[lworld] RFR: 8229897: [lworld] C1 should avoid allocation when reading a field from a flattened field

[Frederic Parain]

On Fri, 28 Aug 2020 14:12:21 GMT, Tobias Hartmann <thartmann at openjdk.org> wrote:

>> Please review this patch which removes temporary heap allocations when reading a field from a flattened field.
>> 
>> During the HIR graph construction, C1 detects sequences of consecutive getfield bytecodes involving flattened fields
>> and optimizes them with a single LoadField or NewInlineTypeInstance directly to the offset of the last nested field.
>> Performance improvement on a simple benchmark
>> 
>>     static inline class Point {
>>         int i = 0, j = 0;
>>     }
>>     static inline class Rectangle {
>>         Point p0 = new Point(), p1 = new Point();
>>     }
>>     static class NamedRectangle {
>>         Rectangle rect = new Rectangle();
>>         String name;
>>     }
>>     static NamedRectangle nr = new NamedRectangle();
>>     @Benchmark
>>     public void testGetfieldChain1(Blackhole blackhole) {
>>         int i = nr.rect.p0.j;                                                              // reads two flattened fields before
>>         reading an int blackhole.consume(i);
>>     }
>> 
>>     @Benchmark
>>     public void testGetfieldChain2(Blackhole blackhole) {
>>         Point p = nr.rect.p1;                                                          // reads a flattened field inside a
>>         flattened field blackhole.consume(p);
>>     }
>> 
>> Without the optimization:
>> Benchmark                             Mode  Samples  Score  Score error  Units
>> o.s.MyBenchmark.testGetfieldChain1    avgt      200  8.696        0.024  ns/op
>> o.s.MyBenchmark.testGetfieldChain2    avgt      200  8.969        0.027  ns/op
>> 
>> With the optimization:
>> Benchmark                             Mode  Samples  Score  Score error  Units
>> o.s.MyBenchmark.testGetfieldChain1    avgt      200  2.134        0.019  ns/op
>> o.s.MyBenchmark.testGetfieldChain2    avgt      200  5.150        0.017  ns/op
>> 
>> 
>> Tested with Mach5, tiers 1 to 3.
>> 
>> Thank you,
>> 
>> Fred
>
> Looks good to me!

Thanks Tobias!

Fred

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

PR: https://git.openjdk.java.net/valhalla/pull/167