[JIT] Unnecessary stack spills when copying large struct from array/ref to array.

[omariom]

.NET 10

static void Copy(WideStruct[] target, WideStruct[] source)
{
    target[0] = source[1];
}

static void Copy(WideStruct[] target, ref WideStruct source)
{
    target[0] = source;
}

[StructLayout(LayoutKind.Sequential, Size = 64)]
struct WideStruct
{
}

Disasm:

; Assembly listing for method TestJITConsoleApp.Program:Copy(TestJITConsoleApp.WideStruct[],TestJITConsoleApp.WideStruct[]) (FullOpts)
; Emitting BLENDED_CODE for generic X64 + VEX on Windows
; FullOpts code
; optimized code
; rsp based frame
; partially interruptible
; No PGO data

G_M000_IG01:                ;; offset=0x0000
       sub      rsp, 104
 
G_M000_IG02:                ;; offset=0x0004
       cmp      dword ptr [rdx+0x08], 1
       jbe      SHORT G_M000_IG04
       vmovdqu  ymm0, ymmword ptr [rdx+0x50]
       vmovdqu  ymmword ptr [rsp+0x28], ymm0  ; <---- spill
       vmovdqu  ymm0, ymmword ptr [rdx+0x70]
       vmovdqu  ymmword ptr [rsp+0x48], ymm0
       cmp      dword ptr [rcx+0x08], 0
       jbe      SHORT G_M000_IG04
       vmovdqu  ymm0, ymmword ptr [rsp+0x28]
       vmovdqu  ymmword ptr [rcx+0x10], ymm0
       vmovdqu  ymm0, ymmword ptr [rsp+0x48]
       vmovdqu  ymmword ptr [rcx+0x30], ymm0
 
G_M000_IG03:                ;; offset=0x003C
       vzeroupper 
       add      rsp, 104
       ret      
 
G_M000_IG04:                ;; offset=0x0044
       call     CORINFO_HELP_RNGCHKFAIL
       int3     
 
; Total bytes of code 74


; Assembly listing for method TestJITConsoleApp.Program:Copy(TestJITConsoleApp.WideStruct[],byref) (FullOpts)
; Emitting BLENDED_CODE for generic X64 + VEX on Windows
; FullOpts code
; optimized code
; rsp based frame
; partially interruptible
; No PGO data

G_M000_IG01:                ;; offset=0x0000
       sub      rsp, 104
 
G_M000_IG02:                ;; offset=0x0004
       vmovdqu  ymm0, ymmword ptr [rdx]
       vmovdqu  ymmword ptr [rsp+0x28], ymm0
       vmovdqu  ymm0, ymmword ptr [rdx+0x20]
       vmovdqu  ymmword ptr [rsp+0x48], ymm0
       cmp      dword ptr [rcx+0x08], 0
       jbe      SHORT G_M000_IG04
       vmovdqu  ymm0, ymmword ptr [rsp+0x28]
       vmovdqu  ymmword ptr [rcx+0x10], ymm0
       vmovdqu  ymm0, ymmword ptr [rsp+0x48]
       vmovdqu  ymmword ptr [rcx+0x30], ymm0
 
G_M000_IG03:                ;; offset=0x0035
       vzeroupper 
       add      rsp, 104
       ret      
 
G_M000_IG04:                ;; offset=0x003D
       call     CORINFO_HELP_RNGCHKFAIL
       int3     
 
; Total bytes of code 67

[dotnet-policy-service]

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

[EgorBo]

At least, the first one should be handled via forward-sub (or early-prop?), today it bails out with

    [000006]:  [000007] is last use of [000006] (V03)  cannot reorder different thrown exceptions

Probably can extract the bound check/nullcheck side-effect and substitute the rest.

cc @dotnet/jit-contrib

[En3Tho]

@EgorBo It can also just use ymm1 for additional storage while preserving similar memory read order, isn't it?