RFR: 8341781: Improve Min/Max node identities

Chen Liang liach at openjdk.org
Thu Oct 10 05:16:14 UTC 2024


On Thu, 10 Oct 2024 02:59:14 GMT, Jasmine Karthikeyan <jkarthikeyan at openjdk.org> wrote:

> Hi all,
> This patch implements some missing identities for Min/Max nodes. It adds static type-based operand choosing for MinI/MaxI, such as the ones that MinL/MaxL use. In addition, it adds simplification for patterns such as `Max(A, Max(A, B))` to `Max(A, B)` and `Max(A, Min(A, B))` to `A`. These simplifications stem from the [lattice identity rules](https://en.wikipedia.org/wiki/Lattice_(order)#As_algebraic_structure). The main place I've seen this pattern is with MinL/MaxL nodes created during loop optimizations. Some examples of where this occurs include BigInteger addition/subtraction, and regex code. I've run some of the existing benchmarks and found some nice improvements:
> 
>                                                                 Baseline                    Patch
> Benchmark                                 Mode  Cnt       Score       Error  Units    Score       Error  Units  Improvement
> BigIntegers.testAdd                       avgt   15      25.096 ±     3.936  ns/op   19.214  ±    0.521  ns/op  (+ 26.5%)
> PatternBench.charPatternCompile           avgt    8     453.727 ±   117.265  ns/op   370.054 ±   26.106  ns/op  (+ 20.3%)
> PatternBench.charPatternMatch             avgt    8     917.604 ±   121.766  ns/op   810.560 ±   38.437  ns/op  (+ 12.3%)
> PatternBench.charPatternMatchWithCompile  avgt    8    1477.703 ±   255.783  ns/op  1224.460 ±   28.220  ns/op  (+ 18.7%)
> PatternBench.longStringGraphemeMatches    avgt    8     860.909 ±   124.661  ns/op   743.729 ±   22.877  ns/op  (+ 14.6%)
> PatternBench.splitFlags                   avgt    8     420.506 ±    76.252  ns/op   321.911 ±   11.661  ns/op  (+ 26.6%)
> 
> I've added some IR tests, and tier 1 testing passes on my linux machine. Reviews would be appreciated!

src/hotspot/share/opto/addnode.hpp line 270:

> 268:   virtual int Opcode() const = 0;
> 269:   virtual int max_opcode() const = 0;
> 270:   virtual int min_opcode() const = 0;

The old comment above 

// all the behavior of addition on a ring.  Only new thing is that we allow
// 2 equal inputs to be equal.

seems outdated.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/21439#discussion_r1794661997


More information about the hotspot-compiler-dev mailing list