huntedbytheirsandSisyphus
53c8531eab
System-headers-only harness timing the five L1 string functions (strlen, strcmp, memcpy, memmove, memset) at 8- and 64-byte inputs. Calls go through volatile function pointers to defeat builtin substitution and LICM hoisting; correctness self-asserts route through the same pointers. Prints FUNC@SIZE ns/call lines. bench_string uses empty per-target CPPFLAGS (no -I include) so it links against whichever libc --with-libc= selects; the bench target runs it always and bench_vlibc only under BENCH_LINK_VLIBC. Both programs relink when config.status changes, so a reconfigure cannot silently reuse a binary from the previous libc. Measured vs glibc 2.44 (median of 3 runs): mem* 1.0-3.3x, strlen/strcmp 3.2-3.9x — word-at-a-time vs AVX2, all sane. Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent) Co-authored-by: Sisyphus <[email protected]>
Benchmarking
Every vlibc component is benchmarked against the software it replaces, per
musts/BENCHMARKING.md. A component that is slower or worse at its task than
the software it replaces is considered failing.
Layout
- Each benchmark harness is a standalone C program in this directory, named
bench_<component>.c. bench_vlibc.cis the skeleton that per-component benchmarks build on.
Running
./autogen.sh # configure (default: link against vlibc)
make bench # build and run the harnesses
# Reference comparison against the libc vlibc replaces:
./autogen.sh --with-libc=glibc
make bench
# musl reference (requires musl-gcc):
./autogen.sh --with-libc=musl
make bench
Adding a benchmark
- Add
bench_<component>.chere and list it inbenchmarks/Makefile.am. - Measure the component against the equivalent glibc (and musl) call.
- Report wall-clock time, throughput, and any binary-size difference.
- If vlibc is slower, the benchmark fails — fix the implementation.