I profiled this in production (on mwdebug1001, using HHVM and dumping the first 40,000 entities from shard 0/4):

Command line:
sudo -u www-data php /srv/mediawiki/multiversion/MWScript.php extensions/Wikidata/extensions/Wikibase/repo/maintenance/dumpJson.php --wiki wikidatawiki --sharding-factor 4 --shard 0 --snippet --limit 40000 --profiler=text

Result extract:
43.96% 713097.698 814618 - Wikibase\DataModel\Deserializers\StatementDeserializer::deserialize

Given our dumps (all shards combined) have a average run time of about 65 hours, we spent almost 29 hours each time in this function… a small saving of 5% would save us more than 100 minutes for each dump run.

I did another test with the unmodified class (on mwdebug1002):
43.58% 727005.407 815558 - Wikibase\DataModel\Deserializers\StatementDeserializer::deserialize (total time 1668083.546ms)

With the optimized script (same command line, on mwdebug1001):
41.61% 692661.326 815574 - Wikibase\DataModel\Deserializers\StatementDeserializer::deserialize (total time 1664704.396ms)

Both scripts ran at (about) the same time and have a very close total run time (especially if you count out the changes in deserialize).

Considering these numbers, we get a speedup of about 5% with the modified version.

Another run after a few more minor modifications (mwdebug1001 and mwdebug1002 side by side, both using db1082, command line as above):

Optimized (mwdebug1001):
43.46% 651201.869 815697 - Wikibase\DataModel\Deserializers\StatementDeserializer::deserialize (total time 1498537.701ms)
Old (mwdebug1002):
44.27% 727679.741 815698 - Wikibase\DataModel\Deserializers\StatementDeserializer::deserialize (total time 1643905.307ms)

That makes a speedup of 11.7% for deserialize, but the fact that the job on mwdebug1001 was 4.4% faster (the time saved by the mentioned optimization counted out) than the one on mwdebug1002 should be considered.

Considering T157380#3009442 the impact of this is potentially way bigger than anticipated above (if the database interaction time is reduced by such a factor). I didn't benchmark it, though.