RFR: 8346664: C2: Optimize mask check with constant offset [v10]

[Emanuel Peter]

On Thu, 30 Jan 2025 14:37:17 GMT, Matthias Ernst <duke at openjdk.org> wrote:

>> Fixes [JDK-8346664](https://bugs.openjdk.org/browse/JDK-8346664): extends the optimization of masked sums introduced in #6697 to cover constant values, which currently break the optimization.
>> 
>> Such constant values arise in an expression of the following form, for example from `MemorySegmentImpl#isAlignedForElement`:
>> 
>> 
>> (base + (index + 1) << 8) & 255
>> => MulNode
>> (base + (index << 8 + 256)) & 255
>> => AddNode
>> ((base + index << 8) + 256) & 255
>> 
>> 
>> Currently, `256` is not being recognized as a shifted value. This PR enables further reduction:
>> 
>> 
>> ((base + index << 8) + 256) & 255
>> => MulNode (this PR)
>> (base + index << 8) & 255
>> => MulNode (PR #6697)
>> base & 255 (loop invariant)
>> 
>> 
>> Implementation notes:
>> * I verified that the originating issue "scaled varhandle indexed with i+1"  (https://mail.openjdk.org/pipermail/panama-dev/2024-December/020835.html) is resolved with this PR.
>> * ~in order to stay with the flow of the current implementation, I refrained from solving general (const & mask)==0 cases, but only those where const == _ << shift.~
>> * ~I modified existing test cases adding/subtracting from the index var (which would fail with current C2). Let me know if would like to see separate cases for these.~
>
> Matthias Ernst has updated the pull request incrementally with one additional commit since the last revision:
> 
>   expect vectorization of i|7 instead of i&7
>   i&7 variant expect "= 0".

test/hotspot/jtreg/compiler/vectorization/TestPopulateIndex.java line 123:

> 121:         }
> 122:         checkResultExprWithIndex3();
> 123:     }

I'm afraid this will now vectorize on some platforms and not on others:
If the number of elements in the vector is <=8 then your optimization probably does not kick in, but with larger vectors, we would get `(i + 8) & 7` folded to `i & 7`, and only at this point does it not vectorize as expected...

In this case, I'm ok with just removing the IR rule. Splitting the IR rule is going to be a pain.

But let's leave a comment here that we might want to fix this in a future RFE.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/22856#discussion_r1935802287