On Mon, Jan 4, 2021, at 04:49, Tom Lane wrote:
>Over the holiday break I've been fooling with some regex performance
>improvements.
Cool! I've also been fooling with regex performance over the years myself, not in the PostgreSQL code, but in general.
More specifically, to first DFA-minimize the regex,
and then to generate LLVMIR directly from the graph.
Perhaps some of the ideas could be interesting to look at.
One idea that I came up with myself is the "merge_linear" step,
where when possible, multiple characters are read in the same operation.
Not sure if other regex JIT engines does this, but it makes quite a difference
for large regexes where you have long strings.
Note: There is no support for capture groups, back-references, etc, but | + * () [] [^] works.
/Joel