This tweaks the code and comments added in commit 36d7748e "Merge pull
request #90 from sarastro-nl/improvement/optimize-SDC" and commit
d46220ca "optimize SDC". Both of those two commit messages are short but
see the discussion re algorithmic changes on PR 87 (yes, 87, not 90):
https://github.com/google/wuffs/pull/87
Also run clang-format.
Also update doc/changelog.md
Compared to before commit d46220ca "optimize SDC", when running
script/manual-test-parse-number-f64.cc and disabling both the
etc__parse_number_f64_eisel_lemire function and the "If both man and
(10 ** exp10) are exactly representable by a double" fast path, the
tests still pass but the number of etc__high_prec_dec__small_lshift and
etc__high_prec_dec__small_rshift calls drop:
Function Before After Ratio
lshift 31174584 25762258 82.64% = 100% - 17.36%
rshift 21592604 20982502 97.17% = 100% - 2.83%
The time taken drops from 15.984s to 13.710s (85.77% = 100% - 14.23%).
To repeat, these are the improvements after disabling the Eisel-Lemire
and other fast paths. Simple Decimal Conversion is the rarely invoked
fallback parse_number_f64 algorithm. The practical improvement (with
Eisel-Lemire etc enabled) is smaller.
On a mid-range x86_64 laptop, processing the 181 MiB citylots.json file
from github.com/zemirco/sf-city-lots-json:
$ g++ -O3 script/process-json-numbers.c
$ time ./a.out -parse-number-f64 < citylots.json
Before/After:
real 0m0.992s
real 0m1.035s
On a mid-range x86_64 laptop, processing the 181 MiB citylots.json file
from github.com/zemirco/sf-city-lots-json:
$ g++ -O3 script/process-json-numbers.c
$ time ./a.out -parse-number-f64 < citylots.json
Before/After:
real 0m0.995s
real 0m0.968s
Processing citylots.json from github.com/zemirco/sf-city-lots-json
$ cc -O3 script/process-json-numbers.c
$ time ./a.out --parse-number-f64 < citylots.json
gcc 9.3 Before/After
real 0m0.638s
real 0m0.609s
clang 9.0 Before/After
real 0m0.692s
real 0m0.685s
x86_64 Broadwell desktop CPU.