RFR: 8347555: [REDO] C2: implement optimization for series of Add of unique value [v5]

[Kangcheng Xu]

On Tue, 4 Mar 2025 08:29:00 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> Kangcheng Xu has updated the pull request incrementally with one additional commit since the last revision:
>> 
>>   remove tri-conditionals
>
> This looks interesting! Thanks @tabjy for the work!
> 
> For now, I just have some drive-through comments about testing.
> 
> Also: would it make sense to have a JMH benchmark to prove that this code change is beneficial enough to warrant the additional complexity?

Added randomized tests per @eme64's suggestions. I'm still looking into micro benchmark with JMH

> test/hotspot/jtreg/compiler/c2/TestSerialAdditions.java line 291:
> 
>> 289:         return i + (x << i) + i; // Expects 64 + 63 + 64 = 191
>> 290:     }
>> 291: }
> 
> Would it make sense to add some randomized patterns, just for result verification?
> 
> You can use `Generators.java` to get interesting values. Of course that would mean not doing IR verification, but at least it would give us better confidence that the values are correct.
> 
> I'm imagining expressions like this:
> `return a * CON1 + a * CON2 + a * CON3 + a * CON4`
> 
> Where the CON are defined as a `public static final` field with a random value generated by `Generators`.
> 
> The advantage of using `Generators` is that it generates powers-of-two more frequently, which seems to be relevant here.

That's a very valid suggestion. Thanks. I updated the test. Please take a look if I understood your intention correctly.

However, it does look like this `a * CON1 + a * CON2 + ...` pattern is picked up by existing the associative optimization first:

https://github.com/openjdk/jdk/blob/a23fb0af65f491ef655ba114fcc8032a09a55213/src/hotspot/share/opto/addnode.cpp#L345-L377

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

PR Comment: https://git.openjdk.org/jdk/pull/23506#issuecomment-2705006453
PR Review Comment: https://git.openjdk.org/jdk/pull/23506#discussion_r1984090578