Constructing a Vector4 from 4 constant floats is very inefficient

[tannergooding]

Code that is in the form of: var value = new Vector4(1, 2, 3, 4) currently generates very inefficient code.

Today, this produces something similar to:

C4E17A100518010000   vmovss   xmm0, dword ptr [reloc @RWD00]
C4E17A100D13010000   vmovss   xmm1, dword ptr [reloc @RWD04]
C4E17A10150E010000   vmovss   xmm2, dword ptr [reloc @RWD08]
C4E17A101D09010000   vmovss   xmm3, dword ptr [reloc @RWD12]
C4E15857E4           vxorps   xmm4, xmm4
C4E15A10E3           vmovss   xmm4, xmm4, xmm3
C4E15973FC04         vpslldq  xmm4, 4
C4E15A10E2           vmovss   xmm4, xmm4, xmm2
C4E15973FC04         vpslldq  xmm4, 4
C4E15A10E1           vmovss   xmm4, xmm4, xmm1
C4E15973FC04         vpslldq  xmm4, 4
C4E15A10E0           vmovss   xmm4, xmm4, xmm0
C4E17828C4           vmovaps  xmm0, xmm4

The JIT should recognize this type of code and instead optimize it to be a single read:

vmovaps xmm0 xmm ptr [address]    ; Use vmovups if address is not aligned

category:cq
theme:vector-codegen
skill-level:intermediate
cost:medium

[tannergooding]

CC. @CarolEidt, @eerhardt

[mikedn]

Eh, I guess I need to look into emitting those 16 bytes data sections a bit more, this issue keeps coming up one way or another :)

[mikedn]

Funny thing that this needs 16 bytes of data section anyway, to store the 4 FP constants. And then generates ~75 bytes of extra code...

[ufcpp]

related to dotnet/csharplang#688 ?

[mikedn]

Yes. Except this one is actually doable. The csharplang suggestion, well, not so much...

[tannergooding]

The csharplang suggestion, well, not so much...

Could you please comment, on the language issue, why you believe this is not doable.

At least from the initial IL tests, using Data declarations to emit structured constants worked as intended, on multiple platforms. So, outside of language design issues (such as what to do for cross-assembly constant declarations), it should be doable (and is actually doable in IL today).