RFR: 8300109: RISC-V: Improve code generation for MinI/MaxI nodes [v2]

[Fei Yang]

On Mon, 16 Jan 2023 09:45:35 GMT, Ludovic Henry <luhenry at openjdk.org> wrote:

> I'm curious to see some JMH results on this change.

Well, not that much (%3 at most) when I run the max & min tests in JMH test [1] on my unmatched board:


Before:
Benchmark          (seed)   Mode  Cnt       Score      Error   Units
MathBench.maxInt        0  thrpt    8   97675.352 ± 4081.422  ops/ms
MathBench.maxLong       0  thrpt    8  105212.075 ± 5305.248  ops/ms
MathBench.minInt        0  thrpt    8  104321.539 ± 4342.072  ops/ms
MathBench.minLong       0  thrpt    8   97597.538 ± 3798.621  ops/ms

After:
Benchmark          (seed)   Mode  Cnt       Score      Error   Units
MathBench.maxInt        0  thrpt    8   97890.688 ± 6732.534  ops/ms
MathBench.maxLong       0  thrpt    8  107836.865 ± 3148.886  ops/ms
MathBench.minInt        0  thrpt    8  107521.081 ± 3992.511  ops/ms
MathBench.minLong       0  thrpt    8   97940.617 ± 3876.168  ops/ms

I think it will depend on the micro-architecture implementation. But the code generated should be always simplified in this case.

[1] https://github.com/openjdk/jdk/blob/master/test/micro/org/openjdk/bench/java/lang/MathBench.java#L301

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

PR: https://git.openjdk.org/jdk/pull/11988