[RyuJIT][LSRA] Take register encoding sizes into account

[CarolEidt]

Consider preferring smaller encoding sizes for high ref count, and larger encoding sizes for small ref count.
(This issue was migrated from the desktop bug tracking system.)
category:cq
theme:register-allocator
skill-level:expert
cost:medium

[mikedn]

A related example I happened to stumble upon:

class Num { public int x, y, z, w; }
static Dictionary<int, Num> Test(int x, int y)
{
    return new Dictionary<int, Num> { { 42, new Num { x = x, y = y, z = 4, w = 2 } } };
}

generates:

E8C78B815F           call     CORINFO_HELP_NEWSFAST
4C8BC0               mov      r8, rax
41897008             mov      dword ptr [r8+8], esi
4189780C             mov      dword ptr [r8+12], edi
41C7401004000000     mov      dword ptr [r8+16], 4
41C7401402000000     mov      dword ptr [r8+20], 2
488BCB               mov      rcx, rbx
BA2A000000           mov      edx, 42
41B902000000         mov      r9d, 2
E84969A35E           call     System.Collections.Generic.Dictionary`2[Int32,__Canon][System.Int32,System.__Canon]:TryInsert(int,ref,ubyte):bool:this

Here LSRA was too eager to move to r8 (presumably selected due to ABI requirements). If the move to r8 was delayed until the actual call the code would have been 4 bytes smaller.

[kunalspathak]

I do not think we will have time to do this in .NET 6.

[kunalspathak]

Fixed by #88151