This optimization is based on the assumption (which at least holds true
for the dill file extracted from the mentioned bug) that only very few
nodes actually have metadata associated with it. For the test file, more
than two million node-offsets is looked up, and only ~7.5 thousand nodes
actually have metadata on it.
What is done here is to assume that most often we ask offsets in order
and that most often the answer is that there's no metadata (when we fall
into this case we can return `false` quickly) and we can skip looking up
the actual metadata. This furthermore - although not benchmark here -
should make it (a lot) faster in the case where we have multiple
subsections as we here only lookup in in one index. Only upon a `true`
response from that we have to lookup in all subsections to find and read
the actual metadata to associate.
Reading the test file without metadata, taking an average of averages
(average of 5 runs of the binary benchmarks output for
"AstFromBinaryEager") takes: 374481.4 us.
Before https://dart-review.googlesource.com/c/sdk/+/391761 reading the
test file with metadata takes: 627120.68 us.
This was about a quarter of a second, or 67+% slower.
With https://dart-review.googlesource.com/c/sdk/+/391761 this number
became: 441784.44 us.
A reduction of ~185 ms, or almost 30%, making the tax of reading
metadata "only" ~18%.
With this CL this number goes to 398011.44 us.
This is a further reduction of ~43 ms, or almost 10% (for a combined
reduction of ~230 ms or 36+%).
The tax of reading metadata is now ~6.3%.
Bug: https://github.com/flutter/flutter/issues/156713
Change-Id: I20bbabecfc8976293fe897528d05fc2145e2a8c0
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/392021
Reviewed-by: Johnni Winther <johnniwinther@google.com>
Commit-Queue: Jens Johansen <jensj@google.com>
Jens Johansen
6d4ebea83b