Change Details

This is a work in progress, currently lacking a method to avoid crossing into unmapped pages. This changeset includes a port of the SIMD implementation of memcmp for amd64 to Aarch64. It also solves an issue with the existing implementation for Aarch64 where the return value is not in accordance with the man page and only returns -1,0 or 1 instead of the byte difference. Performance is better than the existing memcmp implementation borrowed from the Arm Optimized Routines except for long strings. ``` os: FreeBSD arch: arm64 cpu: ARM Neoverse-V1 r1p1 │ memcmpScalar │ memcmpARM │ memcmpSIMD │ │ sec/op │ sec/op vs base │ sec/op vs base │ MemcmpShort 139.41µ ± 2% 63.96µ ± 1% -54.12% (p=0.000 n=20) 25.14µ ± 1% -81.97% (p=0.000 n=20) MemcmpMid 93.38µ ± 3% 12.09µ ± 1% -87.05% (p=0.000 n=20) 10.91µ ± 1% -88.31% (p=0.000 n=20) MemcmpLong 86.722µ ± 7% 4.648µ ± 1% -94.64% (p=0.000 n=20) 4.931µ ± 0% -94.31% (p=0.000 n=20) geomean 104.1µ 15.32µ -85.29% 11.06µ -89.38% │ memcmpScalar │ memcmpARM │ memcmpSIMD │ │ B/s │ B/s vs base │ B/s vs base │ MemcmpShort 855.1Mi ± 2% 1864.0Mi ± 1% +117.99% (p=0.000 n=20) 4742.2Mi ± 1% +454.58% (p=0.000 n=20) MemcmpMid 1.247Gi ± 3% 9.629Gi ± 1% +671.89% (p=0.000 n=20) 10.668Gi ± 1% +755.20% (p=0.000 n=20) MemcmpLong 1.342Gi ± 6% 25.048Gi ± 1% +1765.92% (p=0.000 n=20) 23.608Gi ± 0% +1658.68% (p=0.000 n=20) geomean 1.118Gi 7.600Gi +579.66% 10.53Gi +841.33% os: FreeBSD arch: arm64 cpu: ARM Cortex-A76 r4p1 │ memcmpScalar │ memcmpARM │ memcmpSIMD │ │ sec/op │ sec/op vs base │ sec/op vs base │ MemcmpShort 183.12µ ± 0% 101.29µ ± 0% -44.68% (p=0.000 n=20) 39.96µ ± 0% -78.18% (p=0.000 n=20) MemcmpMid 129.12µ ± 0% 24.55µ ± 1% -80.99% (p=0.000 n=20) 21.48µ ± 0% -83.37% (p=0.000 n=20) MemcmpLong 111.374µ ± 0% 6.288µ ± 0% -94.35% (p=0.000 n=20) 7.593µ ± 0% -93.18% (p=0.000 n=20) geomean 138.1µ 25.01µ -81.89% 18.68µ -86.47% │ memcmpScalar │ memcmpARM │ memcmpSIMD │ │ B/s │ B/s vs base │ B/s vs base │ MemcmpShort 651.0Mi ± 0% 1176.9Mi ± 0% +80.78% (p=0.000 n=20) 2983.1Mi ± 0% +358.23% (p=0.000 n=20) MemcmpMid 923.3Mi ± 0% 4856.0Mi ± 1% +425.95% (p=0.000 n=20) 5551.0Mi ± 0% +501.23% (p=0.000 n=20) MemcmpLong 1.045Gi ± 0% 18.514Gi ± 0% +1671.24% (p=0.000 n=20) 15.332Gi ± 0% +1366.84% (p=0.000 n=20) geomean 863.3Mi 4.656Gi +452.24% 6.233Gi +639.33% ```

This changeset includes a port of the SIMD implementation of memcmp for amd64 to Aarch64. It also solves an issue with the existing implementation for Aarch64 where the return value is not in accordance with the man page and only returns -1,0 or 1 instead of the byte difference. Performance is better than the existing memcmp implementation borrowed from the Arm Optimized Routines except for long strings. ``` os: FreeBSD arch: arm64 cpu: ARM Neoverse-V1 r1p1 │ memcmpARM │ memcmpSIMD │ │ sec/op │ sec/op vs base │ MemcmpShort 63.96µ ± 1% 32.41µ ± 0% -49.33% (p=0.000 n=20) MemcmpMid 12.09µ ± 1% 12.33µ ± 1% +1.98% (p=0.000 n=20) MemcmpLong 4.648µ ± 1% 4.942µ ± 1% +6.32% (p=0.000 n=20) geomean 15.32µ 12.55µ -18.10% │ memcmpARM │ memcmpSIMD │ │ B/s │ B/s vs base │ MemcmpShort 1.820Gi ± 1% 3.592Gi ± 0% +97.35% (p=0.000 n=20) MemcmpMid 9.629Gi ± 1% 9.442Gi ± 1% -1.94% (p=0.000 n=20) MemcmpLong 25.05Gi ± 1% 23.55Gi ± 1% -5.96% (p=0.000 n=20) geomean 7.600Gi 9.279Gi +22.09% ```

This is a work in progress, currently lacking a method to avoid crossing into unmapped pages. This changeset includes a port of the SIMD implementation of memcmp for amd64 to Aarch64. It also solves an issue with the existing implementation for Aarch64 where the return value is not in accordance with the man page and only returns -1,0 or 1 instead of the byte difference. Performance is better than the existing memcmp implementation borrowed from the Arm Optimized Routines except for long strings. ``` os: FreeBSD arch: arm64 cpu: ARM Neoverse-V1 r1p1 │ memcmpScalar │ │ memcmpARM │ memcmpSIMD │ │ sec/op │ sec/op vs base │ sec/op vs base │ MemcmpShort 139.41µ ± 2% 63.96µ ± 1% -54.12% (p=0.000 n=20) 25.141% 32.41µ ± 10% -81.9749.33% (p=0.000 n=20) MemcmpMid 93.38µ ± 3% 12.09µ ± 1% -87.05% (p=0.000 n=20)% 10.912.33µ ± 1% -88.31 +1.98% (p=0.000 n=20) MemcmpLong 86.722µ ± 7% 4.648µ ± 1% -94.64% (p=0.000 n=20)% 4.93142µ ± 0% -94.311% +6.32% (p=0.000 n=20) geomean 104.1µ 15.32µ -85.29% 11.062.55µ -89.38%18.10% │ memcmpScalar │ │ memcmpARM │ memcmpSIMD │ │ B/s │ B/s vs base │ B/s vs base │ MemcmpShort 855.1Mi ± 2% 1864.0M1.820Gi ± 1% +117.99% (p=0.000 n=20) 4742.2M3.592Gi ± 1% +454.580% +97.35% (p=0.000 n=20) MemcmpMid 1.247Gi ± 3% 9.629Gi ± 1% +671.89% (p=0.000 n=20) 10.6689.442Gi ± 1% +755.20-1.94% (p=0.000 n=20) MemcmpLong 1.342Gi ± 6% 25.0485Gi ± 1% +1765.92% (p=0.000 n=20)1% 23.60855Gi ± 0% +1658.681% -5.96% (p=0.000 n=20) geomean 1.118Gi 7.600Gi +579.66% 10.53Gi 9.279Gi +841.33%22.09% os: FreeBSD arch: arm64 cpu: ARM Cortex-A76 r4p1 │ memcmpScalar │ memcmpARM │ memcmpSIMD │ │ sec/op │ sec/op vs base │ sec/op vs base │ MemcmpShort 183.12µ ± 0% 101.29µ ± 0% -44.68% (p=0.000 n=20) 39.96µ ± 0% -78.18% (p=0.000 n=20) MemcmpMid 129.12µ ± 0% 24.55µ ± 1% -80.99% (p=0.000 n=20) 21.48µ ± 0% -83.37% (p=0.000 n=20) MemcmpLong 111.374µ ± 0% 6.288µ ± 0% -94.35% (p=0.000 n=20) 7.593µ ± 0% -93.18% (p=0.000 n=20) geomean 138.1µ 25.01µ -81.89% 18.68µ -86.47% │ memcmpScalar │ memcmpARM │ memcmpSIMD │ │ B/s │ B/s vs base │ B/s vs base │ MemcmpShort 651.0Mi ± 0% 1176.9Mi ± 0% +80.78% (p=0.000 n=20) 2983.1Mi ± 0% +358.23% (p=0.000 n=20) MemcmpMid 923.3Mi ± 0% 4856.0Mi ± 1% +425.95% (p=0.000 n=20) 5551.0Mi ± 0% +501.23% (p=0.000 n=20) MemcmpLong 1.045Gi ± 0% 18.514Gi ± 0% +1671.24% (p=0.000 n=20) 15.332Gi ± 0% +1366.84% (p=0.000 n=20) geomean 863.3Mi 4.656Gi +452.24% 6.233Gi +639.33% ```