test(bench): add llsc_tzconvert_bench fixture for ARM64 CAS livelock repro

Repeated localtime()/__tz_convert() calls exercise the outline-atomics
LDAXR/STXR CAS loop (glibc tzset_lock) that triggered the ARM64
fallback-LL/SC livelock. Wired into generate_config.py's CI matrix; the
CI workflow compiles it fresh (no arch-specific deps beyond libc), so
unlike testdata/take_strings-* it isn't checked in as a binary.

Author: not-matthias

.github/workflows/codspeed.yml

@@ -95,6 +95,12 @@ jobs:
       - name: Install uv
         uses: astral-sh/setup-uv@v5
 
+      # Compiled locally rather than checked in (unlike the take_strings-*
+      # binaries) since it has no arch-specific dependencies beyond libc.
+      - name: Build bench fixtures
+        working-directory: bench/testdata
+        run: cc -O2 -o llsc_tzconvert_bench llsc_tzconvert_bench.c
+
       # Generate the codspeed.yml for this Valgrind version. The script derives
       # the version label from `valgrind --version`, so each matrix job emits its
       # own config (e.g. valgrind.codspeed / valgrind-3.26.0 / valgrind-3.25.1).

.gitignore

@@ -2568,3 +2568,9 @@ test-suite.log
 
 # Capstone build install prefix (built from the third_party/capstone submodule)
 /.capstone
+
+# fake CodSpeed benchmark fixture binary (compiled from testdata/llsc_tzconvert_bench.c)
+/bench/testdata/llsc_tzconvert_bench
+
+# python bytecode cache left behind when something imports generate_config.py as a module
+/bench/__pycache__/

bench/generate_config.py

@@ -24,6 +24,11 @@
     "python3 testdata/test.py",
     "stress-ng --cpu 1 --cpu-ops 10",
     "stress-ng --cpu 4 --cpu-ops 10",
+    # Repeated localtime()/__tz_convert() calls: the guest CAS loop
+    # (outline-atomics LDAXR/STXR helper) that triggered the ARM64
+    # fallback-LL/SC livelock fix. Built by the "Build bench fixtures" CI
+    # step from testdata/llsc_tzconvert_bench.c (not checked in as a binary).
+    "testdata/llsc_tzconvert_bench 5000",
 ]
 
 # Callgrind configurations: (extra args, config name, requires_codspeed). The

bench/testdata/llsc_tzconvert_bench.c

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f428498cf6aa54398cc0c15877680c54e26a1e66e06669cf87b4a55c900e19d3
+size 1188