diff options
author | David Rowley <drowley@postgresql.org> | 2020-11-24 12:04:42 +1300 |
---|---|---|
committer | David Rowley <drowley@postgresql.org> | 2020-11-24 12:04:42 +1300 |
commit | 913ec71d682e99852cc188a7edbdccd02d42b4b3 (patch) | |
tree | 201c5e201df913efd8a9a13a45dfb69523cbf3a7 /src/backend/access/transam/xlogutils.c | |
parent | 697e1d02f53f985992a4f479b0b08fca5c272c6c (diff) | |
download | postgresql-913ec71d682e99852cc188a7edbdccd02d42b4b3.tar.gz postgresql-913ec71d682e99852cc188a7edbdccd02d42b4b3.zip |
Improve compiler code layout in elog/ereport ERROR calls
Here we use a bit of preprocessor trickery to coax supporting compilers
into laying out their generated code so that the code that's in the same
branch as elog(ERROR)/ereport(ERROR) calls is moved away from the hot
path. Effectively, this reduces the size of the hot code meaning that it
can sit on fewer cache lines.
Performance improvements of between 10-15% have been seen on highly CPU
bound workloads using pgbench's TPC-b benchmark.
What's achieved here is very similar to putting the error condition inside
an unlikely() macro. For example;
if (unlikely(x < 0))
elog(ERROR, "invalid x value");
now there's no need to make use of unlikely() here as the common macro
used by elog and ereport will now see that elevel is >= ERROR and make use
of a pg_attribute_cold marked version of errstart().
When elevel < ERROR or if it cannot be determined to be constant, the
original behavior is maintained.
Author: David Rowley
Reviewed-by: Andres Freund, Peter Eisentraut
Discussion: https://postgr.es/m/CAApHDvrVpasrEzLL2er7p9iwZFZ%3DJj6WisePcFeunwfrV0js_A%40mail.gmail.com
Diffstat (limited to 'src/backend/access/transam/xlogutils.c')
0 files changed, 0 insertions, 0 deletions