[heifner]

Builds on Wire-Network/berkeley-softfloat-3#5 and Wire-Network/wire-vcpkg-registry#22, both merged. The registry baseline is abed0cc1, the #22 merge commit, whose tree is identical to the 2a0f6fc4 head this was verified against. The pending wire-sys-vm bump (Wire-Network/wire-vcpkg-registry#23, for Wire-Network/eos-vm#14) merges cleanly on top of abed0cc1; #703 re-pins to the registry commit carrying both.

Berkeley SoftFloat's rounding mode, tininess mode, exception flags and extF80 precision were process globals: the vcpkg port passed no THREAD_LOCAL, so every read-only transaction thread and every sys-vm / sys-vm-jit / sys-vm-oc execution raced on them. gofractally/eos-vm#24 and gofractally/eos-vm#25 fixed this upstream in external/softfloat.cmake, which Wire does not build. The port now exports THREAD_LOCAL (__thread on GCC and Clang for C and C++, __declspec(thread) on MSVC); sysio_chain links softfloat::softfloat PUBLIC, so every chain consumer's compile line carries the define and declares the state with the storage class the library defines it with (the linker rejects a TLS/non-TLS mismatch). __thread is the spelling softfloat.hpp and wasm-jit's Platform.h already use, so the redefinition in wasm_interface.cpp stays identical and silent. The pin also carries the softfloat fix for its arm64-only NEON include-order test, which broke the port build on arm64-osx on the first pin. The second macOS failure was in wire-sysio itself (controller.cpp): since berkeley-softfloat-3@67af45c the softfloat target no longer forces SOFTFLOAT_DISABLE_LEGACY_TYPE_ALIASES on arm64 consumers, and the legacy float32_t/float64_t/float16_t aliases collide with arm_vector_types.h, which Boost.Unordered's arm_neon.h include brings into most chain TUs. Nothing in wire-sysio or the sys-vm headers uses those aliases, so the root CMake now defines the macro for the whole tree (the x86_64 build is unaffected except for the define on every compile line). The generated SysioChain interface (cmake/SysioTester.cmake.in) links the softfloat archive directly and bypasses the port's package, so it now repeats the two definitions the package carries: THREAD_LOCAL=__thread, the storage class the archive was built with, and SOFTFLOAT_DISABLE_LEGACY_TYPE_ALIASES; a build-tree find_package(sysio) consumer sees both on SysioChain.

unittests/softfloat_thread_local_tests.cpp is the regression test for the pin: the calling thread sets softfloat_round_max and raises softfloat_flag_infinite, a worker must start from softfloat_round_near_even with clear flags, set softfloat_round_minMag and raise softfloat_flag_invalid, and neither side may see the other's state. Linked against the 3.0.0#2 state (process globals, no THREAD_LOCAL) all five checks fail: the worker sees round_max and the infinite flag, and the caller ends with minMag and invalid | infinite; against #3 it passes.

Verification (clang 18.1.8, Debug, x64-linux, contracts OFF):

• unit_test --run_test=softfloat_thread_local_tests under --sys-vm, --sys-vm-jit, --sys-vm-oc: *** No errors detected.

• libsoftfloat.a from the pinned port (tree bca9aee, identical at 2a0f6fc4 and abed0cc1, fetched from GitHub): softfloat_roundingMode, softfloat_detectTininess, softfloat_exceptionFlags, extF80_roundingPrecision are TLS GLOBAL in readelf -sW (3.0.0#2 had them as plain OBJECT); nodeop and unit_test reference them as TLS.

• softfloat-config.cmake exports a plain THREAD_LOCAL=__thread; 505 compile lines carry it and 684 carry -DSOFTFLOAT_DISABLE_LEGACY_TYPE_ALIASES. The build has zero warnings.

• unit_test --run_test=softfloat_golden_tests under --sys-vm, --sys-vm-jit, --sys-vm-oc: *** No errors detected (50 cases each).

• unit_test --run_test=wasm_part1_tests,wasm_part2_tests,wasm_part3_tests under --sys-vm, --sys-vm-jit, --sys-vm-oc: *** No errors detected (52 cases each).

• python3 tests/read_only_trx_test.py -p 2 -n 3 --read-only-threads 2: Test succeeded.

• Linux CI was green on both earlier pins (ubuntu24, ubsan, asan, gcc, asserton); the macOS job is the check for the arm64 alias collision, which cannot be built here.

[huangminghuang]

Medium: This definition is directory-scoped, so it does not reach consumers of the generated SysioChain/SysioTester interface. SysioTester.cmake.in links libsoftfloat as a raw archive and exports the SoftFloat include paths without this compile definition. On the documented Apple-Silicon wire-cdt integration path, tester.hpp includes softfloat_types.h before Boost.Unordered pulls in arm_neon.h, reproducing typedef redefinitions for float16_t, float32_t, and float64_t. I reproduced the failure with a Clang 18 AArch64 syntax probe; it passes when this macro is defined. Please export the definition on the generated SysioChain interface, while retaining the in-tree definition as needed.

[heifner]

Exported in 4af9ddb: SysioChain now carries SOFTFLOAT_DISABLE_LEGACY_TYPE_ALIASES and, since it links the archive directly rather than the port's package, also THREAD_LOCAL=__thread, the storage class the archive was built with. Checked with a build-tree find_package(sysio) consumer: both land in INTERFACE_COMPILE_DEFINITIONS. The tree-wide definition stays.

[huangminghuang]

Medium (test coverage): This version bump changes the four SoftFloat state variables from process-global to thread-local, but the current wire-sysio golden tests are serial and never assert rounding or exception-state isolation. The dependency contains a deterministic TLS test, but the registry port configures it with BUILD_TESTING=OFF, so PR CI does not exercise this behavior. Please add a two-thread regression test that assigns distinct rounding modes and exception flags and verifies they do not leak; it should fail with 3.0.0#2 and pass with #3.

[heifner]

Added unittests/softfloat_thread_local_tests in 4af9ddb: the calling thread sets softfloat_round_max and raises softfloat_flag_infinite, a worker must start from the defaults, set softfloat_round_minMag and raise softfloat_flag_invalid, and neither side may see the other's state. Against the #2 archive (no THREAD_LOCAL, process-global state) all five checks fail at runtime (the worker sees round_max and infinite, the caller ends with minMag and invalid | infinite); against #3 it passes, in this tree under --sys-vm, --sys-vm-jit and --sys-vm-oc.

[huangminghuang]

Re-reviewed the current head. The two earlier findings are addressed, and I found no remaining actionable issues.

[huangminghuang]

Reviewed the current diff and relevant SoftFloat, vcpkg registry, build/export, and regression-test context. No actionable findings.