Add fast path to String.Equals

src/coreclr/jit/importercalls.cpp

@@ -4701,6 +4701,7 @@ GenTree* Compiler::impIntrinsic(CORINFO_CLASS_HANDLE    clsHnd,
             case NI_System_SpanHelpers_ClearWithoutReferences:
             case NI_System_SpanHelpers_Memmove:
             {
+                mustExpand = false; // It's fine for these intrinsics to call themselves recursively
                 if (sig->sigInst.methInstCount == 0)
                 {
                     // We'll try to unroll this in lower for constant input.

src/libraries/System.Private.CoreLib/src/System/SpanHelpers.Byte.cs

@@ -7,6 +7,7 @@
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.Arm;
 using System.Runtime.Intrinsics.X86;
 
 namespace System
@@ -754,6 +755,53 @@ internal static unsafe int IndexOfNullByte(byte* searchSpace)
             return (int)(offset + 7);
         }
 
+        private static unsafe bool SequenceEqual_LongInput<TVector>(ref byte first, ref byte second, nuint length)
+            where TVector : struct, ISimdVector<TVector, byte>
+        {
+            Debug.Assert(length >= (nuint)sizeof(TVector) * 3);
+            Debug.Assert(TVector.IsHardwareAccelerated);
+            Debug.Assert(sizeof(TVector) >= 16);
+
+            fixed (byte* pFirst = &first, pSecond = &second)
+            {
+                // First, align pFirst to sizeof(TVector) bytes boundary
+                if (!TVector.EqualsAll(TVector.Load(pFirst), TVector.Load(pSecond)))
+                {
+                    return false;
+                }
+
+                nuint misalignedElements = (nuint)pFirst % (nuint)sizeof(TVector);
+                nuint offset = (nuint)sizeof(TVector) - misalignedElements;
+                nuint lengthToExamine = length - offset - (nuint)sizeof(TVector) * 2;
+                do
+                {
+                    // Since pFirst is aligned, we may consider using NonTemporalAligned loads for it.
+                    TVector firstV1 = TVector.Load(pFirst + offset);
+                    TVector firstV2 = TVector.Load(pFirst + offset + (nuint)sizeof(TVector));
+                    TVector second1 = TVector.Load(pSecond + offset);
+                    TVector second2 = TVector.Load(pSecond + offset + (nuint)sizeof(TVector));
+                    if (!TVector.EqualsAll((firstV1 ^ second1) | (firstV2 ^ second2), TVector.Zero))
+                    {
+                        return false;
+                    }
+                    offset += (nuint)sizeof(TVector) * 2;
+                } while (lengthToExamine > offset);
+
+                // Last iteration
+                {
+                    TVector firstV1 = TVector.Load(pFirst + lengthToExamine);
+                    TVector firstV2 = TVector.Load(pFirst + lengthToExamine + (nuint)sizeof(TVector));
+                    TVector second1 = TVector.Load(pSecond + lengthToExamine);
+                    TVector second2 = TVector.Load(pSecond + lengthToExamine + (nuint)sizeof(TVector));
+                    if (TVector.EqualsAll((firstV1 ^ second1) | (firstV2 ^ second2), TVector.Zero))
+                    {
+                        return true;
+                    }
+                }
+            }
+            return false;
+        }
+
         // Optimized byte-based SequenceEquals. The "length" parameter for this one is declared a nuint rather than int as we also use it for types other than byte
         // where the length can exceed 2Gb once scaled by sizeof(T).
         [Intrinsic] // Unrolled for constant length
@@ -822,6 +870,11 @@ public static unsafe bool SequenceEqual(ref byte first, ref byte second, nuint l
             {
                 if (Vector512.IsHardwareAccelerated && length >= (nuint)Vector512<byte>.Count)
                 {
+                    if (length >= 512) // After this threshold, _LongInput will be faster
+                    {
+                        return SequenceEqual_LongInput<Vector512<byte>>(ref first, ref second, length);
+                    }
+
                     nuint offset = 0;
                     nuint lengthToExamine = length - (nuint)Vector512<byte>.Count;
                     // Unsigned, so it shouldn't have overflowed larger than length (rather than negative)
@@ -852,6 +905,10 @@ public static unsafe bool SequenceEqual(ref byte first, ref byte second, nuint l
                 }
                 else if (Vector256.IsHardwareAccelerated && length >= (nuint)Vector256<byte>.Count)
                 {
+                    if (!Vector512.IsHardwareAccelerated && length >= 256) // After this threshold, _LongInput will be faster
+                    {
+                        return SequenceEqual_LongInput<Vector256<byte>>(ref first, ref second, length);
+                    }
                     nuint offset = 0;
                     nuint lengthToExamine = length - (nuint)Vector256<byte>.Count;
                     // Unsigned, so it shouldn't have overflowed larger than length (rather than negative)
@@ -882,6 +939,14 @@ public static unsafe bool SequenceEqual(ref byte first, ref byte second, nuint l
                 }
                 else if (length >= (nuint)Vector128<byte>.Count)
                 {
+#if !MONO // Mono has performance issues with ISimdVector<T>
+                    if (((!Vector512.IsHardwareAccelerated && !Vector256.IsHardwareAccelerated) || AdvSimd.Arm64.IsSupported) &&
+                        length >= 64) // After this threshold, _LongInput will be faster
+                    {
+                        return SequenceEqual_LongInput<Vector128<byte>>(ref first, ref second, length);
+                    }
+#endif
+
                     nuint offset = 0;
                     nuint lengthToExamine = length - (nuint)Vector128<byte>.Count;
                     // Unsigned, so it shouldn't have overflowed larger than length (rather than negative)

src/libraries/System.Private.CoreLib/src/System/String.Comparison.cs

@@ -26,6 +26,25 @@ private static bool EqualsHelper(string strA, string strB)
             Debug.Assert(strB != null);
             Debug.Assert(strA.Length == strB.Length);
 
+            /*nuint byteLength = (nuint)strA.Length * sizeof(char);
+            ref byte pStrA = ref Unsafe.As<char, byte>(ref strA.GetRawStringData());
+            ref byte pStrB = ref Unsafe.As<char, byte>(ref strB.GetRawStringData());
+
+            // Fast path: if the strings have more than 2 characters, we can use a fast path
+            // where we compare 4 bytes at a time, which gives us either an early mismatch or
+            // aligns pStrA and pStrB to pointer size since String._firstChar has only 4-byte alignment.
+            if (byteLength > (sizeof(char) * 2))
+            {
+                if (Unsafe.ReadUnaligned<uint>(ref pStrA) != Unsafe.ReadUnaligned<uint>(ref pStrB))
+                {
+                    return false;
+                }
+                pStrA = ref Unsafe.Add(ref pStrA, sizeof(char) * 2);
+                pStrB = ref Unsafe.Add(ref pStrB, sizeof(char) * 2);
+                byteLength -= sizeof(char) * 2;
+            }
+            return SpanHelpers.SequenceEqual(ref pStrA, ref pStrB, byteLength);*/
+
             return SpanHelpers.SequenceEqual(
                     ref Unsafe.As<char, byte>(ref strA.GetRawStringData()),
                     ref Unsafe.As<char, byte>(ref strB.GetRawStringData()),