RyuJIT: Unnecessary Copy of System V x86-64 Multi-Register Structure Returning

[fiigii]

According to VTune characterization of TechEmpower with CoreCLR 2.0-beta, the top hot function of System::IO::Pipelines is Pipe.get_buffer, which is called by the top hot function WriteFast in Kestrel.

internal Buffer<byte> Buffer => _writingHead?.Buffer.Slice(_writingHead.End, _writingHead.WritableBytes) ?? Buffer<byte>.Empty;

There are codegen differences between Windows and Ubuntu as shown below:

Windows	Ubuntu
Block 1:	Block 1:
push rdi	push r14
push rsi	push r13
push rbx	push rbx
sub rsp, 0x20	sub rsp, 0x40
mov rsi, rdx	vzeroupper
mov rdi, qword ptr [rcx+0x38]	mov r13, rdi
mov rbx, rdi	lea rdi, ptr [rsp]
test rbx, rbx	mov ecx, 0x10
jnz 0x7ffa9a429019 <Block 4>	xor eax, eax
Block 2:	rep stosd dword ptr [rdi]
mov rcx, rsi	mov rdi, r13
call 0x7ffa9a421410	mov rbx, qword ptr [rdi+0x38]
Block 3:	mov r14, rbx
mov rax, rsi	test r14, r14
add rsp, 0x20	jnz 0x7f52d191a08c <Block 4>
pop rbx	Block 2:
pop rsi	call 0x7f52d1910b90
pop rdi	Block 3:
ret	mov qword ptr [rsp+0x10], rax
	mov qword ptr [rsp+0x18], rdx
	vmovdqu xmm0, xmmword ptr [rsp+0x10]
	vmovdqu xmmword ptr [rsp+0x20], xmm0
	mov rax, qword ptr [rsp+0x20]
	mov rdx, qword ptr [rsp+0x28]
	add rsp, 0x40
	pop rbx
	pop r13
	pop r14
	ret
Block 4:	Block 4:
lea rdx, ptr [rbx+0x28]	lea rax, ptr [r14+0x28]
mov rcx, rdx	mov rdi, rax
mov rax, qword ptr [rcx]	mov rsi, qword ptr [rdi]
mov ecx, dword ptr [rcx+0x8]	mov edi, dword ptr [rdi+0x8]
mov r8d, dword ptr [rdi+0x1c]	mov edx, dword ptr [rbx+0x1c]
mov r9d, r8d	mov ecx, edx
mov r10d, dword ptr [rdi]	mov r8d, dword ptr [rbx]
mov edx, dword ptr [rdx+0xc]	mov eax, dword ptr [rax+0xc]
mov edi, edx	sub eax, edx
sub edi, r8d	lea rdx, ptr [rsp]
lea ebx, ptr [rcx+r9*1]	vxorpd xmm0, xmm0, xmm0
lea rcx, ptr [rsi]	vmovdqu xmmword ptr [rdx], xmm0
mov rdx, rax	add edi, ecx
call CORINFO_HELP_CHECKED_ASSIGN_REF	mov qword ptr [rsp], rsi
Block 5:	mov dword ptr [rsp+0x8], edi
mov dword ptr [rsi+0x8], ebx	mov dword ptr [rsp+0xc], eax
mov dword ptr [rsi+0xc], edi	vmovdqu xmm0, xmmword ptr [rsp]
mov rax, rsi	vmovdqu xmmword ptr [rsp+0x30], xmm0
add rsp, 0x20	mov rax, qword ptr [rsp+0x30]
pop rbx	mov rdx, qword ptr [rsp+0x38]
pop rsi	add rsp, 0x40
pop rdi	pop rbx
ret	pop r13
	pop r14
	ret

There are two codegen issues that make Linux version slower:

1. Redundantly copy the result of calling Buffer<byte>.Empty (highlighted code in Block 3).
2. Redundantly copy the result of calling (inlined) Slice (highlighted code in Block 4).

I will look into fixing these codegen issues.

[RussKeldorph]

@dotnet/jit-contrib

There are a few existing issues that may be related

[fiigii]

@RussKeldorph Most of these existing issues are related to structure argument passing, and this issue is only about returned structures.

[RussKeldorph]

@fiigii Understood. I didn't mean to suggest your issue was a duplicate; just that in theory there might be other issues affected by a fix.

[fiigii]

@RussKeldorph I see, thanks for explaining.

[omariom]

It may be related to First class structs