[arm64] Volatile.Read/Write is 2x faster than "volatile" loads/stores #60232
Description
The following code (same for "loads"):
private volatile int A;
private volatile int B;
public void SetValueInLoop_volatile()
{
for (int i = 0; i < 1000; i++)
{
A = i;
B = i;
}
}; Method VolatileKeyword:SetValueInLoop():this
G_M8342_IG01:
stp fp, lr, [sp,#-16]!
mov fp, sp
G_M8342_IG02:
mov w1, wzr
G_M8342_IG03:
dmb ish ;; <--- full memory barrier
str w1, [x0,#8]
dmb ish ;; <--- full memory barrier
str w1, [x0,#12]
add w1, w1, #1
cmp w1, #1000
blt G_M8342_IG03
G_M8342_IG04:
ldp fp, lr, [sp],#16
ret lr
; Total bytes of code: 48is twice slower than:
private int A;
private int B;
public void SetValueInLoop_VolatileWrite()
{
for (int i = 0; i < 1000; i++)
{
Volatile.Write(ref A, i);
Volatile.Write(ref B, i);
}
}; Method VolatileClass:SetValueInLoop():this
G_M39137_IG01:
stp fp, lr, [sp,#-16]!
mov fp, sp
G_M39137_IG02:
mov w1, wzr
add x2, x0, #8
add x0, x0, #12
G_M39137_IG03:
mov x3, x2
stlr w1, [x3]
mov x3, x0
stlr w1, [x3]
add w1, w1, #1
cmp w1, #1000
blt G_M39137_IG03
G_M39137_IG04:
ldp fp, lr, [sp],#16
ret lr
; Total bytes of code: 56Benchmark (Apple M1, arm64):
| Method | Mean | Error | StdDev |
|----------------------------- |-----------:|--------:|--------:|
| SetValueInLoop_VolatileWrite | 623.7 ns | 0.12 ns | 0.11 ns |
| SetValueInLoop_volatile | 1,257.0 ns | 0.26 ns | 0.23 ns |
Unrelated, but just for fun - same benchmark, same M1 machine but under Rosetta (x64 emulation):
| SetValueInLoop_VolatileWrite | 320.9 ns | 0.11 ns | 0.10 ns |
| SetValueInLoop_volatile | 325.5 ns | 1.31 ns | 1.22 ns |
cc @dotnet/jit-contrib
category:cq
theme:volatile
Metadata
Metadata
Assignees
Labels
Type
Projects
Status