CINC/CSEL not emitted inside the loops instead of jumps over single instruction blocks #96380
Description
Description
It appears that for a straightforward compare -> increment, .NET emits a branch instead of cinc when a method containing such code is inlined inside a loop body.
Analysis
Given a method RuneLength:
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static int RuneLength(in byte value)
{
var lzcnt = (uint)BitOperations.LeadingZeroCount(~((uint)value << 24));
if (lzcnt is 0) lzcnt++;
return (int)lzcnt;
}This compiles to:
G_M000_IG01: ;; offset=0x0000
stp fp, lr, [sp, #-0x10]!
mov fp, sp
G_M000_IG02: ;; offset=0x0008
ldrb w0, [x0]
lsl w0, w0, #24
mvn w0, w0
clz w0, w0
mov w1, #1
cmp w0, #0
csel w0, w0, w1, ne ;; <-- Could have been just cinc with mov 1 elided - we're incrementing zero
G_M000_IG03: ;; offset=0x0024
ldp fp, lr, [sp], #0x10
ret lr
; Total bytes of code 44We can see that .NET emits mov + cmp + csel here. This isn't cinc but it's a good start.
However, if the method is inlined inside a loop body, the codegen is different.
Consider the below method:
static int Iterate(ref byte ptr, ref byte end)
{
var acc = 0;
while (Unsafe.IsAddressLessThan(ref ptr, ref end))
{
var length = RuneLength(in ptr);
acc += length;
ptr = ref Unsafe.Add(ref ptr, length);
}
return acc;
}Instead of the pattern above, the codegen changes to cbnz label and mov 1 which is more compact but is more expensive because it uses branch execution units which have lower throughput per cycle than csel which uses integer units instead (on modern ARM cores like Firestorm or Cortex-X1):
G_M000_IG01: ;; offset=0x0000
stp fp, lr, [sp, #-0x10]!
mov fp, sp
G_M000_IG02: ;; offset=0x0008
mov w2, wzr
cmp x0, x1
bhs G_M000_IG06
align [0 bytes for IG03]
align [0 bytes]
align [0 bytes]
align [0 bytes]
G_M000_IG03: ;; offset=0x0014
ldrb w3, [x0]
lsl w3, w3, #24
mvn w3, w3
clz w3, w3
cbnz w3, G_M000_IG05 ;; <-- Could have been cmp + cinc
G_M000_IG04: ;; offset=0x0028
mov w3, #1
G_M000_IG05: ;; offset=0x002C
add w2, w2, w3
sxtw x3, w3
add x0, x0, x3
cmp x0, x1
blo G_M000_IG03
G_M000_IG06: ;; offset=0x0040
mov w0, w2
G_M000_IG07: ;; offset=0x0044
ldp fp, lr, [sp], #0x10
ret lr
; Total bytes of code 76Configuration
.NET SDK:
Version: 8.0.100
Commit: 57efcf1350
Workload version: 8.0.100-manifests.71b9f198
Runtime Environment:
OS Name: Mac OS X
OS Version: 14.1
OS Platform: Darwin
RID: osx-arm64
Base Path: /usr/local/share/dotnet/sdk/8.0.100/
Regression?
No
Happy New Year!