gimli-rs / gimli / 17785491480

55%

push

github

Optimize unsigned LEB128 reading (#795)

Micro-optimize unsigned LEB128 reading by unrolling parts of the first
loop to improve code gen for the statistically likely case of single
byte values. This improves performance by a small amount as per my
experimentation:

Before:
  test bench_parsing_debug_info            ... bench:  1,854,730.00 ns/iter (+/- 2,237.98)
  test bench_reading_leb128_unsigned_small ... bench:        649.41 ns/iter (+/- 3.82)
  test bench_reading_leb128_unsigned_large ... bench:        203.58 ns/iter (+/- 1.51)

After:
  test bench_parsing_debug_info            ... bench:  1,809,890.00 ns/iter (+/- 2,453.82)
  test bench_reading_leb128_unsigned_small ... bench:        607.84 ns/iter (+/- 0.63)
  test bench_reading_leb128_unsigned_large ... bench:        198.43 ns/iter (+/- 0.20)

Adjust u16 in a way to make the first check symmetrical, though this
change does not appear to have any measurable performance impact. Also
add a test for a LEB128 encoded value that we want to make sure we don't
regress on by treating parsing differently.

Signed-off-by: Daniel Müller <deso@posteo.net>

7 of 7 new or added lines in 1 file covered. (100.0%)

6651 of 11625 relevant lines covered (57.21%)

1.48 hits per line