[CALCITE-7722] RexSimplify IS [NOT] NULL on a safe operator with Strong policy ANY and unsafe operands can be further simplified by rubenada · Pull Request #5184 · apache/calcite

rubenada · 2026-08-17T14:24:32Z

Jira Link

Changes Proposed

RexSimplify.simplifyIsNotNull / simplifyIsNull currently bail out of the whole simplification when the input RexCall is not fully safe (i.e. isSafeExpression(a) == false). This is stricter than necessary for operators with Strong.Policy.ANY, where IS [NOT] NULL(f(x, y, ...)) is semantically equivalent to IS [NOT] NULL(x) OR/AND IS [NOT] NULL(y) OR/AND ... — the operator itself does not need to be evaluated to compute the result.

Example (regression for downstream projects such as Hive):

Before (≤ 1.34):
IS NOT NULL(CAST(key AS DOUBLE) + 1.0) → IS NOT NULL(CAST(key AS DOUBLE))

After (≥ 1.35):
IS NOT NULL(CAST(key AS DOUBLE) + 1.0) → (unchanged)

The rewrite is dropped because CAST(key AS DOUBLE) + 1.0 is a non-lossless cast wrapped in a +, so isSafeExpression returns false — even though + is Strong.ANY and the distribution is a valid rewrite regardless of the outer call's safety.

Proposed fix:

In the Strong.Policy.ANY branch, replace the full-tree safety requirement with a shallow safety check on the outer call. Because the branch rewraps the input as IS [NOT] NULL(operand_i) and joins the results with OR/AND, add a per-operand guard to prevent RexCall.isAlwaysTrue()/isAlwaysFalse() from silently collapsing a rewrapped IS [NOT] NULL(op) whose operand is typed non-nullable but not fully safe (which would otherwise erase a throwing subexpression such as 1 / 0).

Strong.Policy.NOT_NULL and CUSTOM continue to require full-tree safety, since those branches drop the subtree entirely.

… operations if they do not affect the nullability

mihaibudiu · 2026-08-17T17:20:30Z

    case ANY:
      // "f" is a strong operator, so "f(operand0, operand1) IS NOT NULL"
-      // simplifies to "operand0 IS NOT NULL AND operand1 IS NOT NULL"
+      // simplifies to "operand0 IS NOT NULL AND operand1 IS NOT NULL".


this comment could be shorter

comment simplified

mihaibudiu · 2026-08-17T17:20:48Z

+      if (!SafeRexVisitor.INSTANCE.isShallowSafe(a)) {
+        return simplifiedResult;
+      }
+      // The peel only preserves runtime semantics when each operand can be


and so can this one

comment simplified

mihaibudiu · 2026-08-17T17:21:20Z

    case ANY:
      // "f" is a strong operator, so "f(operand0, operand1) IS NULL" simplifies
-      // to "operand0 IS NULL OR operand1 IS NULL"
+      // to "operand0 IS NULL OR operand1 IS NULL". This branch PRESERVES the


And I think this one too

comment simplified

mihaibudiu · 2026-08-17T17:23:02Z

    checkSimplifyUnchanged(div(cast(vVarchar(), tInt(false)), nullInt));
  }

+  /** Test cases for peeling {@code IS [NOT] NULL} across a strong outer


Maybe you can follow the established pattern for test case JavaDoc, "Test case for [CALCITE-7722]" etc?

mihaibudiu · 2026-08-17T17:23:31Z

+    // IS NULL(CAST(10/0 AS BIGINT)) stays as IS NULL(10/0)
+    // after the lossless-CAST strip; the DIVIDE is not
+    // shallow-safe, so no further distribution occurs.
+    checkSimplify(isNull(cast(div(vIntNotNull(), literal(0)), tBigInt())),


can we have some tests using checked arithmetic or arithmetic on intervals?

good point; added both cases

julianhyde

-1

This change should use Strong.

It is frustrating when you log an issue and provide a PR at the same time. It removes the chance to have design discussions.

rubenada · 2026-08-17T17:46:45Z

@julianhyde the patch is actually using Strong. I have update the Jira to make it more accurate with the actual fix proposal

thomasrebele

I find the idea of "shallow-safety" interesting. It would be nice if we could support simplifying ((1/0)+1) IS NOT NULL.

Side note: This PR makes me think that there are some limitations in the current code for representing the properties of the operators (e.g., CALCITE-7264). I think there was a discussion somewhere whether the possibility that a RexNode may throw an exception could be included in the type system. The safety and shallow-safety of an operator would be attached to the operator itself instead of a visitor (see Julian's comment).

The concept of "safety" and "shallow-safety" could be collapsed, i.e., an operator is safe iff its evaluation does not throw an exception. A RexNode expression is safe iff all its operators are safe. Might be a bit easier to understand than "shallow-safety".

thomasrebele · 2026-08-18T09:16:27Z

+
+    // The outer PLUS is shallow-safe, but the div(1, 0) is not, so the peel is therefore suppressed
+    checkSimplifyUnchanged(isNotNull(plus(div(literal(1), literal(0)), vIntNotNull())));
+    checkSimplifyUnchanged(isNull(plus(div(literal(1), literal(0)), vIntNotNull())));


The previous version of the PR stated the reason for not simplifying this: otherwise it would be simplified to true because RexCall#isAlwaysTrue() returns true for IS NULL(1/0). This sounds like a bug in RexCall.isAlwaysTrue() to me. Could we make RexCall#isAlwaysTrue check for the safety of the IS [NOT] NULL operand? Potentially in a follow-up ticket if this turns out to be too complex.

Correct, the problem is that RexCall#isAlwaysTrue/isAlwaysFalse of an IS [NOT] NULL on an expression that is of a non-nullable type will always return a short-circuited true/false, without considering if the expression is safe, so we can hide a rutime issue; e.g. IS NOT NULL of a division by zero of type non-nullable would be collapsed to true. IMO this is a different improvement, I think this should be handled separately, as a follow-up ticket.

UPDATE: on a second thought, maybe if we include it in here we can simplify the current patch, checking....

UPDATE II: actually, fixing the RexCall isAlwaysTrue / isAlwaysFalse simplification for IS [NOT] NULL so that it considers isSafe makes the patch simpler, and enables more simplifications; so I've just applied it on the last commit. Thanks for pointing this out @thomasrebele !

…n for IS [NOT] NULL

sonarqubecloud · 2026-08-18T12:42:35Z

Quality Gate passed

Issues
1 New issue
0 Accepted issues

Measures
0 Security Hotspots
86.7% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

julianhyde · 2026-08-18T17:48:26Z

+1

(This overrides my earlier '-1'. Thank you to @rubenada for explaining in Jira the purpose of this change.)

julianhyde · 2026-08-18T17:55:24Z

@thomasrebele Yes, we could use better terminology. "Safe", "Strict" and "Strong" are parallel concepts, dealing with exceptions (non-termination), whether arguments get evaluated, and null values.

It's necessary to distinguish an operator's propagation characteristics (e.g. whether it returns null if and only if both its arguments are null) from the characteristics of an expression (e.g. whether it may return null, or may throw).

Even if you are able to find a good terminology that the community agrees on, implementing it is a challenge - you would have to modify the source code, potentially renaming classes, and deal with the fact that there are closed issues and commits that cannot be retrospectively changed.

[CALCITE-7722] Simplify IS NULL / IS NOT NULL expressions by removing…

4a3058c

… operations if they do not affect the nullability

rubenada mentioned this pull request Aug 17, 2026

HIVE-29641: Upgrade Calcite to 1.42.0 apache/hive#6523

Open

mihaibudiu reviewed Aug 17, 2026

View reviewed changes

julianhyde requested changes Aug 17, 2026

View reviewed changes

rubenada changed the title ~~[CALCITE-7722] Simplify IS NULL / IS NOT NULL expressions by removing operations if they do not affect the nullability~~ [CALCITE-7722] RexSimplify IS [NOT] NULL over a Strong.ANY call is no longer simplified when the call itself is unsafe Aug 17, 2026

rubenada changed the title ~~[CALCITE-7722] RexSimplify IS [NOT] NULL over a Strong.ANY call is no longer simplified when the call itself is unsafe~~ [CALCITE-7722] RexSimplify IS [NOT] NULL on a safe operator with Strong policy ANY and unsafe operands can be further simplified Aug 18, 2026

Improve tests, simplify comments

2953c94

thomasrebele reviewed Aug 18, 2026

View reviewed changes

Consider isSafe on RexCall isAlwaysTrue / isAlwaysFalse simplificatio…

52811f0

…n for IS [NOT] NULL

thomasrebele reviewed Aug 18, 2026

View reviewed changes

Comment thread core/src/test/java/org/apache/calcite/rex/RexProgramTest.java

Add clarification comment on test

73e650f

Uh oh!

Conversation

rubenada commented Aug 17, 2026

Jira Link

Changes Proposed

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

julianhyde left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rubenada commented Aug 17, 2026

Uh oh!

thomasrebele left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rubenada Aug 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

sonarqubecloud Bot commented Aug 18, 2026

Quality Gate passed

Uh oh!

julianhyde commented Aug 18, 2026

Uh oh!

julianhyde commented Aug 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

julianhyde left a comment •

edited

Loading

rubenada Aug 18, 2026 •

edited

Loading