While modernizing some code in HexConverter to C# 11 ""u8 syntax:
|
Vector128<byte> asciiTable = (casing == Casing.Upper) ? |
|
Vector128.Create((byte)'0', (byte)'1', (byte)'2', (byte)'3', |
|
(byte)'4', (byte)'5', (byte)'6', (byte)'7', |
|
(byte)'8', (byte)'9', (byte)'A', (byte)'B', |
|
(byte)'C', (byte)'D', (byte)'E', (byte)'F') : |
|
Vector128.Create((byte)'0', (byte)'1', (byte)'2', (byte)'3', |
|
(byte)'4', (byte)'5', (byte)'6', (byte)'7', |
|
(byte)'8', (byte)'9', (byte)'a', (byte)'b', |
|
(byte)'c', (byte)'d', (byte)'e', (byte)'f'); |
to become:
Vector128<byte> asciiTable = Vector128.Create((casing == Casing.Upper) ?
"0123456789ABCDEF"u8 : "0123456789abcdef"u8);I noticed a significant regression in codegen. The culprit appears to be Vector128.Create's ReadOnlySpan<T> overload.
Repro
# bash on linux-x64
$ dotnet7 --version
7.0.100-preview.7.22358.13
$ dotnet7 new classlib -n VectorCreation
$ cat > VectorCreation/Class1.cs << EOF
using System.Runtime.Intrinsics;
public class D
{
public Vector128<byte> CreateWithROSpan() => Vector128.Create("0123456789ABCDEF"u8);
public Vector128<byte> CreateWithByteArray() => Vector128.Create(
(byte)'0', (byte)'1', (byte)'2', (byte)'3',
(byte)'4', (byte)'5', (byte)'6', (byte)'7',
(byte)'8', (byte)'9', (byte)'A', (byte)'B',
(byte)'C', (byte)'D', (byte)'E', (byte)'F');
}
EOF
$ dotnet7 publish VectorCreation -c Release -o dist --use-current-runtime -p:PublishAot=true -p:LangVersion=preview
$ gdb dist/VectorCreation.so -batch \
-ex 'set disassembly-flavor intel' \
-ex 'disassemble VectorCreation_D__CreateWithByteArray' \
-ex 'disassemble VectorCreation_D__CreateWithROSpan' \
-ex 'disassemble S_P_CoreLib_System_Runtime_Intrinsics_Vector128__Create_14<UInt8>'Codegen
Dump of assembler code for function VectorCreation_D__CreateWithByteArray:
0x0000000000211e10 <+0>: movups xmm0,XMMWORD PTR [rip+0x111689] # 0x3234a0
0x0000000000211e17 <+7>: movups XMMWORD PTR [rsi],xmm0
0x0000000000211e1a <+10>: mov rax,rsi
0x0000000000211e1d <+13>: ret
End of assembler dump.
vs.
Dump of assembler code for function VectorCreation_D__CreateWithROSpan:
0x0000000000211df0 <+0>: mov rdi,rsi
0x0000000000211df3 <+3>: lea rsi,[rip+0x10f33e] # 0x321138
0x0000000000211dfa <+10>: mov edx,0x10
0x0000000000211dff <+15>: jmp 0x273a50 <S_P_CoreLib_System_Runtime_Intrinsics_Vector128__Create_14<UInt8>>
End of assembler dump.
Dump of assembler code for function S_P_CoreLib_System_Runtime_Intrinsics_Vector128__Create_14<UInt8>:
0x0000000000273a50 <+0>: push rbp
0x0000000000273a51 <+1>: mov rbp,rsp
0x0000000000273a54 <+4>: cmp edx,0x10
0x0000000000273a57 <+7>: jl 0x273a64 <S_P_CoreLib_System_Runtime_Intrinsics_Vector128__Create_14<UInt8>+20>
0x0000000000273a59 <+9>: movups xmm0,XMMWORD PTR [rsi]
0x0000000000273a5c <+12>: movups XMMWORD PTR [rdi],xmm0
0x0000000000273a5f <+15>: mov rax,rdi
0x0000000000273a62 <+18>: pop rbp
0x0000000000273a63 <+19>: ret
0x0000000000273a64 <+20>: mov edi,0x6
0x0000000000273a69 <+25>: call 0x17ff40 <S_P_CoreLib_System_ThrowHelper__ThrowArgumentOutOfRangeException_0>
0x0000000000273a6e <+30>: int3
End of assembler dump.
Side note: with -p:IlcInstructionSet=avx2, we get additional vzerouppers (in CreateWithByteArray as well), which seems to be related to #11496.
While modernizing some code in
HexConverterto C# 11""u8syntax:runtime/src/libraries/Common/src/System/HexConverter.cs
Lines 99 to 107 in d344e72
I noticed a significant regression in codegen. The culprit appears to be Vector128.Create's
ReadOnlySpan<T>overload.Repro
Codegen
vs.
Side note: with
-p:IlcInstructionSet=avx2, we get additionalvzerouppers (inCreateWithByteArrayas well), which seems to be related to #11496.