Right now nin is not fast enough to be usable as a ninja replacement. Thankfully we can directly compare performance and find the obvious areas to optimize.
Generate a memory dump of allocated objects.
go test -memprofile mem.prof
# or
nin -memprofile mem.prof
# then
go tool pprof -http :8010 mem.prof
Visit http://localhost:8010/ui/flamegraph?si=alloc_objects
Saves a sample at ~100Hz. For short runs it may not be useful enough.
go test -cpuprofile mem.prof
# or
nin -cpuprofile cpu.prof
# then
go tool pprof -http :8010 cpu.prof
Visit http://localhost:8010/ui/flamegraph?si=cpu
Hint: Use -http 0.0.0.0:8010 to be remotely accessible.
Tracing keeps much more information by tracing everything that happens. It is more useful for short runs.
go test -trace=trace.out -bench=.
# or
nin -trace trace.out
go tool trace -http=:33071 trace.out
Runs all Go benchmark on the current commit and the previous one, then compares with benchstat. Use ba:
go install github.com/maruel/pat/cmd/...@latest
ba
This runs both ninja's performance tests and nin's equivalent for an apple-to-apple comparison.
./compare_against_ninja.sh
I wrote a small tool to post process go tool objdump output:
go install github.com/maruel/pat/cmd/...@latest
Disassemble a single function:
disfunc -f 'nin.CanonicalizePath$' -pkg ./cmd/nin | less -R
List all bound checks:
boundcheck -pkg ./cmd/nin | less -R
List all symbols in the binary and their size:
go tool nm -size nin | grep github.com/maruel/nin
See the compiler optimization process in a web page:
GOSSAFUNC=nin.CanonicalizePath go build ./cmd/nin && open ssa.html
Escape analysis:
go build -gcflags="-m" .