This function is used in the named timezone ctor and was using QByteArray::split(),
followed by size checks and a linear scan for invalid chars per section. The use of
split() resulted in a lot of memory allocations and, unsurprisingly, bad performance.
The new code just performs one linear scan through the byte array, calculating
section sizes on the fly.
Benchmark results (with the test data in tst_QTimeZone::isValidId_data()) show
typical speedups of ~10x for valid IDs:
RESULT : tst_QTimeZone::isValidId_bench():"minimal middle":
- 0.00036 msecs per iteration (total: 95, iterations: 262144)
+ 0.000035 msecs per iteration (total: 74, iterations: 2097152)
Even in the sweet-spot case of the old code---a space character anywhere in the
string, checked for before the split---the new code is anywhere between slightly
faster and not much slower:
RESULT : tst_QTimeZone::isValidId_bench():"invalid char ' ' front":
- 0.000011 msecs per iteration (total: 94, iterations: 8388608)
+ 0.000010 msecs per iteration (total: 86, iterations: 8388608)
RESULT : tst_QTimeZone::isValidId_bench():"invalid char ' ' middle":
- 0.000014 msecs per iteration (total: 62, iterations: 4194304)
+ 0.000016 msecs per iteration (total: 69, iterations: 4194304)
RESULT : tst_QTimeZone::isValidId_bench():"invalid char ' ' back":
- 0.000018 msecs per iteration (total: 79, iterations: 4194304)
+ 0.000023 msecs per iteration (total: 98, iterations: 4194304)
This is not surprising, as the space character was singled out for a fast-exit
check before. For any other invalid character, the new version is anywhere from
15x to 35x faster:
RESULT : tst_QTimeZone::isValidId_bench():"invalid char ? front":
- 0.00034 msecs per iteration (total: 91, iterations: 262144)
+ 0.000010 msecs per iteration (total: 87, iterations: 8388608)
RESULT : tst_QTimeZone::isValidId_bench():"invalid char ? middle":
- 0.00036 msecs per iteration (total: 96, iterations: 262144)
+ 0.000016 msecs per iteration (total: 68, iterations: 4194304)
RESULT : tst_QTimeZone::isValidId_bench():"invalid char ? back":
- 0.00035 msecs per iteration (total: 94, iterations: 262144)
+ 0.000021 msecs per iteration (total: 92, iterations: 4194304)
If there was a deeper reason to single out the space character, that fast-exit
path can easily be restored.
This function is often used in conjunction with availableTimeZoneIds(), which
currently vastly dominates the runtime of the function calling both, but I'll
add another optimization for the common use-case of just checking for a time-zone's
existence in a subsequent commit.
Change-Id: Ife1d096fcd39464083ea464c23e49ad98fabf345
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
0e7100d8c0