[External] : Re: What is the purpose of loop predication lower bounds check

[Cao, Joshua]

Yes, I agree that an array can be zero sized.

I still do not see a reason to check `data.length <= 0`.
`data.length u<= iters - 1`, which is the upper bounds check, should be
sufficient to determine whether we should trap.


________________________________
From: Vladimir Kozlov <vladimir.kozlov at oracle.com>
Sent: Thursday, February 1, 2024 12:02:15 PM
To: Cao, Joshua; hotspot-compiler-dev at openjdk.org
Subject: RE: [EXTERNAL] [External] : Re: What is the purpose of loop predication lower bounds check

CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.



In general we usually generate different checks during parsing and
optimize/collapse/remove/merge them later based on full information
after parsing.

On 2/1/24 11:55 AM, Vladimir Kozlov wrote:
> If array is external it could be 0 sized.
> With locally allocated array the check should collapse unless we missing
> such optimization.
>
> On 2/1/24 11:52 AM, Cao, Joshua wrote:
>> If I use the same example but with negative stride:
>>
>>
>> ```
>>
>> for (int i = iters - 1; i >= 0; --i) {
>>
>>    blackhole(data[i]);
>>
>> }
>> ```
>>
>> We still generate these two checks:
>> ```
>>
>> if (data.length <= 0) { trap() }
>> if (data.length <= iters - 1) { trap() }
>>
>> ```
>>
>>
>> So even with negative stride, the upper bounds check is sufficient by
>> itself.
>>
>>
>> I am playing around with stride/scale/offset and I cannot come up with an
>>
>> example where the lower bounds check is important.
>>
>> ------------------------------------------------------------------------
>> *From:* hotspot-compiler-dev <hotspot-compiler-dev-retn at openjdk.org>
>> on behalf of Vladimir Kozlov <vladimir.kozlov at oracle.com>
>> *Sent:* Thursday, February 1, 2024 11:44:29 AM
>> *To:* hotspot-compiler-dev at openjdk.org
>> *Subject:* RE: [EXTERNAL] What is the purpose of loop predication
>> lower bounds check
>> CAUTION: This email originated from outside of the organization. Do
>> not click links or open attachments unless you can confirm the sender
>> and know the content is safe.
>>
>>
>>
>> Initial index value is variable.
>>
>> Reverse (negative stride) processing of array?
>>
>> On 2/1/24 11:31 AM, Cao, Joshua wrote:
>>> Can I get some help with https://bugs.openjdk.org/browse/JDK-8325146?
>> <https://bugs.openjdk.org/browse/JDK-8325146?>
>>> What is
>>> the purpose of the lower bounds check for loop predication? It seems
>>> like an
>>> unnecessary if-check for any example I can come up with.
>>>
>>> Whatever the reason might be, I think there are cases in which we can
>>> ellide
>>> the lower bounds check, like in the example in the JBS issue.
>>>
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-compiler-dev/attachments/20240201/8c29a4d4/attachment.htm>