[
https://issues.apache.org/jira/browse/CALCITE-3224?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16897886#comment-16897886
]
Feng Zhu commented on CALCITE-3224:
-----------------------------------
Yes, [~zabetak], I agree with you. It brings heavy work to review.
When I worked on this feature, I also found it is hard to make just "local"
changes on current code base.
On one hand, the new visitor-based manner returns Expression pair as result,
which contains more information. Different algorithm results it is difficult to
reuse current methods.
On the other hand, RexToLixTranslator (1000+ lines) and RexImpTable
(2000+lines) are "huge" class. For my personal point, separating some inner
classes and static functions seems to be more friendly and readability.
Let's get some feedbacks first and see if we can come up with solution.
> New RexNode-to-Expression CodeGen Implementation
> ------------------------------------------------
>
> Key: CALCITE-3224
> URL: https://issues.apache.org/jira/browse/CALCITE-3224
> Project: Calcite
> Issue Type: Bug
> Components: core
> Affects Versions: 1.20.0
> Reporter: Feng Zhu
> Assignee: Feng Zhu
> Priority: Critical
> Labels: pull-request-available
> Attachments: codegen.png
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> h3. *Background*
> Current RexNode-to-Expression implementation relies on BlockBuilder's
> incorrect “optimizations” to inline unsafe operations. As illustrated in
> CALCITE-3173, when this cooperation is broken in some special cases, it will
> cause exceptions like NPE, such as CALCITE-3142, CALCITE-3143, CALCITE-3150.
> Though we can fix these problems under current implementation framework
> with some efforts like the PR in CALCITE-3142, the logic will become more and
> more complex. To pursue a thorough and elegant solution, we implement a new
> one. Moreover, it also ensures the correctness for non-optimized code.
> h3. *Major Features*
> * *Visitor Pattern*: Each RexNode will be visited only once in a bottom-up
> way, rather than recursively visiting a RexNode many times with different
> NullAs settings.
> * *Conditional Semantic*: It can naturally guarantee the correctness even
> without BlockBuilder’s “optimizings”. Each line of code generated for a
> RexNode is null safe.
> * *Interface Compatibility*: The implementation only updates
> _RexToLixTranslator_ and _RexImpTable_. Interfaces such as CallImplementor
> keep unchanged.
> h3. *Implementation*
> For each RexNode, the visitor will generally generate two declaration
> statements, one for value and one for nullable. The code snippet is like:
> {code:java}
> {valueVariable} = {valueExpression}
> {isNullVariable} = {isNullExpression}
> {code}
> The visitor’s result will be the variable pair (*_isNullVariable_*,
> *_valueVariable_*).
> *Other changes:*
> (1) ReImplement different RexCall implementations (e.g., CastImplementor,
> BinaryImplementor and etc.) as seperated files and remove them into the newly
> created package _org.apache.calcite.adapter.enumerable.rex,_ and organize
> them in RexCallImpTable.
> (2) move some util functions into EnumUtils.
> h3. *Example Demonstration*
> Take a simple test case as example, in which the "commission" column is
> nullable.
> {code:java}
> @Test public void testNPE() {
> CalciteAssert.hr()
> .query("select \"commission\" + 10 as s\n"
> + "from \"hr\".\"emps\"")
> .returns("S=1010\nS=510\nS=null\nS=260\n");
> }
> {code}
> The codegen progress and non-optimized code are demonstrated in the figure
> below.
> !codegen.png!
> # When visiting *RexInputRef (commission)*, the visitor generates three
> lines of code, the result is a pair of ParameterExpression (*_input_isNull_*,
> *_input_value_*).
> # Then the visitor visits *RexLiteral (10)* and generates two lines of code.
> The result is (*_literal_isNull_*, *_literal_value_*).
> # After that, when visiting *RexCall(Add)*, (_*input_isNull*_,
> _*input_value*_) and (_*literal_isNull*_, _*literal_value*_) can be used to
> implement the logic. The visitor also generates two lines of code and return
> the variable pair.
> In the end, the result Expression is constructed based on
> (_*binary_call_isNull*_, _*binary_call_value*_)
--
This message was sent by Atlassian JIRA
(v7.6.14#76016)
