RFR: 8328528: C2 should optimize long-typed parallel iv in an int counted loop [v24]

[Kangcheng Xu]

On Mon, 21 Oct 2024 06:07:14 GMT, Christian Hagedorn <chagedorn at openjdk.org> wrote:

>> Kangcheng Xu has updated the pull request incrementally with one additional commit since the last revision:
>> 
>>   update comment pseudo code, improve readability with explicit skip
>
> src/hotspot/share/opto/loopnode.cpp line 3973:
> 
>> 3971: // conversions are required:
>> 3972: //
>> 3973: //     long iv2 = ((long) phi * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))
> 
> Thanks for updating the example. I guess we should use `iv` consistently - missed that to do in my example before:
> Suggestion:
> 
> //    int a = init2;
> //    for (int iv = init; iv < limit; iv += stride_con) {
> //      a += stride_con2;
> //    }
> //
> // and transforms it to:
> //
> //    int iv2 = init2
> //    int iv = init
> //    loop:
> //      if ( iv >= limit ) goto exit
> //      iv += stride_con
> //      iv2 = init2 + (iv - init) * (stride_con2 / stride_con)
> //      goto loop
> //    exit:
> //    ...
> //
> // Such transformation introduces more optimization opportunities. In this
> // particular example, the loop can be eliminated entirely given that
> // `stride_con2 / stride_con` is exact  (i.e., no remainder). Checks are in
> // place to only perform this optimization if such a division is exact. This
> // example will be transformed into its semantic equivalence:
> //
> //     int iv2 = (iv * stride_con2 / stride_con) + (init2 - (init * stride_con2 / stride_con))
> //
> // which corresponds to the structure of transformed subgraph.
> //
> // However, if there is a mismatch between types of the loop and the parallel
> // induction variable (e.g., a long-typed IV in an int-typed loop), type
> // conversions are required:
> //
> //     long iv2 = ((long) iv * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))

Sorry I thought you were only refering to transform code in that last review. Yes, it's better to keep names consistent across. Thanks for pointing out!

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

PR Review Comment: https://git.openjdk.org/jdk/pull/18489#discussion_r1808176476