RFR: 8269559: AArch64: Implement string_compare intrinsic in SVE

[Nick Gasson]

On Mon, 16 Aug 2021 20:59:55 GMT, TatWai Chong <github.com+78814694+tatwaichong at openjdk.org> wrote:

> This patch implements string_compare intrinsic in SVE.
> It supports all LL, LU, UL and UU comparisons.
> 
> As we haven't found an existing benchmark to measure performance impact,
> we created a benchmark derived from the test [1] for this evaluation.
> This benchmark is attached to this patch.
> 
> Besides, remove the unused temporary register `vtmp3` from the existing
> match rules for StrCmp.
> 
> The result below shows all varients can be benefited largely.
> Command: make exploded-test TEST="micro:StringCompareToDifferentLength"
> 
> Benchmark    (size)  Mode  Cnt   Score Speedup Units
> compareToLL      24  avgt   10      1.0x       ms/op
> compareToLL      36  avgt   10      1.0x       ms/op
> compareToLL      72  avgt   10      1.0x       ms/op
> compareToLL     128  avgt   10      1.4x       ms/op
> compareToLL     256  avgt   10      1.8x       ms/op
> compareToLL     512  avgt   10      2.7x       ms/op
> compareToLU      24  avgt   10      1.6x       ms/op
> compareToLU      36  avgt   10      1.8x       ms/op
> compareToLU      72  avgt   10      2.3x       ms/op
> compareToLU     128  avgt   10      3.8x       ms/op
> compareToLU     256  avgt   10      4.7x       ms/op
> compareToLU     512  avgt   10      6.3x       ms/op
> compareToUL      24  avgt   10      1.6x       ms/op
> compareToUL      36  avgt   10      1.7x       ms/op
> compareToUL      72  avgt   10      2.2x       ms/op
> compareToUL     128  avgt   10      3.3x       ms/op
> compareToUL     256  avgt   10      4.4x       ms/op
> compareToUL     512  avgt   10      6.1x       ms/op
> compareToUU      24  avgt   10      1.0x       ms/op
> compareToUU      36  avgt   10      1.0x       ms/op
> compareToUU      72  avgt   10      1.4x       ms/op
> compareToUU     128  avgt   10      2.2x       ms/op
> compareToUU     256  avgt   10      2.6x       ms/op
> compareToUU     512  avgt   10      3.7x       ms/op
> 
> [1] https://github.com/openjdk/jdk/blob/master/test/hotspot/jtreg/compiler/intrinsics/string/TestStringCompareToDifferentLength.java

src/hotspot/cpu/aarch64/aarch64.ad line 16515:

> 16513: instruct string_compareUL(iRegP_R1 str1, iRegI_R2 cnt1, iRegP_R3 str2, iRegI_R4 cnt2,
> 16514:                         iRegI_R0 result, iRegP_R10 tmp1, iRegL_R11 tmp2,
> 16515:                         vRegD_V0 vtmp1, vRegD_V1 vtmp2, vRegD_V2 vtmp3, rFlagsReg cr)

I think vtmp3 (=V2) is still used by the non-SVE compare-long-strings stub? (see `generate_compare_long_string_different_encoding`)

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

PR: https://git.openjdk.java.net/jdk/pull/5129