Invalid free error in UnrollerTestGroup/UnrollerTest.TestAllConvert/SVE2_128

[RossComputerGuy]

libhwy> 0 of 1 size 8 0.000000 -> 0
libhwy> 0 of 1 size 8 0 -> 0.000000
libhwy> 0 of 3 size 8 0.000000 -> 0
libhwy> 1 of 3 size 8 0.250000 -> 0
libhwy> 2 of 3 size 8 0.500000 -> 0
libhwy> 0 of 3 size 8 0 -> 0.000000
libhwy> 1 of 3 size 8 0 -> 0.000000
libhwy> 2 of 3 size 8 0 -> 0.000000
libhwy> free(): invalid pointer
libhwy> 
libhwy> 
libhwy> 99% tests passed, 1 tests failed out of 1530
libhwy> 
libhwy> Total Test time (real) =   7.70 sec
libhwy> 
libhwy> The following tests FAILED:
libhwy>         1519 - UnrollerTestGroup/UnrollerTest.TestAllConvert/SVE2_128  # GetParam() = 16777216 (Subprocess aborted)
libhwy> Errors while running CTest
libhwy> FAILED: [code=8] CMakeFiles/test.util
libhwy> cd /build/source && /nix/store/vai8b7k3jqg1b82fs13x4r3qr1hi6h8x-cmake-4.1.2/bin/ctest
libhwy> ninja: build stopped: subcommand failed.

Using v1.0.7 on aarch64-linux

[jan-wassenberg]

Hi, thanks for reporting. That is quite an old release. We generally fix issues for the current release. Are you also seeing this at HEAD?

[jan-wassenberg]

Closing, feel free to reopen :)

[Thyre]

I just came across this, as we had a test build failure with v1.4.0 as well:

2802/2815 Test #2802: UnrollerTestGroup/UnrollerTest.TestAllConvert/SVE2_128 ......................................................................Subprocess aborted***Exception:   0.00 sec
Running main() from /dev/shm/swmanage/jupiter/googletest/1.17.0/GCCcore-15.2.0/googletest-1.17.0/googletest/src/gtest_main.cc
Note: Google Test filter = UnrollerTestGroup/UnrollerTest.TestAllConvert/SVE2_128
[==========] Running 1 test from 1 test suite.
[----------] Global test environment set-up.
[----------] 1 test from UnrollerTestGroup/UnrollerTest
[ RUN      ] UnrollerTestGroup/UnrollerTest.TestAllConvert/SVE2_128
free(): invalid pointer

This didn't occur in a prior build on the same system, or in a build on another system with very similar hardware.
We did port over #3107 as a patch, but otherwise this is a standard v1.4.0 release.

Maybe there's a rare (race) condition where this test failure occurs?

The system was a login node of JUPITER, i.e. one 72 core Grace Hopper chip, RHEL 9.8 and ~512 GiB of memory.
We used GCC 15.2.0, Google Test v1.17.0 and CMake 4.2.1.