callgrind SIGILLs on AVX-512 despite the GLIBC_TUNABLES workaround

[Verdenroz]

Affected crate(s)

soothfast-measure

Version

0.2.0

Steps to reproduce

1. On a machine with glibc 2.44, valgrind 3.25.1, and an AVX-512 CPU.
2. cargo bench --no-run -p soothfast-registry --bench soothfast
3. Run callgrind the way callgrind.rs does, with the tunable it sets:

GLIBC_TUNABLES=glibc.cpu.hwcaps=-AVX512F,-AVX512VL,-AVX512BW,-AVX512DQ,-AVX512CD \
  valgrind --tool=callgrind --callgrind-out-file=/dev/null <bench binary> --list

Expected behavior

The tunable disables glibc's AVX-512 routines so callgrind can run the binary. That is what the workaround at soothfast-measure/src/callgrind.rs:16-19 exists for, and its docstring says so: "an AVX-512-compiled glibc loader SIGILLs before main()".

Actual behavior

SIGILL at the same address with and without the tunable, so the workaround no longer suppresses anything on this glibc.

vex amd64->IR: unhandled instruction bytes: 0x62 0xF1 0x7F 0x48 0x7F 0x84 0x24 0x30 0x0 0x0
valgrind: Unrecognised instruction at address 0x40bdecd.
Process terminating with default action of signal 4 (SIGILL)

The leading 0x62 is the EVEX prefix. The bytes decode as a 512-bit store, vmovdqu32 [rsp+0x30], zmm0. Valgrind 3.25.1's VEX frontend does not decode EVEX. The binary runs fine natively, and callgrind runs a trivial C program on the same machine without complaint.

Impact

probe() fails, gating falls back to walltime, and complexity sweeps over sizes(16, 64, 256) become noise. On unmodified master, three consecutive gate -p soothfast-spec runs each failed a different bench: bench_openapi_diff x2.48 then x3.22 then x3.45 against a x2.5 limit, with bench_openapi_document and bench_serialize_yaml swapping pass/fail between runs. On a machine in this state, any local pass or fail on a callgrind-gated claim is walltime noise.

Environment

valgrind 3.25.1, glibc 2.44, rustc 1.98.0, x86_64 with avx512f/bw/cd/dq/vl.