Avoid char array allocation in Starlark format

[fmeum]

Positional access via String#charAt is slightly faster than precreating the char array for Latin-1 strings and much faster for UTF-8 strings. It allocates less in both cases.

Also adds a --latin1 flag to the Benchmarks tool that allows benchmarking against Bazel's way of parsing Starlark files.

Before:

INFO: Running command line: bazel-bin/src/test/java/net/starlark/java/eval/Benchmarks --filter bench_format --seconds 20
File src/test/java/net/starlark/java/eval/testdata/bench_string.star:
benchmark                        ops     cpu/op    wall/op   steps/op   alloc/op
bench_format                16777215     1.46µs     1.46µs          7     2.26KB

INFO: Running command line: bazel-bin/src/test/java/net/starlark/java/eval/Benchmarks --filter bench_format --seconds 20 --latin1
File src/test/java/net/starlark/java/eval/testdata/bench_string.star:
benchmark                        ops     cpu/op    wall/op   steps/op   alloc/op
bench_format                16777215     1.47µs     1.46µs          7     2.26KB

After:

INFO: Running command line: bazel-bin/src/test/java/net/starlark/java/eval/Benchmarks --filter bench_format --seconds 20
File src/test/java/net/starlark/java/eval/testdata/bench_string.star:
benchmark                        ops     cpu/op    wall/op   steps/op   alloc/op
bench_format                16777215     1.79µs     1.79µs          7     1.67KB

INFO: Running command line: bazel-bin/src/test/java/net/starlark/java/eval/Benchmarks --filter bench_format --seconds 20 --latin1
File src/test/java/net/starlark/java/eval/testdata/bench_string.star:
benchmark                        ops     cpu/op    wall/op   steps/op   alloc/op
bench_format                16777215     1.37µs     1.37µs          7     1.67KB

[fmeum]

@tetromino Friendly ping

[tetromino]

Based on some internal testing on bazel query $VERY_LARGE_TARGET, this delivers a measurable CPU time savings, but does not translate into a statistically significant effect on overall wall time.

My only concern is in bench_string.star - you were running "".format(...), so the measurements you showed are probably not representative.

[fmeum]

Thanks, I fixed and reran the benchmark. It now shows that the new implementation is only better with the Latin-1 string encoding, but worse in UTF-8 standalone Starlark. Not sure whether that makes it a good idea.

[tetromino]

Interesting. In that case, we may want different behavior depending on whether the string is Latin1...

[fmeum]

Interesting. In that case, we may want different behavior depending on whether the string is Latin1...

You will probably be interested in #24417 :-)