Order of initialization in async kickoff method with params blocks Jit zero init eliding #45295
Description
Currently Rosyln zeros the the AsyncMethodBuilder field with .locals init and then assigns the result of .Create to that field.
When .Create is inlined it is also zeros; so the assignment is a second set of zeroing.
The Jit has an optimization where it will remove the second set of zeroing dotnet/runtime#36918 ("Optimization to remove redundant zero initializations"); however this only kicks in if there aren't any preceding non-zero assignments.
e.g.
using System.IO.Pipelines;
public async ValueTask<ReadResult> OuterMethod(object o)
{
return await InnerMethod(o);
static async ValueTask<ReadResult> InnerMethod(object o)
{
return default;
}
}G_M51821_IG01:
push rsi
sub rsp, 144
vzeroupper ; zero
xor rax, rax ; |
mov qword ptr [rsp+28H], rax ; |
vxorps xmm4, xmm4 ; |
mov rax, -96 ; |
vmovdqa xmmword ptr [rsp+rax+90H], xmm4 ; |
vmovdqa xmmword ptr [rsp+rax+A0H], xmm4 ; |
vmovdqa xmmword ptr [rsp+rax+B0H], xmm4 ; |
add rax, 48 ; |
jne SHORT -5 instr ; V
mov rsi, rdx
;; bbWeight=1 PerfScore 8.58
G_M51821_IG02:
mov gword ptr [rsp+28H], r8 ; save param ** blocks redundant zero elimination **
xor ecx, ecx ; re-zero
vxorps xmm0, xmm0 ; |
vmovdqu xmmword ptr [rsp+38H], xmm0 ; |
vmovdqu xmmword ptr [rsp+48H], xmm0 ; |
mov qword ptr [rsp+58H], rcx ; V
mov dword ptr [rsp+30H], -1
lea rcx, [rsp+28H]
call System.Runtime.CompilerServices.AsyncMethodBuilderCore:Start(byref)
lea rcx, bword ptr [rsp+38H]
mov rdx, rsi
call System.Runtime.CompilerServices.AsyncValueTaskMethodBuilder`1[ReadResult][System.IO.Pipelines.ReadResult]:get_Task():System.Threading.Tasks.ValueTask`1[ReadResult]:this
mov rax, rsiIf the call to .Create() was moved to the first step; prior to saving parameters to the fields so it should not block the redundant zero elimination optimization.
From dotnet/runtime#2325 (comment)
Regarding your second example where we first assign to a gc field and then initialize the struct field: currently my optimization doesn't handle that. It's possible to extend it to handle that case but that would require keeping track of field initializations and may be a little tricky.
However, now the answer to your question
Or would changing it to move all the zeroing above that first assignment to non-zero (which would be a much more complex refactoring in Roslyn) help?
is yes.
If the struct field zero initialization was done before the first assignment, it would be eliminated by the new optimization.