Inefficient codegen for casts between same size types.

[tannergooding]

As per dotnet/coreclr#20788 (comment), using BitConverter.SingleToInt32Bits, BitConverter.Int32BitsToSingle, BitConverter.DoubleToInt64Bits, and BitConverter.Int64BitsToDouble generates some "inefficient" code.

Currently BitConverter.SingleToInt32Bits is generating:

vmovss   dword ptr [rsp+14H], xmm0
mov      eax, dword ptr [rsp+14H]

When it could generate:

vmovd eax, xmm0

Currently BitConverter.Int32BitsToSingle is generating:

mov      dword ptr [rsp+0CH], eax
vmovss   xmm0, dword ptr [rsp+0CH]

When it could generate:

vmovd xmm0, eax

The same logic applies to double <-> long, but using the rax register and vmovq.

• For x86, it can use the movq xmm, [m64] or movq [m64], xmm encoding

category:cq
theme:casts
skill-level:intermediate
cost:medium

[GrabYourPitchforks]

Related PR: #6864, which makes use of this API

[tannergooding]

#33476 introduced a fix which works around the issue, but we are leaving this issue open to track the JIT getting a more general codegen improvement that isn't specific to BitConverter.