NullCheckOp

[Thomas Wuerthinger]

Tom,

We can either introduce a new field in the TargetDescription class or disable all implicit null checks when TargetDescription.implicitNullCheckLimit equals 0.

- thomas

On 02 Jan 2014, at 23:16, Deneau, Tom <tom.deneau at amd.com> wrote:

> Thomas --
> 
> To overcome the problem mentioned below, I made the following one line change in GuardLoweringPhase.
> However, I am wondering if there should be a target-specific boolean indicating whether
> a target could support implicitNullChecks rather than just a single global option.
> That way, if you are compiling for two different targets (say cpu and gpu), one could
> use implicit null checks and one could use explicit.
> 
> -- Tom D.
> 
>     3.1 --- a/graal/com.oracle.graal.phases.common/src/com/oracle/graal/phases/common/GuardLoweringPhase.java	Tue Dec 24 16:03:03 2013 -0600
>     3.2 +++ b/graal/com.oracle.graal.phases.common/src/com/oracle/graal/phases/common/GuardLoweringPhase.java	Mon Dec 30 08:03:01 2013 -0600
>     3.3 @@ -143,7 +143,7 @@
>     3.4          protected void processNode(Node node) {
>     3.5              if (node instanceof GuardNode) {
>     3.6                  GuardNode guard = (GuardNode) node;
>     3.7 -                FixedWithNextNode lowered = guard.lowerGuard();
>     3.8 +                FixedWithNextNode lowered = (OptImplicitNullChecks.getValue() ? guard.lowerGuard() : null);
>     3.9                  if (lowered != null) {
>    3.10                      replaceCurrent(lowered);
>    3.11                  } else {
>    3.12                      lowerToIf(guard);
>    3.13                  }
> 
>> -----Original Message-----
>> From: Deneau, Tom
>> Sent: Saturday, December 28, 2013 9:00 AM
>> To: 'Thomas Wuerthinger'
>> Cc: graal-dev at openjdk.java.net; Christian Thalinger
>> Subject: RE: NullCheckOp
>> 
>> Thomas --
>> 
>> I tried -G:-OptImplicitNullChecks, it did not make any difference.
>> I think the problem is here:
>> 
>>    private static void processBlock(Block block, SchedulePhase
>> schedule, int implicitNullCheckLimit) {
>>        if (OptImplicitNullChecks.getValue() && implicitNullCheckLimit >
>> 0) {
>>            new
>> UseImplicitNullChecks(implicitNullCheckLimit).processNodes(block,
>> schedule);
>>        }
>>        new LowerGuards(block).processNodes(block, schedule);
>>    }
>> 
>> In that we still go thru LowerGuards.processNodes which eventually gets
>> to GuardNode.lowerGuard and the if block conditions are all met here so
>> the NullCheckNode is created...
>> 
>>    public FixedWithNextNode lowerGuard() {
>>        if (negated() && condition() instanceof IsNullNode) {
>>            IsNullNode isNull = (IsNullNode) condition();
>>            NullCheckNode nullCheck = graph().add(new
>> NullCheckNode(isNull.object()));
>>            setCondition(null);
>>            if (isNull.usages().isEmpty()) {
>>                isNull.safeDelete();
>>            }
>>            return nullCheck;
>>        }
>> 
>>        return null;
>>    }
>> 
>> I have confirmed that if I force GuardNode.lowerGuard to just return
>> null, the explicit null comparison and deopt node works correctly.
>> 
>> -- Tom
>> 
>> 
>> 
>>> -----Original Message-----
>>> From: Thomas Wuerthinger [mailto:thomas.wuerthinger at oracle.com]
>>> Sent: Saturday, December 28, 2013 6:41 AM
>>> To: Deneau, Tom
>>> Cc: graal-dev at openjdk.java.net; Christian Thalinger
>>> Subject: Re: NullCheckOp
>>> 
>>> Tom,
>>> 
>>> The flag "OptImplicitNullChecks" controls whether the compiler tries
>>> to create implicit null check instructions. If you set it to false,
>>> the code generated is equivalent to a comparison of the value against
>>> null and a deoptimization if the comparison yields true.
>>> 
>>> - thomas
>>> 
>>> On 24 Dec 2013, at 03:45, Christian Thalinger
>>> <christian.thalinger at oracle.com> wrote:
>>> 
>>>> I'm not an expert on this but it seems that a NullCheckOp (or a
>>> NullCheckNode) only does an implicit null check; there is no branch
>>> information attached to it.
>>>> 
>>>> In order to do an explicit exception you'd need to add nodes to the
>>> graph that throw the exception.  Not sure if something like this
>>> exists already.
>>>> 
>>>> On Dec 23, 2013, at 1:24 PM, Deneau, Tom <tom.deneau at amd.com> wrote:
>>>> 
>>>>> I am trying to get null checks working on the hsail backend.
>>>>> What is required of NullCheckOp in the HSAILLIRGenerator?
>>>>> On Hsail, we can't support an implicit exception for this like
>>>>> amd64
>>> does  so I think we would just want to end up with an explicit compare
>>> and branch if null to something that would be similar to the code
>>> emitted for a deoptimizing node.  How much of that does NullCheckOp
>>> have to do?
>>>>> 
>>>>> -- Tom
>>>> 
>>> 
> 
>