CLMUL instruction set

[newpavlov]

Fixes #318

Depends on rust-lang/rust#48126.

[alexcrichton]

Thanks! The verification program is reporting

• intel cpuid pclmulqdq not in pclmul for _mm_clmulepi64_si128

Perhaps the feature in Rust itself could be called pclmulqdq?

[newpavlov]

We have 3 options for feature name:

• clmul - shortest one, coherent with intrinsic name _mm_clmulepi64_si128, matches with instruction set name
• pclmul - used by LLVM
• pclmulqdq - longest one, matches with CPUID and instruction mnemonic.

I've used pclmul because it required only simple whitelisting, while other options will require more involved changes, which I am currently don't know how to implement.

[alexcrichton]

So far I think our intention is to provide a direct mapping from what vendors (Intel in this case) says. So far everything (except bmi/bmi1) has aligned with LLVM but I think for new ones we'll want to align with Intel rather than LLVM.

[newpavlov]

Can you please list necessary changes for adding pclmulqdq feature which will use pclmul LLVM feature under the hood? (I think it's better to discuss in the relevant Rust PR)

[alexcrichton]

Sure! I'll comment on the Rust PR

[alexcrichton]

@newpavlov I think these should be in nightly now, right? If so, want to rebase to kick CI?

[newpavlov]

I will change imm8: u8 to imm8: i32 to be coherent with other intrinsics.

Is it worth to open an issue for discussing potential change of i32 to u8 or i8 for immediate 8 bits arguments? I understand i32 was chosen to be coherent with headers, but I personally don't think we should follow them in this case.

[alexcrichton]

Nah this is the territory of the RFC, so if you strongly feel we should be diverging from the headers I'd recommend reading over the discussion on the RFC and leaving a comment if you still feel that way. For now though, yeah, let's stick to the headers here.

[newpavlov]

Hm, not sure how to fix this. In some cases _mm_clmulepi64_si128(a, b, 0x00) gets compiled to pclmullqlqdq %xmm1,%xmm0 and in others to pclmulqdq $0x0, %xmm1, %xmm0 . IIUC both mnemonics are equivalent.

[alexcrichton]

Ah no worries, wanna only run that test on Linux and leave a fixme?

[newpavlov]

Ah, indeed I could've run OS specific tests. I think last commit should fix it.

[alexcrichton]

Awesome, thanks!