RFR: 8317121: vector_masked_load instruction is moved too early after JDK-8286941 [v2]
Aleksey Shipilev
shade at openjdk.org
Wed Oct 11 17:11:33 UTC 2023
On Tue, 3 Oct 2023 16:10:27 GMT, Vladimir Kozlov <kvn at openjdk.org> wrote:
>> [JDK-8286941](https://bugs.openjdk.org/browse/JDK-8286941) changed how masked vector operations are created and processed. It calls `LoadNode::Ideal()` which tries to find best place in code for it.
>> Unfortunately some places do not expect to have vector access to modified memory.
>> In `LoadNode::find_previous_arraycopy()` the call to `ArrayCopyNode::modifies()` assumes that it access only one element and did not take into account that a vector may access wider memory.
>> The same issue exists with Unsafe mismatching (bigger size) access.
>>
>> The fix is to take into account size of memory access.
>>
>> New regression test is added which show the issue.
>>
>> Tested tier1-5. xcomp, stress. Testing tier6-7 in progress.
>
> Vladimir Kozlov has updated the pull request incrementally with one additional commit since the last revision:
>
> Be more specific in exception message which test failed
I think the original JBS issue got closed somehow, could it be open again?
https://bugs.openjdk.org/browse/JDK-8317121
-------------
PR Comment: https://git.openjdk.org/jdk/pull/16015#issuecomment-1758130307
More information about the hotspot-compiler-dev
mailing list