diff --git a/src/System.Private.CoreLib/shared/System/SpanHelpers.Char.cs b/src/System.Private.CoreLib/shared/System/SpanHelpers.Char.cs index 409c01044eaf..475d59fc44b9 100644 --- a/src/System.Private.CoreLib/shared/System/SpanHelpers.Char.cs +++ b/src/System.Private.CoreLib/shared/System/SpanHelpers.Char.cs @@ -311,101 +311,226 @@ public static unsafe int IndexOf(ref char searchSpace, char value, int length) } [MethodImpl(MethodImplOptions.AggressiveOptimization)] - public static unsafe int IndexOfAny(ref char searchSpace, char value0, char value1, int length) + public static int IndexOfAny(ref char searchSpace, char value0, char value1, int length) { Debug.Assert(length >= 0); - fixed (char* pChars = &searchSpace) - { - char* pCh = pChars; - char* pEndCh = pCh + length; + nint offset = 0; + nint lengthToExamine = length; - if (Vector.IsHardwareAccelerated && length >= Vector.Count * 2) + if (Sse2.IsSupported) + { + // Avx2 branch also operates on Sse2 sizes, so check is combined. + // Needs to be double length to allow us to align the data first. + if (length >= Vector128.Count * 2) { - // Figure out how many characters to read sequentially until we are vector aligned - // This is equivalent to: - // unaligned = ((int)pCh % Unsafe.SizeOf>()) / elementsPerByte - // length = (Vector.Count - unaligned) % Vector.Count - const int elementsPerByte = sizeof(ushort) / sizeof(byte); - int unaligned = ((int)pCh & (Unsafe.SizeOf>() - 1)) / elementsPerByte; - length = (Vector.Count - unaligned) & (Vector.Count - 1); + lengthToExamine = UnalignedCountVector128(ref searchSpace); } - - SequentialScan: - while (length >= 4) + } + else if (Vector.IsHardwareAccelerated) + { + // Needs to be double length to allow us to align the data first. + if (length >= Vector.Count * 2) { - length -= 4; + lengthToExamine = UnalignedCountVector(ref searchSpace); + } + } - if (pCh[0] == value0 || pCh[0] == value1) - goto Found; - if (pCh[1] == value0 || pCh[1] == value1) - goto Found1; - if (pCh[2] == value0 || pCh[2] == value1) - goto Found2; - if (pCh[3] == value0 || pCh[3] == value1) - goto Found3; + SequentialScan: + // In the non-vector case lengthToExamine is the total length. + // In the vector case lengthToExamine first aligns to Vector, + // then in a second pass after the Vector lengths is the + // remaining data that is shorter than a Vector length. + int lookUp; + while (lengthToExamine >= 4) + { + ref char current = ref Add(ref searchSpace, offset); + + lookUp = current; + if (value0 == lookUp || value1 == lookUp) + goto Found; + lookUp = Add(ref current, 1); + if (value0 == lookUp || value1 == lookUp) + goto Found1; + lookUp = Add(ref current, 2); + if (value0 == lookUp || value1 == lookUp) + goto Found2; + lookUp = Add(ref current, 3); + if (value0 == lookUp || value1 == lookUp) + goto Found3; + + offset += 4; + lengthToExamine -= 4; + } - pCh += 4; - } + while (lengthToExamine > 0) + { + lookUp = Add(ref searchSpace, offset); + if (value0 == lookUp || value1 == lookUp) + goto Found; - while (length > 0) + offset += 1; + lengthToExamine -= 1; + } + + // We get past SequentialScan only if IsHardwareAccelerated or intrinsic .IsSupported is true. However, we still have the redundant check to allow + // the JIT to see that the code is unreachable and eliminate it when the platform does not have hardware accelerated. + if (Avx2.IsSupported) + { + if (offset < length) { - length--; + lengthToExamine = GetCharVector256SpanLength(offset, length); + if (lengthToExamine > 0) + { + Debug.Assert(length - offset >= Vector256.Count); - if (pCh[0] == value0 || pCh[0] == value1) - goto Found; + Vector256 values0 = Vector256.Create(value0); + Vector256 values1 = Vector256.Create(value1); + do + { + Vector256 search = LoadVector256(ref searchSpace, offset); + // Bitwise Or to combine the flagged matches for the second value to our match flags + int matches = Avx2.MoveMask( + Avx2.Or( + Avx2.CompareEqual(values0, search), + Avx2.CompareEqual(values1, search)) + .AsByte()); + // Note that MoveMask has converted the equal vector elements into a set of bit flags, + // So the bit position in 'matches' corresponds to the element offset. + if (matches == 0) + { + // Zero flags set so no matches + offset += Vector256.Count; + lengthToExamine -= Vector256.Count; + continue; + } + + // Find bitflag offset of first match and add to current offset, + // flags are in bytes so divide for chars + return (int)(offset + (BitOperations.TrailingZeroCount(matches) / sizeof(char))); + } while (lengthToExamine > 0); + } - pCh++; + lengthToExamine = GetCharVector128SpanLength(offset, length); + if (lengthToExamine > 0) + { + Debug.Assert(length - offset >= Vector128.Count); + + Vector128 values0 = Vector128.Create(value0); + Vector128 values1 = Vector128.Create(value1); + Vector128 search = LoadVector128(ref searchSpace, offset); + + // Same method as above + int matches = Sse2.MoveMask( + Sse2.Or( + Sse2.CompareEqual(values0, search), + Sse2.CompareEqual(values1, search)) + .AsByte()); + if (matches == 0) + { + // Zero flags set so no matches + offset += Vector128.Count; + // Don't need to change lengthToExamine here as we don't use its current value again. + } + else + { + // Find bitflag offset of first match and add to current offset, + // flags are in bytes so divide for chars + return (int)(offset + (BitOperations.TrailingZeroCount(matches) / sizeof(char))); + } + } + + if (offset < length) + { + lengthToExamine = length - offset; + goto SequentialScan; + } } + } + else if (Sse2.IsSupported) + { + if (offset < length) + { + lengthToExamine = GetCharVector128SpanLength(offset, length); - // We get past SequentialScan only if IsHardwareAccelerated is true. However, we still have the redundant check to allow - // the JIT to see that the code is unreachable and eliminate it when the platform does not have hardware accelerated. - if (Vector.IsHardwareAccelerated && pCh < pEndCh) + Vector128 values0 = Vector128.Create(value0); + Vector128 values1 = Vector128.Create(value1); + while (lengthToExamine > 0) + { + Debug.Assert(length - offset >= Vector128.Count); + + Vector128 search = LoadVector128(ref searchSpace, offset); + + // Same method as above + int matches = Sse2.MoveMask( + Sse2.Or( + Sse2.CompareEqual(values0, search), + Sse2.CompareEqual(values1, search)) + .AsByte()); + if (matches == 0) + { + // Zero flags set so no matches + offset += Vector128.Count; + lengthToExamine -= Vector128.Count; + continue; + } + + // Find bitflag offset of first match and add to current offset, + // flags are in bytes so divide for chars + return (int)(offset + (BitOperations.TrailingZeroCount(matches) / sizeof(char))); + } + + if (offset < length) + { + lengthToExamine = length - offset; + goto SequentialScan; + } + } + } + else if (Vector.IsHardwareAccelerated) + { + if (offset < length) { - // Get the highest multiple of Vector.Count that is within the search space. - // That will be how many times we iterate in the loop below. - // This is equivalent to: length = Vector.Count * ((int)(pEndCh - pCh) / Vector.Count) - length = (int)((pEndCh - pCh) & ~(Vector.Count - 1)); + lengthToExamine = GetCharVectorSpanLength(offset, length); - // Get comparison Vector Vector values0 = new Vector(value0); Vector values1 = new Vector(value1); - while (length > 0) + while (lengthToExamine > 0) { - // Using Unsafe.Read instead of ReadUnaligned since the search space is pinned and pCh is always vector aligned - Debug.Assert(((int)pCh & (Unsafe.SizeOf>() - 1)) == 0); - Vector vData = Unsafe.Read>(pCh); - var vMatches = Vector.BitwiseOr( - Vector.Equals(vData, values0), - Vector.Equals(vData, values1)); - if (Vector.Zero.Equals(vMatches)) + Debug.Assert(length - offset >= Vector.Count); + + Vector search = LoadVector(ref searchSpace, offset); + var matches = Vector.BitwiseOr( + Vector.Equals(search, values0), + Vector.Equals(search, values1)); + if (Vector.Zero.Equals(matches)) { - pCh += Vector.Count; - length -= Vector.Count; + offset += Vector.Count; + lengthToExamine -= Vector.Count; continue; } + // Find offset of first match - return (int)(pCh - pChars) + LocateFirstFoundChar(vMatches); + return (int)(offset + LocateFirstFoundChar(matches)); } - if (pCh < pEndCh) + if (offset < length) { - length = (int)(pEndCh - pCh); + lengthToExamine = length - offset; goto SequentialScan; } } - - return -1; - Found3: - pCh++; - Found2: - pCh++; - Found1: - pCh++; - Found: - return (int)(pCh - pChars); } + return -1; + Found3: + return (int)(offset + 3); + Found2: + return (int)(offset + 2); + Found1: + return (int)(offset + 1); + Found: + return (int)offset; } [MethodImpl(MethodImplOptions.AggressiveOptimization)]