You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@calcite.apache.org by "Zoltan Haindrich (JIRA)" <ji...@apache.org> on 2018/10/08 21:27:00 UTC
[jira] [Comment Edited] (CALCITE-2615) When simplifying NOT-AND-OR,
RexSimplify incorrectly applies predicates deduced for operands to the same
operands
[ https://issues.apache.org/jira/browse/CALCITE-2615?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16642472#comment-16642472 ]
Zoltan Haindrich edited comment on CALCITE-2615 at 10/8/18 9:26 PM:
--------------------------------------------------------------------
I've found a smaller sample:
{code}
@Test
public void testSimplifyNotAnd3() {
final RexNode e = or(
le(
vBool(1),
literal(true)),
eq(
literal(false),
eq(literal(false), vBool(1))));
final String expected = "TODO";
checkSimplify(e, expected);
}
{code}
the following is happening:
* simpifyOrTerms adds predicate {{not(vBool1 <= true)}} => {{(vBool1 > true)}}
* I think the main problem is that the presence of this predicate also implies that {{vBool1 is NOT NULL}}
* the comparision {{=(false, vBool(1))}} is evaluated to false; because of range checks in simplifyUsingPredicates
* {{=(false, vBool1)}} becames false
** {{(vBool1 <= true) or (false = false)}}
* rest of the OR is discarded
I think the main problem is that simplifyUsingPredicates exploited that vBool1 may not be null.
I'm not sure why; but the unknownAs guard is missing for simplifyOrTerms call.
was (Author: kgyrtkirk):
I've found a smaller sample:
{code}
@Test
public void testSimplifyNotAnd3() {
final RexNode e = or(
le(
vBool(1),
literal(true)),
eq(
literal(false),
eq(literal(false), vBool(1))));
final String expected = "TODO";
checkSimplify(e, expected);
}
{code}
the following is happening:
* simpifyOrTerms adds predicate {{not(vBool1 <= true)}} => {{(vBool1 > true)}}
* I think the main problem is that the presence of this predicate also implies that {{vBool1 is NOT NULL}}
* the comparision {{=(false, vBool(1))}} is evaluated to false; because of range checks in simplifyUsingPredicates
* {{=(false, vBool1)}} becames false
** { (vBool1 <= true) or (false = false) }}
* rest of the OR is discarded
I think the main problem is that simplifyUsingPredicates exploited that vBool1 may not be null.
I'm not sure why; but the unknownAs guard is missing for simplifyOrTerms call.
> When simplifying NOT-AND-OR, RexSimplify incorrectly applies predicates deduced for operands to the same operands
> -----------------------------------------------------------------------------------------------------------------
>
> Key: CALCITE-2615
> URL: https://issues.apache.org/jira/browse/CALCITE-2615
> Project: Calcite
> Issue Type: Task
> Reporter: Julian Hyde
> Assignee: Julian Hyde
> Priority: Major
> Labels: newbie
>
> When simplifying NOT-AND-OR, RexSimplify incorrectly applies predicates deduced for operands to the same operands.
> Here is the test case (add it to RexProgramTest):
> {code:java}
> @Test public void testSimplifyNotAnd() {
> final RexNode e = not(
> and(
> gt(and(vBool(1), literal(true)),
> or(literal(true), literal(true), literal(false))),
> gt(isNotNull(vBool(0)), eq(literal(false), vBool(1))),
> or(ne(literal(true), literal(false)),
> ge(vInt(0), literal((Integer) null)))));
> final String expected = "TODO";
> checkSimplify(e, expected);
> }
> {code}
> When you run it, verify will find a combination of assignments such that the simplified expression returns a different result than the original.
> The test case is not minimal; sorry. Maybe it reproduces with NOT-OR.
> The bug is in [RexSimplify.simplifyAndTerms|https://github.com/apache/calcite/blob/6b3844c0634792263a5073b8ea93565fb3415f41/core/src/main/java/org/apache/calcite/rex/RexSimplify.java#L412]. Put a breakpoint at that line to see the problem. It passes through the operands once, building a list of predicates. Then it passes through the operands again, simplifying each operand. Thus operand1 is simplified using a list of predicates that includes the predicate 'not operand1'. Clearly wrong. [~kgyrtkirk], I warned that this was possible in our discussions.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)