]> granicus.if.org Git - postgresql/commit
Fix comparison of similarity to threshold in GIST trigram searches.
authorTom Lane <tgl@sss.pgh.pa.us>
Mon, 20 Jun 2016 14:49:19 +0000 (10:49 -0400)
committerTom Lane <tgl@sss.pgh.pa.us>
Mon, 20 Jun 2016 14:49:19 +0000 (10:49 -0400)
commit9c852566a3cf4ede40e22e4ca216d12cd4a27117
tree9cae7919739b245f16d5bbda6cf2ac0e75fe59c0
parent47981a4665dfc9c3808366dff9971daedba32e98
Fix comparison of similarity to threshold in GIST trigram searches.

There was some very strange code here, dating to commit b525bf77, that
purported to work around an ancient gcc bug by forcing a float4 comparison
to be done as int instead.  Commit 5871b8848 broke that when it changed
one side of the comparison to "double" but left the comparison code alone.
Commit f576b17cd doubled down on the weirdness by introducing a "volatile"
marker, which had nothing to do with the actual problem.

Guess that the gcc bug, even if it's still present in the wild, was
triggered by comparison of float4's and can be avoided if we store the
result of cnt_sml() into a double before comparing to the double "nlimit".
This will at least work correctly on non-broken compilers, and it's way
more readable.

Per bug #14202 from Greg Navis.  Add a regression test based on his
example.

Report: <20160620115321.5792.10766@wrigleys.postgresql.org>
contrib/pg_trgm/expected/pg_trgm.out
contrib/pg_trgm/sql/pg_trgm.sql
contrib/pg_trgm/trgm_gist.c