Optimize constant AVX512/AVX2 vectors with broadcast

[EgorBo]

Vector512<byte> HexLUT() => 
    Vector512.Create("0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef"u8);

Produces:

; Method Tests:HexLUT():System.Runtime.Intrinsics.Vector512`1[ubyte]:this (FullOpts)
       vzeroupper 
       vmovups  zmm0, zmmword ptr [reloc @RWD00]
       vmovups  zmmword ptr [rdx], zmm0
       mov      rax, rdx
       vzeroupper 
       ret      
RWD00  	dq	
3736353433323130h, 6665646362613938h, 
3736353433323130h, 6665646362613938h, 
3736353433323130h, 6665646362613938h, 
3736353433323130h, 6665646362613938h
; Total bytes of code: 26

so it saves the whole thing to data section while it should be smart enough to only store a single lane and broadcast it. It is quite often with AVX2 and AVX512 we still work with 128bit lanes and duplicate constants per lane.

Expected codegen:

; Method Tests:HexLUT():System.Runtime.Intrinsics.Vector512`1[ubyte]:this (FullOpts)
       vzeroupper 
       broadcasti32x4 zmm0, xmmword ptr [reloc @RWD00]
       vmovups  zmmword ptr [rdx], zmm0
       mov      rax, rdx
       vzeroupper 
       ret      
RWD00  	dq	
3736353433323130h, 6665646362613938h

[EgorBo]

Additional/Alternative action is to expose crossplat api Vector512<T> Create<T>(Vector128<T>)

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Vector512<byte> HexLUT() => 
    Vector512.Create("0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef"u8);

Produces:

; Method Tests:HexLUT():System.Runtime.Intrinsics.Vector512`1[ubyte]:this (FullOpts)
       vzeroupper 
       vmovups  zmm0, zmmword ptr [reloc @RWD00]
       vmovups  zmmword ptr [rdx], zmm0
       mov      rax, rdx
       vzeroupper 
       ret      
RWD00  	dq	
3736353433323130h, 6665646362613938h, 
3736353433323130h, 6665646362613938h, 
3736353433323130h, 6665646362613938h, 
3736353433323130h, 6665646362613938h
; Total bytes of code: 26

so it saves the whole thing to data section while it should be smart enough to only store a singline lane and broadcast it. It is quite often with AVX2 and AVX512 we still work with 128bit lanes and duplicate constants per lane.

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[Ruihan-Yin]

Hi, @EgorBo, @DeepakRajendrakumaran and I will work on this issue, and hopefully we could make a public PR soon.

[Ruihan-Yin]

After #92017, JIT is generating the expected codegen described in this issue:
https://godbolt.org/z/YWW6K43na