On Mon, 1 Nov 2021 11:23:16 GMT, Aleksey Shipilev <sh...@openjdk.org> wrote:
> This blocks JDK-8276215: `StrictMath` intrinsics are handled peculiarly by > giving failing intrinsics a second chance to match against the similar `Math` > intrinsics. This has interesting consequence for matchers: we can match the > native `StrictMath.sqrt` to non-native intrinsic for `Math.sqrt`. Interpreter > would then have to disambiguate the two. It could be made simpler and more > consistent. > > For `min`/`max` methods, `StrictMath` already delegates to `Math` methods, so > we can just drop the intrinsics for them. `sqrt` is harder to delegate, > because it is `native` and a part of public API, so we can instead do the > proper special intrinsic for it. > > There seem to be no performance regressions with this patch at least on Linux > x86_64: > > > $ CONF=linux-x86_64-server-release make test TEST="micro:StrictMathBench" > > Benchmark Mode Cnt Score Error Units > > ### Before > > StrictMathBench.minDouble thrpt 4 230921.558 ± 234.238 ops/ms > StrictMathBench.minFloat thrpt 4 230932.303 ± 126.721 ops/ms > StrictMathBench.minInt thrpt 4 230917.256 ± 73.008 ops/ms > StrictMathBench.minLong thrpt 4 194460.828 ± 178.079 ops/ms > > > StrictMathBench.maxDouble thrpt 4 230983.180 ± 161.211 ops/ms > StrictMathBench.maxFloat thrpt 4 230969.290 ± 277.500 ops/ms > StrictMathBench.maxInt thrpt 4 231033.581 ± 200.015 ops/ms > StrictMathBench.maxLong thrpt 4 194590.744 ± 114.295 ops/ms > > > StrictMathBench.sqrtDouble thrpt 4 230722.037 ± 2222.080 ops/ms > > ### After > > StrictMathBench.minDouble thrpt 4 230976.625 ± 67.338 ops/ms > StrictMathBench.minFloat thrpt 4 230896.021 ± 270.434 ops/ms > StrictMathBench.minInt thrpt 4 230859.741 ± 403.147 ops/ms > StrictMathBench.minLong thrpt 4 194456.673 ± 111.557 ops/ms > > StrictMathBench.maxDouble thrpt 4 230890.776 ± 89.924 ops/ms > StrictMathBench.maxFloat thrpt 4 230918.334 ± 63.160 ops/ms > StrictMathBench.maxInt thrpt 4 231059.128 ± 51.224 ops/ms > StrictMathBench.maxLong thrpt 4 194488.210 ± 495.224 ops/ms > > StrictMathBench.sqrtDouble thrpt 4 231023.703 ± 247.330 ops/ms > > > Additional testing: > - [x] `StrictMath` benchmarks > - [x] Linux x86_64 fastdebug `tier1` So we have _dsqrt and_dsqrt_strict, which must be functionally identical, but we provide both names because they're part of a public API. I think this deserves an explanatory comment in the code. ------------- PR: https://git.openjdk.java.net/jdk/pull/6184