Complete avx512vbmi2

[minybot]

expandloadu_epi16,epi8
mask_compressstoreu_epi16,epi8

[rust-highfive]

r? @Amanieu

(rust-highfive has picked a reviewer for you, use r? to override)

[jhorstmann]

Nice, I also had those on my todo list but did not get around to implementing them yet. Note that there are also llvm intrinsics for expand load (examples 1, 2) which might lead to a bit simpler code and automatically support 32-bit pointers.

[minybot]

Nice, I also had those on my todo list but did not get around to implementing them yet. Note that there are also llvm intrinsics for expand load (examples 1, 2) which might lead to a bit simpler code and automatically support 32-bit pointers.

Thanks. I will check those links.
After this, I will try to finish avx512bw left ones if I have time during the weekend.
Thanks for your help for avx512f.

[minybot]

Nice, I also had those on my todo list but did not get around to implementing them yet. Note that there are also llvm intrinsics for expand load (examples 1, 2) which might lead to a bit simpler code and automatically support 32-bit pointers.

I check LLVM code. We need to use asm! for expandload because there is "i1".
llvm.masked.expandload.v32i16(i16* %1, <32 x i1> %2, <32 x i16> %0)

[jhorstmann]

There are two variants, the architecture independent ones with i1 masks and the avx512 specific ones, like llvm.x86.avx512.mask.expand.load.w.512(i8* %addr, <32 x i16> %data, i32 %mask). I don't know whether the avx512 specific ones are deprecated, that would probably make asm the better option.

[minybot]

There are two variants, the architecture independent ones with i1 masks and the avx512 specific ones, like llvm.x86.avx512.mask.expand.load.w.512(i8* %addr, <32 x i16> %data, i32 %mask). I don't know whether the avx512 specific ones are deprecated, that would probably make asm the better option.

In this situation, what should we do? If there is a LLVM one, I prefer the LLVM one.