The Swift test job pulls in grpc-swift-2 -> grpc-swift-nio-transport ->
swift-nio-ssl -> swift-crypto, and this package ships no Package.resolved,
so that whole graph is resolved afresh on every run. On 2026-10-15 it
resolved swift-crypto 5.0.1 (released three days earlier) and "Test Swift
Linux (6.2)" turned red before a single line of FlatBuffers' Swift code was
compiled:
Sources/Crypto/HPKE/Ciphersuite/HPKE-AEAD.swift:27:6: error: unknown
attribute 'nonexhaustive'
5.x declares swift-tools-version 6.2, which is why only the 6.2 job selects
it; the 6.0 and 6.1 jobs prune it and keep building. The 6.2.x toolchains
then reject the `@nonexhaustive` that 5.x uses, a combination swift-crypto
does not exercise because it builds on nightly.
Restrict the test-only graph to 4.0.0 ..< 4.4.0, which resolves to 4.3.0:
the last 4.x release that declares swift-tools-version 6.0 and so builds
with every toolchain in the matrix. 4.4+ asks for 6.1 and would break the 6.0
job, and the whole 4.x line is free of `@nonexhaustive`.
Co-authored-by: Derek Bailey <derekbailey@google.com>
Commit 6dda43bb (#9425) added `alignas(uint64_t)` to embedded binary schemas
generated by `flatc --bfbs-gen-embed`. Regenerate `tests/monster_test_bfbs_generated.h`
via `scripts/generate_code.py` to keep checked-in test artifacts in sync and
unblock CI Check Generated Code.
GitOrigin-RevId: 3786479bebd33365edc91aa13388cdf95d7313d9
In `src/idl_gen_cpp.cpp`, code generated by `flatc --bfbs-gen-embed` emits `static const uint8_t bfbsData[...] = { ... };`. Because `uint8_t` has 1-byte alignment, the array may be placed at an unaligned address in `.rodata`, triggering alignment traps under Clang UBSan on Arm when accessed via FlatBuffers root accessors.
Add `alignas(uint64_t)` to the generated `bfbsData` array declaration in `GenerateBinarySchema` so embedded binary schemas are 8-byte aligned across all architectures.
PiperOrigin-RevId: 1003962636
Restrict the `__linux__` check for `FLATBUFFERS_LOCALE_INDEPENDENT` in `include/flatbuffers/base.h` to `(!defined(__linux__) || defined(__GLIBC__))` so Linux with glibc enables `strtod_l`/`strtof_l`.
PiperOrigin-RevId: 1002266254
* [Swift] Verify size-prefixed roots from reader position
Validate the size prefix before advancing ByteBuffer.reader, then use the active reader position consistently for root, file identifier, and returned-object verification. Cover malformed decoy roots, oversized prefixes, short identifiers, and valid prefixed identifiers.
* Use wrapping arithmetic in Swift verifier
---------
Co-authored-by: Darren Carreras <283775510+carrerasdarren-cell@users.noreply.github.com>
* Migrating from Xctests to swift testing
This migrates to the new Swift testing framework,
which would allow us to always use the latest tech
from swift moving forward.
* Updates flag to make sure that Wasm testing works
* Migrate to swift 6.0 & swift-gRPC 2.0
The following migrates to swift 6.0, and also
migrate to swift-grpc 2.0 that uses swift-nio
under the hood to provide nicer API and async await
Adds sendable to enum & update @_implementationOnly imports to use internal imports
* Address PR comments regarding misspelling & proper method naming.
VerifyKey() returns true on the first non-zero byte instead of
checking for a null terminator. This causes VerifyBuffer() to accept
FlexBuffers with non-null-terminated keys. Subsequent access to those
keys via strlen()/strcmp() reads out of bounds.
The condition if (*p++) should be if (!*p++) — return true
when a null terminator is found, not when any non-zero byte is found.
Confirmed with AddressSanitizer: heap-buffer-overflow in strlen()
after VerifyBuffer() returns true on a corrupted buffer.
Copy the repo-root LICENSE into the Python package directory for the duration of setup() so license_files = LICENSE remains valid without using deprecated parent-directory paths.
Remove the staged copy after the build completes.
* fix: correct operator precedence in ForAllFields reverse iteration
The expression `size() - i + 1` evaluates as `(size() - i) + 1` due to
left-to-right associativity, producing an out-of-bounds index when
reverse=true. For a vector of size N, the first iteration (i=0) accesses
index N+1, which is 2 past the last valid index.
Changed to `size() - (i + 1)` to match the correct implementation
already present in bfbs_gen.h:192.
Bug: CWE-125 (Out-of-bounds Read), CWE-783 (Operator Precedence Error)
* test: add ForAllFieldsReverseTest for reverse iteration correctness
Verify that ForAllFields with reverse=true iterates fields in
descending ID order. Tests both Stat (3 fields) and Monster
(many fields with non-sequential definition order) tables.
---------
Co-authored-by: Tulgaa <tulgaa.kek@gmail.com>
* Add fallible try_* API for FlatBufferBuilder
This is to support error propagation from Allocator trait. The Allocator
grow_downwards() method returns Result<(), Self::Error>, but
FlatBufferBuilder panics via .expect() when allocation fails instead of
propagating the error.
* Add rust fallible API docs
A blob is an array of bytes and has no intrinsic alignment (i.e. the
alignment is 1). The alignment of the existing flexbuffers blob is
solely affected by the width of the integer needed to store the blob's
size: that integer's width becomes the alignment of the blob.
The proposed AlignedBlob function here piggybacks on this effect and
simply uses a user-defined alignment for the width of the integer that
stores the blob's size; this automatically imparts that same alignment
on the blob itself. (The width is bounded below by the actual width
needed to store the blob's size.)
The ability to control the alignment of a blob is important for use
cases in which the blob itself stores structured data that we want to
access without further copies (e.g. other flatbuffer messages).
String default values parsed from .fbs schemas are un-escaped by the IDL
parser (e.g., \x22 becomes a raw " byte), but code generators embed these
raw values directly into generated source code string literals. This allows
specially crafted .fbs files to break out of string literals and inject
arbitrary code into generated C++, Rust, TypeScript, and Swift source.
Fix by adding EscapeCodeGenString() helper that re-escapes string content
before embedding, and applying it to all 7 affected injection points across
5 code generators (C++, Rust, TypeScript, Swift, FBS).
Resolves the TODO comments in idl_gen_cpp.cpp and idl_gen_rust.cpp.
* fix(flatbuffers): use manual impl Default for struct object types
* fix: handle bool and float zero literals in struct object Default impl
* fix: regenerate all test bindings with generate_code.py
* fix: data type check on swift build
* fix: test large array on struct and enum
Anshul Nigham