Hello,
On Wed, Aug 16, 2023 at 1:27 PM Amit Langote <amitlangote09@gmail.com> wrote:
> I will post a new version after finishing working on a few other
> improvements I am working on.
Sorry about the delay. Here's a new version.
I found out that llvmjit_expr.c additions have been broken all along,
I mean since I rewrote the JsonExpr evaluation code to use soft error
handling back in January or so. For example, I had made CoerceiViaIO
evaluation code (EEOP_IOCOERCE ExprEvalStep) invoked by JsonCoercion
node's evaluation to pass an ErrorSaveContext to the type input
functions so that any errors result in returning NULL instead of
throwing the error. Though the llvmjit_expr.c code was not modified
to do the same, so the SQL/JSON query functions would return wrong
results when JITed. I have made many revisions to the JsonExpr
expression evaluation itself, not all of which were reflected in the
llvmjit_expr.c counterparts. I've fixed all that in the attached.
I've broken the parts to teach the CoerceViaIO evaluation code to
handle errors softly into a separate patch attached as 0001.
Other notable changes in the SQL/JSON query functions patch (now 0002):
* Significantly rewrote the parser changes to make it a bit more
readable than before. My main goal was to separate the code for each
JSON_EXISTS_OP, JSON_QUERY_OP, and JSON_VALUE_OP such that the
op-type-specific behaviors are more readily apparent by reading the
code.
* Got rid of JsonItemCoercions struct/node, which contained a
JsonCoercion field to store the coercion expressions for each JSON
item type that needs to be coerced to the RETURNING type, in favor of
using List of JsonCoercion nodes. That resulted in simpler code in
many places, most notably in the executor / llvmjit_expr.c.
--
Thanks, Amit Langote
EDB: http://www.enterprisedb.com
