On Fri, 15 Oct 2021 16:14:25 GMT, Andrew Haley <a...@openjdk.org> wrote:
>> Optimize the new Math.unsignedMultiplyHigh using the x86 mul instruction. >> This change show 1.87X improvement on a micro benchmark. > > src/hotspot/share/opto/mulnode.cpp line 468: > >> 466: } >> 467: >> 468: >> //============================================================================= > > MulHiLNode::Value() and UMulHiLNode::Value() seem to be identical. Perhaps > some refactoring would be in order, maybe make a common shared routine. Sure, will do the refactoring to use a shared routine. > test/micro/org/openjdk/bench/java/lang/MathBench.java line 547: > >> 545: return Math.unsignedMultiplyHigh(long747, long13); >> 546: } >> 547: > > As far as I can tell, the JMH overhead dominates when trying to measure the > latency of events in the nanosecond range. `unsignedMultiplyHigh` should have > a latency of maybe 1.5-2ns. Is that what you saw? Yes, the JMH overhead was dominating the measurement of latency. The latency observed for `unsignedMultiplyHigh` was 2.3ns with the intrinsic enabled. ------------- PR: https://git.openjdk.java.net/jdk/pull/5933
