When to use#

You must A/B a code path between two versions of a CMake/Ninja project (for example a git worktree at the old commit and the working tree), no benchmark in the repo covers it, and you must not add one to either tree.

Steps#

1. Write the timing program in a scratch folder. Use only the public headers. A helper type from the tests (such as a test sink) is not in the library, so define a small one in the program.

2. In each tree, build a benchmark target that links the same library, so the libraries are current: cmake --build <tree>/build/build-release --target <bench>.

3. Take that target's compile and link lines from Ninja and swap in your source, object and output:

   cd <tree>/build/build-release
   cc=$(ninja -t commands src/<bench> | grep -- "-c .*<bench>.cpp" | tail -1)
   ln=$(ninja -t commands src/<bench> | tail -1)
   eval "$(echo "$cc" | sed "s#-o [^ ]*<bench>.cpp.o#-o $OBJ#; s#-c [^ ]*<bench>.cpp#-c $SRC#; s#-MD -MT [^ ]* -MF [^ ]*##")"
   eval "$(echo "$ln" | sed "s#: && ##; s#&& :##; s#src/CMakeFiles/<bench>.dir/__/benchmarks/<bench>.cpp.o#$OBJ#; s#-o src/<bench>#-o $EXE#; s#@loader_path/../lib#$PWD/lib#")"
   

   The last sed term points the rpath at the tree's own lib/, so each binary loads its own libraries.

4. Give each run fresh input (copy the data into a new folder per run), so a cache that one run writes does not serve the next.

5. Alternate A and B runs and compare medians. Run nothing else heavy meanwhile.

Evidence#

• dftracer-utils stage 10b (2026-09-27): qc_bench.cpp built against a54be906 and the working tree through index_bench's lines. The first try failed with error: unknown type name 'StringSink' (a type from tests/trace/views/test_view_common.h); a local NullSink : ExportSink fixed it. Three alternating runs gave medians within 1.3% between the builds.