Investigate adjusting herustics for unrolled block copies/initialization #82529
Description
Zero initialization becomes expensive with two bytes struct fields are involved:
Given:
unsafe struct S1
{
fixed byte a[10];
int b;
fixed byte c[23];
fixed byte d[24];
fixed byte e[25];
}
unsafe struct S2
{
fixed short a[10];
int b;
fixed short c[23];
fixed short d[24];
fixed short e[25];
}Unlike S1, initializing S2 does not inline the CORINFO_HELP_MEMSET call and fail to make use of vectors:
; S1 X1() { S1 s = default; return s; }
C:X1():S1:this:
sub rsp, 24
vzeroupper
mov rax, qword ptr [(reloc)]
mov qword ptr [rsp+10H], rax
xor eax, eax
vxorps ymm0, ymm0
vmovdqu ymmword ptr[rsi], ymm0
vmovdqu ymmword ptr[rsi+20H], ymm0
vmovdqu xmmword ptr [rsi+40H], xmm0
mov qword ptr [rsi+50H], rax
mov rax, rsi
lea rdi, [(reloc)]
mov rdi, qword ptr [rdi]
cmp qword ptr [rsp+10H], rdi
je SHORT G_M61359_IG03
call [CORINFO_HELP_FAIL_FAST]
G_M61359_IG03:
nop
add rsp, 24
ret
; S2 X2() { S2 s = default; return s; }
C:X2():S2:this:
push rbx
sub rsp, 16
mov rax, qword ptr [(reloc)]
mov qword ptr [rsp+08H], rax
mov rbx, rsi
xor esi, esi
mov rdi, rbx
mov edx, 168
call [CORINFO_HELP_MEMSET]
mov rax, rbx
lea rdi, [(reloc)]
mov rdi, qword ptr [rdi]
cmp qword ptr [rsp+08H], rdi
je SHORT G_M46095_IG03
call [CORINFO_HELP_FAIL_FAST]
G_M46095_IG03:
nop
add rsp, 16
pop rbx
ret The codegen can try to match that of C++ compiler:
X2(): # @X2()
mov rax, rdi
vxorps xmm0, xmm0, xmm0
vmovups ymmword ptr [rdi + 128], ymm0
vmovups ymmword ptr [rdi + 96], ymm0
vmovups ymmword ptr [rdi + 64], ymm0
vmovups ymmword ptr [rdi + 32], ymm0
vmovups ymmword ptr [rdi], ymm0
mov qword ptr [rdi + 160], 0
vzeroupper
ret