Improve XmlDictionaryWriter UTF8 encoding performance

src/libraries/System.Private.DataContractSerialization/src/System.Private.DataContractSerialization.csproj

@@ -156,10 +156,12 @@
     <Reference Include="System.Collections.Specialized" />
     <Reference Include="System.Linq" />
     <Reference Include="System.Memory" />
+    <Reference Include="System.Numerics.Vectors" />
     <Reference Include="System.Reflection.Emit.ILGeneration" />
     <Reference Include="System.Reflection.Emit.Lightweight" />
     <Reference Include="System.Reflection.Primitives" />
     <Reference Include="System.Runtime" />
+    <Reference Include="System.Runtime.Intrinsics" />
     <Reference Include="System.Runtime.Serialization.Formatters" />
     <Reference Include="System.Runtime.Serialization.Primitives" />
     <Reference Include="System.Text.Encoding.Extensions" />

src/libraries/System.Private.DataContractSerialization/src/System/Xml/XmlStreamNodeWriter.cs

@@ -2,7 +2,10 @@
 // The .NET Foundation licenses this file to you under the MIT license.
 
 using System.IO;
+using System.Numerics;
 using System.Text;
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
 using System.Runtime.Serialization;
 using System.Threading.Tasks;
 
@@ -337,32 +340,28 @@ protected unsafe void UnsafeWriteUnicodeChars(char* chars, int charCount)
 
         protected unsafe int UnsafeGetUnicodeChars(char* chars, int charCount, byte[] buffer, int offset)
         {
-            char* charsMax = chars + charCount;
-            while (chars < charsMax)
+            if (BitConverter.IsLittleEndian)
             {
-                char value = *chars++;
-                buffer[offset++] = (byte)value;
-                value >>= 8;
-                buffer[offset++] = (byte)value;
+                new ReadOnlySpan<byte>((byte*)chars, sizeof(char) * charCount)
+                    .CopyTo(buffer.AsSpan(offset));
             }
+            else
+            {
+                char* charsMax = chars + charCount;
+                while (chars < charsMax)
+                {
+                    char value = *chars++;
+                    buffer[offset++] = (byte)value;
+                    buffer[offset++] = (byte)(value >> 8);
+                }
+            }
+
             return charCount * 2;
         }
 
         protected unsafe int UnsafeGetUTF8Length(char* chars, int charCount)
         {
-            char* charsMax = chars + charCount;
-            while (chars < charsMax)
-            {
-                if (*chars >= 0x80)
-                    break;
-
-                chars++;
-            }
-
-            if (chars == charsMax)
-                return charCount;
-
-            return (int)(chars - (charsMax - charCount)) + (_encoding ?? s_UTF8Encoding).GetByteCount(chars, (int)(charsMax - chars));
+            return (_encoding ?? s_UTF8Encoding).GetByteCount(chars, charCount);
         }
 
         protected unsafe int UnsafeGetUTF8Chars(char* chars, int charCount, byte[] buffer, int offset)
@@ -372,38 +371,51 @@ protected unsafe int UnsafeGetUTF8Chars(char* chars, int charCount, byte[] buffe
                 fixed (byte* _bytes = &buffer[offset])
                 {
                     byte* bytes = _bytes;
-                    byte* bytesMax = &bytes[buffer.Length - offset];
                     char* charsMax = &chars[charCount];
 
-                    while (true)
+                    if (Sse41.IsSupported && charCount >= Vector128<short>.Count)
+                    {
+                        Vector128<short> mask = Vector128.Create(unchecked((short)0xff80));
+                        char* simdLast = chars + charCount - Vector128<short>.Count;
+
+                        while (chars < simdLast)
+                        {
+                            Vector128<short> v = *(Vector128<short>*)chars;
+                            if (!Sse41.TestZ(v, mask))
+                                goto NonAscii;
+
+                            Sse2.StoreScalar((long*)bytes, Sse2.PackUnsignedSaturate(v, v).AsInt64());
+                            bytes += Vector128<short>.Count;
+                            chars += Vector128<short>.Count;
+                        }
+
+                        Vector128<short> v2 = Sse2.LoadVector128((short*)simdLast);
+                        if (!Sse41.TestZ(v2, mask))
+                            goto NonAscii;
+
+                        Sse2.StoreScalar((long*)(_bytes + charCount - sizeof(long)), Sse2.PackUnsignedSaturate(v2, v2).AsInt64());
+                        return charCount;
+                    }
+                    // Fast path for small strings, skip and use Encoding.GetBytes for larger strings since it is faster even for the all-Ascii case
+                    else if (charCount < 16)
                     {
                         while (chars < charsMax)
                         {
                             char t = *chars;
                             if (t >= 0x80)
-                                break;
+                                goto NonAscii;
 
                             *bytes = (byte)t;
                             bytes++;
                             chars++;
                         }
 
-                        if (chars >= charsMax)
-                            break;
-
-                        char* charsStart = chars;
-                        while (chars < charsMax && *chars >= 0x80)
-                        {
-                            chars++;
-                        }
-
-                        bytes += (_encoding ?? s_UTF8Encoding).GetBytes(charsStart, (int)(chars - charsStart), bytes, (int)(bytesMax - bytes));
-
-                        if (chars >= charsMax)
-                            break;
+                        return charCount;
                     }
 
-                    return (int)(bytes - _bytes);
+                NonAscii:
+                    byte* bytesMax = _bytes + buffer.Length - offset;
+                    return (int)(bytes - _bytes) + (_encoding ?? s_UTF8Encoding).GetBytes(chars, (int)(charsMax - chars), bytes, (int)(bytesMax - bytes));
                 }
             }
             return 0;

src/libraries/System.Runtime.Serialization.Xml/System.Runtime.Serialization.Xml.sln

@@ -1,4 +1,8 @@
-Microsoft Visual Studio Solution File, Format Version 12.00
+
+Microsoft Visual Studio Solution File, Format Version 12.00
+# Visual Studio Version 17
+VisualStudioVersion = 17.2.32616.157
+MinimumVisualStudioVersion = 10.0.40219.1
 Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "TestUtilities", "..\Common\tests\TestUtilities\TestUtilities.csproj", "{CBA80130-6773-4DF9-995C-DC6CBED89CB5}"
 EndProject
 Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "Microsoft.Win32.Primitives", "..\Microsoft.Win32.Primitives\ref\Microsoft.Win32.Primitives.csproj", "{E5DB95E1-94AA-405C-9FFE-09B1E2498EE2}"
@@ -45,6 +49,10 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "src", "src", "{DB29DBEF-FA4
 EndProject
 Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "gen", "gen", "{DE71D38E-4154-477C-9C27-3FA4ADB4098F}"
 EndProject
+Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "System.Numerics.Vectors", "..\System.Numerics.Vectors\ref\System.Numerics.Vectors.csproj", "{EFE0C13B-6902-4FC9-91DD-F180420B36C8}"
+EndProject
+Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "System.Runtime.Intrinsics", "..\System.Runtime.Intrinsics\ref\System.Runtime.Intrinsics.csproj", "{7F8A8D87-B49A-4C7B-8474-AC83F8CFD40B}"
+EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
 		Debug|Any CPU = Debug|Any CPU
@@ -127,30 +135,40 @@ Global
 		{DF2255F4-F671-4C15-9100-D8079992E19D}.Debug|Any CPU.Build.0 = Debug|Any CPU
 		{DF2255F4-F671-4C15-9100-D8079992E19D}.Release|Any CPU.ActiveCfg = Release|Any CPU
 		{DF2255F4-F671-4C15-9100-D8079992E19D}.Release|Any CPU.Build.0 = Release|Any CPU
+		{EFE0C13B-6902-4FC9-91DD-F180420B36C8}.Debug|Any CPU.ActiveCfg = Debug|Any CPU
+		{EFE0C13B-6902-4FC9-91DD-F180420B36C8}.Debug|Any CPU.Build.0 = Debug|Any CPU
+		{EFE0C13B-6902-4FC9-91DD-F180420B36C8}.Release|Any CPU.ActiveCfg = Release|Any CPU
+		{EFE0C13B-6902-4FC9-91DD-F180420B36C8}.Release|Any CPU.Build.0 = Release|Any CPU
+		{7F8A8D87-B49A-4C7B-8474-AC83F8CFD40B}.Debug|Any CPU.ActiveCfg = Debug|Any CPU
+		{7F8A8D87-B49A-4C7B-8474-AC83F8CFD40B}.Debug|Any CPU.Build.0 = Debug|Any CPU
+		{7F8A8D87-B49A-4C7B-8474-AC83F8CFD40B}.Release|Any CPU.ActiveCfg = Release|Any CPU
+		{7F8A8D87-B49A-4C7B-8474-AC83F8CFD40B}.Release|Any CPU.Build.0 = Release|Any CPU
 	EndGlobalSection
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE
 	EndGlobalSection
 	GlobalSection(NestedProjects) = preSolution
 		{CBA80130-6773-4DF9-995C-DC6CBED89CB5} = {41101B02-36C9-476B-98D5-1A6E105BBF4A}
-		{8B069551-9B95-464E-BB40-C56817506FEC} = {41101B02-36C9-476B-98D5-1A6E105BBF4A}
-		{8FF5E841-29F6-4DB7-A4F8-9281FBDA0B9C} = {41101B02-36C9-476B-98D5-1A6E105BBF4A}
 		{E5DB95E1-94AA-405C-9FFE-09B1E2498EE2} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
 		{7DF41C40-FE5D-41DF-B106-3DD77BE4D4B5} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
+		{1392041A-E2CA-4553-BEAF-363974651B81} = {DB29DBEF-FA4E-4334-AFB8-BFB2DA82D1DE}
 		{E3347E75-EAE8-4E6B-98D1-7230B1EE5450} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
 		{5EE18CED-28AE-4415-B5A3-C31123BF57E1} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
 		{E813073E-07A7-4C88-A505-484CB33C9DC4} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
+		{76AC3DDD-2B38-489F-A8B0-8E43054595DB} = {DB29DBEF-FA4E-4334-AFB8-BFB2DA82D1DE}
+		{7D7457FD-B88C-4375-926D-7D46C71E34A7} = {DE71D38E-4154-477C-9C27-3FA4ADB4098F}
+		{D5FF2DBA-F304-4ACB-8F82-B8F9321E22A9} = {DE71D38E-4154-477C-9C27-3FA4ADB4098F}
 		{DAD8EBB8-A1D6-4E8F-A334-D7F0273280D1} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
 		{0C045A64-AE30-47CC-A931-5B5C6C9EF06D} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
 		{19F785D2-F7A4-41AB-9301-A6AD7E40B238} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
-		{9759BE1C-98A0-4319-AC82-D432002BD66B} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
-		{DF2255F4-F671-4C15-9100-D8079992E19D} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
-		{1392041A-E2CA-4553-BEAF-363974651B81} = {DB29DBEF-FA4E-4334-AFB8-BFB2DA82D1DE}
-		{76AC3DDD-2B38-489F-A8B0-8E43054595DB} = {DB29DBEF-FA4E-4334-AFB8-BFB2DA82D1DE}
 		{6FD10BE0-24C8-456E-8B9A-FD101C05C961} = {DB29DBEF-FA4E-4334-AFB8-BFB2DA82D1DE}
-		{7D7457FD-B88C-4375-926D-7D46C71E34A7} = {DE71D38E-4154-477C-9C27-3FA4ADB4098F}
-		{D5FF2DBA-F304-4ACB-8F82-B8F9321E22A9} = {DE71D38E-4154-477C-9C27-3FA4ADB4098F}
+		{8B069551-9B95-464E-BB40-C56817506FEC} = {41101B02-36C9-476B-98D5-1A6E105BBF4A}
+		{8FF5E841-29F6-4DB7-A4F8-9281FBDA0B9C} = {41101B02-36C9-476B-98D5-1A6E105BBF4A}
+		{9759BE1C-98A0-4319-AC82-D432002BD66B} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
 		{6E942A4A-405E-4AAD-89A7-006358A8A004} = {DE71D38E-4154-477C-9C27-3FA4ADB4098F}
+		{DF2255F4-F671-4C15-9100-D8079992E19D} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
+		{EFE0C13B-6902-4FC9-91DD-F180420B36C8} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
+		{7F8A8D87-B49A-4C7B-8474-AC83F8CFD40B} = {18E62E91-73A2-48AE-BEFF-CE7C64DF759D}
 	EndGlobalSection
 	GlobalSection(ExtensibilityGlobals) = postSolution
 		SolutionGuid = {EE9FB522-4B73-4E3E-B63D-C21826BB7B5D}

src/libraries/System.Runtime.Serialization.Xml/tests/XmlDictionaryWriterTest.cs

@@ -3,6 +3,7 @@
 
 using System;
 using System.Collections.Generic;
+using System.Diagnostics;
 using System.IO;
 using System.Linq;
 using System.Runtime.Serialization;
@@ -320,6 +321,71 @@ public static void FragmentTest()
         Assert.False(FragmentHelper.CanFragment(writer));
     }
 
+    [Fact]
+    public static void XmlBaseWriter_WriteString()
+    {
+        const byte Chars8Text = 152;
+        const byte Chars16Text = 154;
+        MemoryStream ms = new MemoryStream();
+        XmlDictionaryWriter writer = (XmlDictionaryWriter)XmlDictionaryWriter.CreateBinaryWriter(ms);
+        writer.WriteStartElement("root");
+
+        int[] lengths = new[] { 7, 8, 9, 15, 16, 17, 31, 32, 36, 258 };
+        byte[] buffer = new byte[lengths.Max() + 1];
+
+        foreach (var length in lengths)
+        {
+            string allAscii = string.Create(length, null, (Span<char> chars, object _) =>
+            {
+                for (int i = 0; i < chars.Length; ++i)
+                    chars[i] = (char)(i % 128);
+            });
+            string multiByteLast = string.Create(length, null, (Span<char> chars, object _) =>
+            {
+                for (int i = 0; i < chars.Length; ++i)
+                    chars[i] = (char)(i % 128);
+                chars[^1] = '\u00E4'; // '�' - Latin Small Letter a with Diaeresis. Latin-1 Supplement.
+            });
+
+            int numBytes = Encoding.UTF8.GetBytes(allAscii, buffer);
+            Assert.True(numBytes == length, "Test setup wrong - allAscii");
+            ValidateWriteText(ms, writer, allAscii, expected: buffer.AsSpan(0, numBytes));
+
+            numBytes = Encoding.UTF8.GetBytes(multiByteLast, buffer);
+            Assert.True(numBytes == length + 1, "Test setup wrong - multiByte");
+            ValidateWriteText(ms, writer, multiByteLast, expected: buffer.AsSpan(0, numBytes));
+        }
+
+        static void ValidateWriteText(MemoryStream ms, XmlDictionaryWriter writer, string text, ReadOnlySpan<byte> expected)
+        {
+            writer.Flush();
+            ms.Seek(0, SeekOrigin.Begin);
+            ms.SetLength(0);
+            writer.WriteString(text);
+            writer.Flush();
+
+            ms.TryGetBuffer(out ArraySegment<byte> arraySegment);
+            ReadOnlySpan<byte> buffer = arraySegment;
+
+            if (expected.Length <= byte.MaxValue)
+            {
+                Assert.Equal(Chars8Text, buffer[0]);
+                Assert.Equal(expected.Length, buffer[1]);
+                buffer = buffer.Slice(2);
+            }
+            else if (expected.Length <= ushort.MaxValue)
+            {
+                Assert.Equal(Chars16Text, buffer[0]);
+                Assert.Equal(expected.Length, (int)(buffer[1]) | ((int)buffer[2] << 8));
+                buffer = buffer.Slice(3);
+            }
+            else
+                Assert.Fail("test use to long length");
+
+            AssertExtensions.SequenceEqual(expected, buffer);
+        }
+    }
+
     private static bool ReadTest(MemoryStream ms, Encoding encoding, ReaderWriterFactory.ReaderWriterType rwType, byte[] byteArray)
     {
         ms.Position = 0;