, and keep the full matrix for
workflow_dispatch, which is how the nightly schedule and release
verification invoke it. Only macOS is trimmed: its org-wide
concurrency pool of 50 runners saturates during peak hours and has
been delaying release jobs, while Linux and Windows share a pool with
headroom.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit df977545f07404ac6b49b6992a4ab78c2f2b0f9a)
The cache step hardcoded snapshot-*, which no longer matches when this
reusable workflow is called with a release archname such as
ruby-3.4.0-rc1 from ruby/actions, turning cache save and restore into
a no-op.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 674e93935c846f129329e78c4270015ed1805e80)
Backport of the June 2026 net-imap security releases to the v0.4 line.
CVE-2026-47240
CVE-2026-47242
CVE-2026-47241
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Since Git 2.46, commands like `git commit` spawn a detached
`git maintenance run --auto` process, which briefly creates
`.git/objects/maintenance.lock`. Specs copying a just-committed
repository with FileUtils.cp_r can hit the lock file disappearing
between readdir and lstat, causing flaky Errno::ENOENT failures on
macOS CI runners with newer Git.
https://github.com/ruby/ruby/actions/runs/28764006107/job/85284933867
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
`/..%5c../` resolves the backslash as a path separator on Windows, so the
vendored FileHandler returns 404 instead of the expected 200 and logs a
nondisclosure warning, breaking the daily Windows snapshot CI. This vendored
test only remains on ruby_3_3; 3.4 and later removed it together with WEBrick.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
openssl@1.1 is no longer available on the macos-26 runner image. Build
against openssl@3, matching master (18404af9ec0), so the macos-26 check
job can install its dependencies.
The macos-15-intel runner image ships a pre-installed Ruby under
/usr/local whose default-gem dir collides with the build's on the 3.3.0
ABI path. During `make check`, the test-tool autoloads rake and scans
default-gem stubs there, hitting a stale bundler stub and failing with
ENOENT. Build under an isolated prefix on that runner only. Backport of
master #17339.
The macos-15 runner image ships a pre-installed Ruby under /usr/local
whose default-gem directory shares the 3.3.0 ABI path with the snapshot
build. With the default prefix, RDoc generation scans that directory and
dies on a stale bundler default-gem stub left by the runner's RubyGems
self-update. Build, install and verify under an isolated prefix instead.
(cherry picked from commit dc24122d900c32a14db56a4f1def4f79027a24ee)
Dir.each_child yields entry names, so stat/digest/children ran against
the working directory instead of HOME. Join dir and name first.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Now that win32/Makefile.sub passes XINCFLAGS so the gmp headers under
--with-opt-dir are visible to cl.exe (bug #22078), the gmp build no
longer fails. Restore gmp to the vcpkg install to match master.
Make macos-26 the default runner for the tarball build verification and
keep macos-15, macos-15-intel, and macos-14 as regression coverage.
(cherry picked from commit 5c762e786bb14ec46a15e1073e1d806fa9a39a1c)
Replace secrets: inherit with explicit Slack webhook secrets, which are
the only secrets the ubuntu/macos/windows builds use. tarball-non-development
uses no secrets, so drop inherit there entirely.
(cherry picked from commit daea48078cd029b16ef5a4a31fe7fedf72bb00f0)
The subsequent git steps only fetch from the public origin and operate
on local branches, so the token does not need to be persisted.
(cherry picked from commit f00e912db0d175ffe61d7195f40c10838be1c826)
