On 11/09/2020 11:23, Jason Tatton wrote: > For the x86 implementation there may be two further improvements we > can make in order to improve performance of both the StringUTF16 and > StringLatin1 indexOf(char) intrinsics: > > 1. Make use of AVX-512 instructions.
Is this really a good idea? When the processor detects Intel AVX instructions, additional voltage is applied to the core. With the additional voltage applied, the processor could run hotter, requiring the operating frequency to be reduced to maintain operations within the TDP limits. The higher voltage is maintained for 1 millisecond after the last Intel AVX instruction completes, and then the voltage returns to the nominal TDP voltage level. https://computing.llnl.gov/tutorials/linux_clusters/intelAVXperformanceWhitePaper.pdf So, if StringLatin1.indexOf(char) executes enough to make a difference to any real-world program, the effect may well be to slow down the clock for all of the code that does not use AVX. > 2. For “short” Strings (see below), I think it may be possible to modify the > existing algorithm to still use SSE SIMD > instructions instead of a loop. -- Andrew Haley (he/him) Java Platform Lead Engineer Red Hat UK Ltd. <https://www.redhat.com> https://keybase.io/andrewhaley EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671
