SQL: Early rejection of query with constant fail (e.g., division by zero)

[philrz]

The sqllogictests reveal a subtle difference in behavior where Postgres sniffs out the constant division by zero in this query at planning time:

$ psql postgres -c "SELECT DISTINCT + 31 + + NULLIF ( - CAST ( NULL AS REAL ), - + 93 + - 22 / CASE COUNT ( * ) WHEN - NULLIF ( + 11, - 22 * 52 + 20 / + COALESCE ( - COALESCE ( 0, - 31 / 65 ), - 8 ) ) THEN - COUNT ( * ) ELSE NULL END ) AS col1;"
ERROR:  division by zero

$ echo $?
1

Whereas SuperDB surfaces it during query execution, which also creates follow-on error values seemingly due to failed attempts to do things with the root "divide by zero" error:

$ super -version
Version: 292f4415e

$ super -c "SELECT DISTINCT + 31 + + NULLIF ( - CAST ( NULL AS REAL ), - + 93 + - 22 / CASE COUNT ( * ) WHEN - NULLIF ( + 11, - 22 * 52 + 20 / + COALESCE ( - COALESCE ( 0, - 31 / 65 ), - 8 ) ) THEN - COUNT ( * ) ELSE NULL END ) AS col1;"
{col1:error({message:"?-operator: bool predicate required",on:error({message:"type incompatible with unary '-' operator",on:error("divide by zero")})})}

$ echo $?
0

Details

Repro is with super commit 292f441.

It's my understanding that seeing this same result with EXPLAIN is confirmation that it's being sniffed out at planning time:

$ psql postgres -c "EXPLAIN SELECT DISTINCT + 31 + + NULLIF ( - CAST ( NULL AS REAL ), - + 93 + - 22 / CASE COUNT ( * ) WHEN - NULLIF ( + 11, - 22 * 52 + 20 / + COALESCE ( - COALESCE ( 0, - 31 / 65 ), - 8 ) ) THEN - COUNT ( * ) ELSE NULL END ) AS col1;"
ERROR:  division by zero

I became aware of this effect because it's a sqllogictest among several dozen that seemingly should have been hinted as skipif postgresql but was not. I've made a post in the SQLite Forum to see what others make of it. For our own testing purposes I intend to skip these kinds of queries in our nightly automation going forward.

I'm not necessarily advocating that we need match the Postgres behavior here, but figured I'd surface it in the event there's something to be learned or documented here.

[philrz]

Reacting to the follow-on errors, @mccanne noted that some work has been done in the semantic analysis to relate an expression node back to the AST. This would allow us to have the errors point to the spots in the original query text that triggered them.