Question on arithmetic overflow detection support in C2

[Krystal Mok]

Hi Remi,

On Tue, Jun 19, 2012 at 3:19 PM, Rémi Forax <forax at univ-mlv.fr> wrote:

> I'm not sure to understand why you need a pair of values knowing that
> methods like Math.addExact() throws an ArithmeticException if it overflows
> so in my opinion, only one value is needed.
>
> We need to model the implementation logic of Math.addExact() in C2's IR,
and we have to represent it with basic IR nodes. There's no node in this IR
that could "do addition and throw exception on overflow". We could
certainly make one, but then it'd be a macro node, and has to be expanded
(lowered) into the more basic nodes later on.

There's only one example I know of in C2 that carries very heavy logic all
the way into instruction selection, which is the STTNI optimizations for
some String methods.

As Vladimir and John already said, the IR in C2 usually produce only one
data out edge (in other words, produces only a single value). Data nodes
that produces multiple values are handled by projection nodes.


> I think the easy way is to check if their is a catch(ArithmeticException)
> that is
> transformed to a goto (the VM already does that) and to generate an add +
> a jump on overflow
> for that case.
>
> That's actually not as easy as it seems. Well, it's the matter of where to
put the logic into.


> BTW, I'm very happy and I suppose that Charles (Nutter) and the Jython
> team is happy too,
> that you guys take a look to that optimization.
>
> I'm very happy to find something new to work on, so that I can get up to
speed with C2. Hopefully I can get something running for you in time :-)

BTW, have you tried out Graal for custom optimizations?

- Kris


> cheers,
> Rémi
>
>
> On 06/19/2012 06:43 AM, Vladimir Kozlov wrote:
>
>> Vladimir Kozlov wrote:
>>
>>> It will be difficult. Almost all data nodes produce only one result and
>>>
>>
>> I meant only one data edge (not result) is produced.
>>
>>  this rule is used in all parts of C2 (from parser to RA). One exception
>>> is conditional stores which produce flag and memory state (look in
>>> macro.cpp in expand_allocate_common() method). This case handled specially
>>> everywhere. And you do need to add overflow and non-overflow pair to
>>> BoolTest values.
>>>
>>> Vladimir
>>>
>>> Krystal Mok wrote:
>>>
>>>> Hi all,
>>>>
>>>> I'm doing some experiments in C2, trying to add support in the IR for
>>>> arithmetic overflow detection. Looks like it's quite involved, though.
>>>> Any suggestions on how it could be done would be greatly appreciated :-)
>>>>
>>>> My questions:
>>>>
>>>> 1. The structure of IfNode: does the IR have to be in the form of
>>>> If_proj->If->Bool->Cmp
>>>>
>>>> Although there's code in idealize_test() (in ifnode.cpp) that checks if
>>>> iff->in(1)->is_Bool(), other places seem to assume the sole data input to
>>>> If is a Bool, e.g. the transformation in Matcher::find_shared().
>>>>
>>>> What if I wanted to add some node type other than Bool as the test
>>>> input to If? Something like (If (CheckOverflow expr)), does that feel right?
>>>>
>>>> 2. The BoolTest kinds, does it feel right to add a pair of kinds for
>>>> overflow? Like BoolTest::ovf and BoolTest::noovf. They'll still fit in the
>>>> 3-bit encoding.
>>>>
>>>> If this pair is added, then to make it useful, the input to Bool
>>>> wouldn't always be Cmp anymore; instead it could be any arithmetic nodes
>>>> that modifies the condition flags as a side-effect.
>>>> And...that doesn't seem to fit in C2's IR right now, since the only
>>>> kind of node that models DEF of the condition flags are the Cmp nodes.
>>>>
>>>> 3. As a quick-n-dirty experiment, I tried not to touch the
>>>> machine-independent IR, and instead modify only the ad file to pattern
>>>> match the current Java implementation of Match.addExact(int, int) to make
>>>> it emit a jo (jump-if-overflow) instruction for overflow detection, but it
>>>> failed to compile.
>>>> The patch is here: https://gist.github.com/**3818aa3acbc78e28e7fa<https://gist.github.com/3818aa3acbc78e28e7fa>
>>>>
>>>> (Aside: the Ideal graph of the overflow check expression is like this:
>>>> http://dl.iteye.com/upload/**picture/pic/114476/674e0a78-**
>>>> 5515-3da3-8ea6-6eedbeb32f8e.**png<http://dl.iteye.com/upload/picture/pic/114476/674e0a78-5515-3da3-8ea6-6eedbeb32f8e.png>
>>>> This is the state right before matching the matcher rules)
>>>>
>>>> ADLC didn't complain, but the generated jmpAddOverflowNode::Expand()
>>>> function had compilation errors, saying:
>>>>
>>>> ../generated/adfiles/ad_x86_**64_expand.cpp: In member function
>>>> ‘virtual MachNode* jmpAddOverflowNode::Expand(**State*, Node_List&,
>>>> Node*)’:
>>>> ../generated/adfiles/ad_x86_**64_expand.cpp:5915: error: ‘num8’ was
>>>> not declared in this scope
>>>>
>>>> Apparently I was doing something wrong in the ad file, but I don't
>>>> quite get what it was.
>>>>
>>>> Regards,
>>>> Kris
>>>>
>>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20120619/52891dcc/attachment-0001.html