Re: Bug in row_number() optimization

 

Yeah, you're right.  In this example the window function row_number()
goes into pass-through mode after the second evaluation because its
run condition does not hold true any more.  The remaining run would just
return the result from the second evaluation, which is stored in
econtext->ecxt_aggvalues[wfuncno].

If int8 is configured as pass-by-ref, the precomputed value from the
second evaluation is actually located in a memory area from context
ecxt_per_tuple_memory, with its pointer stored in ecxt_aggvalues.  As
this memory context is reset once per tuple, we would be prone to wrong
results.

I tried with memory context ecxt_per_query_memory when evaluating
window function in the case where int8 is configured as pass-by-ref and
I can see the problem vanishes.  I'm using the changes as below

--- a/src/backend/executor/nodeWindowAgg.c
+++ b/src/backend/executor/nodeWindowAgg.c
@@ -1027,8 +1027,14 @@ eval_windowfunction(WindowAggState *winstate, WindowStatePerFunc perfuncstate,
 {
        LOCAL_FCINFO(fcinfo, FUNC_MAX_ARGS);
        MemoryContext oldContext;
-
-       oldContext = MemoryContextSwitchTo(winstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory);
+       MemoryContext evalWfuncContext;
+
+#ifdef USE_FLOAT8_BYVAL
+       evalWfuncContext = winstate->ss.ps.ps_ExprContext->ecxt_per_tuple_memory;
+#else
+       evalWfuncContext = winstate->ss.ps.ps_ExprContext->ecxt_per_query_memory;
+#endif
+       oldContext = MemoryContextSwitchTo(evalWfuncContext);

Thanks
Richard