Kristofer Karlsson found that t4205's EXPENSIVE tests run several times slower on macOS than on Linux in CI. His measurements attribute much of the time to tr(1). The macOS tr processes input one character at a time through getwchar() and putwchar(), even in the C locale. For the more than 2GB t4205 generates, tr alone used 95 to 123 seconds of CPU time on macOS runners, against about 6 seconds on Linux.

The three-patch series renames test-genzeros.c to test-genbytes.c, adds a genbytes command and uses it instead of genzeros | tr. The third patch folds genzeros into genbytes, spelling the NUL byte as "NUL", and converts the remaining callers. Karlsson asked in the cover letter whether that patch is wanted, or whether genzeros should stay as a thin wrapper.

Junio Hamano objected to splitting the rename from the modification. He later withdrew his first concern, that test-tool genbytes would not work after patch 1, since the dispatcher still recognises genzeros. His point about the split stands, though. He listed ten test scripts that would fail because they could not find test-tool genzeros, and suggested squashing patches 1 and 2 and formatting with git format-patch -M3 so the diff shows as a rename with modification.

Hamano also suggested that genbytes take the byte value as exactly two lowercase hex digits, so NUL needs no special case. He argued that for a test helper simplicity matters more than ergonomics. He also asked whether forcing LANG=C LC_ALL=C on the tr invocation would help. The thread does not show a reply to that question or a reroll.