RFR: 8324655: Identify integer minimum and maximum patterns created with if statements [v3]

[Emanuel Peter]

On Thu, 22 Feb 2024 16:33:13 GMT, Jasmine Karthikeyan <jkarthikeyan at openjdk.org> wrote:

>> Hi all, I've created this patch which aims to convert common integer mininum and maximum patterns created using if statements into Min and Max nodes. These patterns are usually in the form of `a > b ? a : b` and similar, as well as patterns such as `if (a > b) b = a;`. While this transform doesn't generally improve code generation it's own, it simplifies control flow and creates new opportunities for vectorization.
>> 
>> I've created a benchmark for the PR, and I've attached some data from my (Zen 3) machine:
>> 
>>                                                 Baseline                     Patch            Improvement
>> Benchmark                   Mode   Cnt  Score      Error  Units    Score       Error  Units
>> IfMinMax.testReductionInt   avgt   15  500.307 ±  16.687  ns/op    509.383 ±  32.645  ns/op   (no change)*
>> IfMinMax.testReductionLong  avgt   15  493.184 ±  17.596  ns/op    513.587 ±  28.339  ns/op   (no change)*
>> IfMinMax.testSingleInt      avgt   15    3.588 ±   0.540  ns/op      2.965 ±   1.380  ns/op   (no change)
>> IfMinMax.testSingleLong     avgt   15    3.673 ±   0.128  ns/op      3.506 ±   0.590  ns/op   (no change)
>> IfMinMax.testVectorInt      avgt   15  340.425 ±  13.123  ns/op     59.689 ±   7.509  ns/op   + 5.7x
>> IfMinMax.testVectorLong     avgt   15  326.420 ±  15.554  ns/op    117.190 ±   5.622  ns/op   + 2.8x
>> 
>> 
>> * After writing this benchmark I discovered that the compiler doesn't seem to create some simple min/max reductions, even when using Math.min/max() directly. Is this known or should I create a followup RFE for this?
>> 
>> The patch passes tier 1-3 testing on linux x64. Reviews or comments would be appreciated!
>
> Jasmine Karthikeyan has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains four additional commits since the last revision:
> 
>  - Merge master
>  - Apply changes from review
>  - Don't transform highly predictable branches
>  - Convert integer min/max patterns to Min/Max nodes

> But there are also secondary effects like computing both sides of cmov instead of one side of the branch, that would be hard to catch in benchmarks.

I guess you would have to create a branch that has either high latency (many chained ops) or that would saturate the CPU pipeline?

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

PR Comment: https://git.openjdk.org/jdk/pull/17574#issuecomment-1963904552