ci: enable SIMD instructions for x86_64 targets

[Itsusinn]

🤔 This is a ...

• Performance optimization

[Itsusinn]

https://en.wikipedia.org/wiki/SSE2
https://en.wikipedia.org/wiki/SSE3
https://en.wikipedia.org/wiki/Advanced_Vector_Extensions

[Itsusinn]

https://github.com/Watfaq/clash-rs/releases/tag/v0.1.17

❯ elfx86exts clash-x86_64-unknown-linux-gnu
File format and CPU architecture: Elf, X86_64
MODE64 (call)
CMOV (cmova)
SSE1 (xorps)
SSE2 (movdqa)
BMI (tzcnt)
AES (aeskeygenassist)
AVX (vmovups)
NOVLX (vmovups)
AVX2 (vpbroadcastb)
SSSE3 (pshufb)
PCLMUL (pclmulqdq)
SSE41 (pblendw)
SSE3 (lddqu)
SHA (sha1rnds4)
BMI2 (shlx)
AVX512 (kmovw)
VLX (vpcmpltud)
ADX (adox)
3DNow (femms)
NOT64BITMODE (xchg)
Instruction set extensions used: 3DNow, ADX, AES, AVX, AVX2, AVX512, BMI, BMI2, CMOV, MODE64, NOT64BITMODE, NOVLX, PCLMUL, SHA, SSE1, SSE2, SSE3, SSE41, SSSE3, VLX
CPU Generation: Unknown

Seems SIMD is enabled by default on x86_64 targets.
But I wonder what will happen if old cpus use those binaries.

[Itsusinn]

It might be introduced by ring

[ibigbug]

It looks like we shouldn't target a specific CPU which would cause degradation on other platforms? Unless we build binary for each cpu ?

https://github.com/rust-lang/portable-simd/blob/master/beginners-guide.md

[Itsusinn]

It looks like we shouldn't target a specific CPU which would cause degradation on other platforms? Unless we build binary for each cpu ?

I use target CPU because there are about 5～8 target features need to be written. Using target cpu is shorter.

And I dont think it downgarde perf on other cpu. The sandy bridge is quite a modern CPU. It supports sse1 to sse4 , aes and sha instructions which includes default instrctions we are using now.

[ibigbug]

Okay makes sense

[VendettaReborn]

It sounds like a easy-to-obtain improvement, but I'd like to share Rob pike's programming principles
from my view, the real problem is, we even don't know where the bottleneck is, and we don't have any measures except the performance boost in theory.
Btw, I'm also thinking about ways to compare clash-rs's performance with clash.meta and sing-box. It would be fun. Maybe we can move to discussions.

[Itsusinn]

from my view, the real problem is, we even don't know where the bottleneck is, and we don't have any measures except the performance boost in theory.

We do need some performance characterization methods.But I think the performance of clash-rs is limited to user actual network condition.(CN2, BGP etc.) The perf test on local loopback could never be reach in realworld.

[VendettaReborn]

that's true, the network condition between test machines must be stable enough, in another words, they'd better exist under the same LAN

Need figure out a perf baseline