db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mamta Satoor <msat...@gmail.com>
Subject Re: [PATCH] Derby-479 Passing the return of a RETURN NULL ON NULL INPUT function to another function call throws linkage error.
Date Tue, 14 Feb 2006 18:20:09 GMT
Please find my replies inline. Hopefully they answer your questions.

On 2/13/06, Daniel John Debrunner <djd@apache.org> wrote:

> Mamta Satoor wrote:
> > Hi,
> >
> > I have a patch for this bug. The patch description in brief is as
> follows.
>
> Thanks for the full description.
>
> >
> > I have taken an example to explain the patch. Consider following sql
> > function.(Notice it is defined with RETURNS NULL ON NULL INPUT)
> >
> > CREATE FUNCTION RN_abs(A int) RETURNS int
> > EXTERNAL NAME 'java.lang.Math.abs(int)'
> > RETURNS NULL ON NULL INPUT
> > LANGUAGE JAVA PARAMETER STYLE JAVA;
> > VALUES CAST( RN_abS(RN_abs(90)) AS INTEGER);
> >
> > First let's start out by a VALUES query that does work fine.
> > VALUES RN_abs(90)
> >
> > During the bind phase of VALUES RN_abs(90), following nodes are created
> > JavaToSQL ( StaticMethodCall (SQLToJava (NumericConstantNode for
> > constant 90) ) )
> >
> > During code generation, StaticMethodCallNode, for each of the
> > parameters, calls generateOneParameter. In this eg, there is only one
> > parameter and the node for that parameter is SQLToJavaValueNode. This
> > node first generates the SQL value by calling generate on
> > NumericConstantNode. NumericConstantNode generates code to get
> > NumberDataValue for constant 90. This is the right type for the SQL
> > domain but since this needs to be passed into Java domain, we need to do
> > some casting. The java domain value would be to get a primitive int from
> > NumberDataValue. To achieve this, SQLToJavaValueNode in it's
> > generateJavaValue generates the code to get primitive value from
> > NumberDataValue. In this case, it will be generating code to call
> > getInt() on NumberDataValue so that we have the correct type for the
> > java domain.
> >
> > StaticMethodCallNode after generating each parameter checks if the
> > argument type to SQL function is same
> > as the parameter type to the Java method. In this case, the SQL argument
> > type is java.lang.Integer and Java method parameter type is int. Because
> > of this, it casts the item on the top of the stack to Java method
> > parameter type.
>
> I got lost in this paragraph. What do you mean by 'the SQL argument type
> is java.lang.Integer ...'? In the previous paragraph you said the
> SQLToJavaValueNode got the primitive directly from the NumberDataValue,
> so I'm lost as to where java.lang.Integer comes into it.
> Could you maybe point to the methods/line numbers where you think this
> is happening?


My comment was based on code MethodCallNode.generateParameters() line
numbers 535 through 549 as shown below

   // type from the SQL-J expression
   String argumentType = getParameterTypeName( methodParms[param] );

   // type of the method
   String parameterType = expectedTypes[param];

   if (!parameterType.equals(argumentType))
   {
    // since we reached here through method resolution
    // casts are only required for primitive types.
    // In any other case the expression type must be assignable
    // to the parameter type.
    if (classInspector.primitiveType(parameterType)) {

     mb.cast(parameterType);

    } else {

In the same MethodCallNode class, getParameterTypeName is defined as follows

 static public String getParameterTypeName( JavaValueNode param )
  throws StandardException
 {
  String argumentType;

  // RESOLVE - shouldn't this logic be inside JavaValueNode ??
  // I.e. once the value is primitive then its java type name is its
  // primitive type name.
  if (param.isPrimitiveType()) { argumentType = param.getPrimitiveTypeName();
}
  else { argumentType = param.getJavaTypeName(); }

  return argumentType;
 }

So, getParameterTypeName returns the type as "java.lang.Integer" to
generateParameters at line 536. generateParameters compares this type with
the type of the Java method which is "int", line 541. Since they don't
match, on line 549, it does casting of the top of the stack to the Java
method type. The inside StaticMethodCallNode finds the top of stack to be
"int" type already (because NumberDataValue's  returns int as it's Java
primitive value) and hence the casting done on 549 for inside
StaticMethodCallNode, although a no-op, is not a problem. But in the case of
the outside StaticMethodCallNode, on line 549, it finds that the top of the
stack is of the type java.lang.Integer (because of return null on null
input), the casting on line 549 fails for outside StaticMethodCallNode
because it is trying to java.lang.Integer to int. I hope this explanation is
helpful.

> I thought the generated code was something like this for a parameter in
> an returns null on null input function.
>
>   DataValueDescriptor pdvd = <somthing>
>
>   if (pdvd.isNull())
>       forceReturnNull = true; // actually generated field name
>   else
>       // needed due to limitation in byte code compiler
>       forceReturnNull = forceReturnNull;
>
>   pdvd.getInt();
>
> See SQLToJavaValueNode.generateReturnsNullOnNullCheck()


You are right about what gets generated, which is primitive int type, at
this point. But the conversion of int to java.lang.Integer happens after all
the parameters code generation is finished.
StaitcMethodCallNode.generationExpression line numbers 892 through 922 and
lines 928 through 942 convert int type to java.lang.Integer if sql function
is defined with return null on null input. And that is why we end up with
java.lang.Integer.


> > After taking care of parameter code generation, StaticMethodCallNode
> > checks if the sql function is defined to return null if any of its
> > arguments are null. This is true for our specific sql function. To take
> > care of this, StaticMethodNode needs to have a return type of Object
> > rather than primitive int type. After generating code to call
> > java.lang.Math.abs, it generates the code to return java.lang.Math.abs's
> > return value as java.lang.Integer This is where VALUES (
> > RN_abS(RN_abs(90))) runs into problem.
> >
> > So, now let's look at the problem sql.
> > VALUES ( RN_abS(RN_abs(90)));
> >
> > In the bind phase, following nodes are created for VALUES (
> > RN_abS(RN_abs(90)))
> > JavaToSQL ( StaticMethodCall ( StaticMethodCall (SQLToJava
> > (NumericConstantNode for constant 90) ) )  )
>
> Just to check, this is really the case, the outer node StaticMethodCall
> wraps the inner one directly? Have the SQL2Java and Java2SQL nodes that
> were generated been removed? I think this might be the problem, they
> should not have been removed, because the returns null on null input
> really means that part of the function call executes in the SQL domain,
> and cannot be short cut by using its return in a direct call to Java
> method.


Yes, during the parsing phase, MethodCallNode.addParams(Vector) on lines 217
through 220 removes SQL2Java and Java2SQL and your guess is right about that
being the problem. When I removed that optimization by commenting out lines
217 through 221, the VALUES (RN_abS(RN_abs(90))) runs fine. I ran derbyall
with this change (and removed by changes that I had proposed in my earlier
mail) and the tests ran fine.

> In generate phase, SQLToJavaValueNode generates code for SQL domain
> NumberDataValue first. But since the value needs to be returned to Java
> domain, it generates code to call getInt on NumberDataValue to get int
> for the Java domain. The inside StaticMethodCallNode finds that SQL
> function argument type is not same as Java parameter type and hence it
> generates code to cast the stack value to Java method parameter type.
Same as above, I can't see this in the code.

> StaticMethodNode then generates the code to call java.lang.Math.abs and
> it converts the int type of the Java method return value into
> java.lang.Integer because the sql function has been defined to return
> null on null input. In order to be able to return null, the return type
> can't be primitive, it has to be an Object which in this case is
> java.lang.Integer. So far the code generation is same as for simple
> VALUES RN_abs(90). The problem happens during code generation for
> outside StaticMethodCallNode. Outside StaticMethodCallNode is expecting
> an int type but inside StaticMethodNode returned java.lang.Integer to
> cover return null on null state definition of the sql function. This
> type mismatch is what causes the linkage exception.

The question is why is the StaticMethodNode expecting an int type when
the return type is Integer?

> I have resolved this type mismatch for outside sql function call by
> retrieving int value from java.lang.Integer with the help of the
> DataValueFactory, as shown below (I have added the code for this in
> MethodCallNode.generateParameters .

What happens when the return from the inner function was null with your
change? Is the outer method called? I think you need some tests with
this case. I think the patch does not fix the problem.

Dan.

> The new code has !!! in front of
> them in following code)
>    if (!parameterType.equals(argumentType))
>    {
>     // since we reached here through method resolution
>     // casts are only required for primitive types.
>     // In any other case the expression type must be assignable
>     // to the parameter type.
>     if (classInspector.primitiveType(parameterType)) {
> !!!     //The type on the stack for this parameter will not be primitive
> if this parameter
> !!!     //is a call to another sql function which is defined to return
> null on null input.
> !!!     //To take care of such a case, first generate the code to
> extract primitive type
> !!!     //from the value on the top of the stack. Derby479. eg
> !!!     //CREATE FUNCTION RN_abs(A int) RETURNS int
> !!!     //EXTERNAL NAME 'java.lang.Math.abs(int)'
> !!!     //RETURNS NULL ON NULL INPUT
> !!!     //LANGUAGE JAVA PARAMETER STYLE JAVA;
> !!!     //VALUES ( RN_abS(RN_abs(90)));
> !!!     //Inside RN_abs is returning a java.lang.Math so that it can
> return null if the argument
> !!!     //to it is null. But inside RN_abS is expecting primitive type
> int as input argument.
> !!!     //Following piece of code extract primitive value from the
> Object value.
> !!!     TypeId temp =
> methodParms[param].getJSQLType().getSQLType().getTypeId();
> !!!     acb.generateDataValue(mb, getTypeCompiler(temp), (LocalField)
null);
> !!!     mb.callMethod(VMOpcode.INVOKEINTERFACE,
> ClassName.DataValueDescriptor ,
> !!!       getTypeCompiler(temp).getPrimitiveMethodName(),
> !!!       getTypeCompiler(temp).getCorrespondingPrimitiveTypeName(), 0);
>      mb.cast(parameterType);
>     } else {
> Because of these changes to MethodCallNode, mb.cast doesn't run into
> casting exception anymore. I have attached the code changes to the JIRA
> entry.
>
> svn stat
> M
> java\engine\org\apache\derby\impl\sql\compile\ActivationClassBuilder.java
> M      java\engine\org\apache\derby\impl\sql\compile\MethodCallNode.java
> M      java\engine\org\apache\derby\impl\sql\compile\CursorNode.java
>
M      java\engine\org\apache\derby\impl\sql\compile\SQLToJavaValueNode.java
> M
>
java\testing\org\apache\derbyTesting\functionTests\tests\lang\functions.sql
> M
> java\testing\org\apache\derbyTesting\functionTests\master\functions.out
>
> I ran derbyall on my Windows XP with Sun's jdk14 and everything ran fine.
>
> Please send any feedback you might have.
>
> Mamta
>

Mime
View raw message