Consuming static readonly Vector128<T> fields results in non-optimal codegen #13514
Description
Sample code:
private static readonly Vector128<sbyte> Vector0x0F = Vector128.Create((sbyte)0x0F);
public static int GetMask(Vector128<sbyte> data)
{
return Sse2.MoveMask(Sse2.Add(data, Vector0x0F));
}Codegen:
00007ffc`19901370 c5f877 vzeroupper
00007ffc`19901373 c5f91001 vmovupd xmm0,xmmword ptr [rcx]
00007ffc`19901377 48b8482ca75ea8010000 mov rax,1A85EA72C48h
00007ffc`19901381 488b00 mov rax,qword ptr [rax] ; this dereference could be avoided
00007ffc`19901384 c5f9fc4008 vpaddb xmm0,xmm0,xmmword ptr [rax+8]
00007ffc`19901389 c5f9d7c0 vpmovmskb eax,xmm0
00007ffc`1990138d c3 retSince the static ctor has already been run, the JIT could be optimized to burn into the codegen the exact memory location of where the Vector0x0F field lives. This would allow eliding the dereference highlighted in the above codegen sample.
/cc @tannergooding
category:cq
theme:basic-cq
skill-level:expert
cost:medium