NonBacktracking Regex optimizations (#102655)

- Raise limits for when to switch from DFA to NFA and for when to fail to build because the predicted NFA size is too large to handle.
- Cache nullability information for each minterm to make computing it cheaper.
- Cache minterm information for all characters that map to a minterm other than zero, avoiding expensive computation per step for each non-ASCII character encountered.
- Reduce inner hot loop overhead for patterns not containing an end Z anchor and for having fewer than 256 minterms (more than that is rare).
- Reduce the frequency of timeout checks that were both costly and unnecessarily frequent to achieve the goal

---------

Co-authored-by: ieviev <ieviev@users.noreply.github.com>
Co-authored-by: Stephen Toub <stoub@microsoft.com>
Co-authored-by: Dan Moseley <danmose@microsoft.com>

Author: 3 people

src/libraries/System.Text.RegularExpressions/src/System.Text.RegularExpressions.csproj

@@ -72,6 +72,8 @@
     <Compile Include="System\Text\RegularExpressions\Symbolic\MatchingState.cs" />
     <Compile Include="System\Text\RegularExpressions\Symbolic\DoublyLinkedList.cs" />
     <Compile Include="System\Text\RegularExpressions\Symbolic\ISolver.cs" />
+    <Compile Include="System\Text\RegularExpressions\Symbolic\MatchReversalKind.cs"/>
+    <Compile Include="System\Text\RegularExpressions\Symbolic\MatchReversal.cs"/>
     <Compile Include="System\Text\RegularExpressions\Symbolic\MintermClassifier.cs" />
     <Compile Include="System\Text\RegularExpressions\Symbolic\MintermGenerator.cs" />
     <Compile Include="System\Text\RegularExpressions\Symbolic\RegexNodeConverter.cs" />

src/libraries/System.Text.RegularExpressions/src/System/Text/RegularExpressions/RegexReplacement.cs

@@ -5,7 +5,6 @@
 using System.Collections.Generic;
 using System.Diagnostics;
 using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
 
 #pragma warning disable CS8500 // takes address of managed type
 

src/libraries/System.Text.RegularExpressions/src/System/Text/RegularExpressions/Symbolic/BitVectorSolver.cs

@@ -10,11 +10,11 @@ internal sealed class BitVectorSolver : ISolver<BitVector>
         internal readonly MintermClassifier _classifier;
         private readonly BitVector[] _mintermVectors;
 
-        public BitVectorSolver(BDD[] minterms, CharSetSolver solver)
+        public BitVectorSolver(BDD[] minterms)
         {
             _minterms = minterms;
 
-            _classifier = new MintermClassifier(minterms, solver);
+            _classifier = new MintermClassifier(minterms);
 
             var singleBitVectors = new BitVector[minterms.Length];
             for (int i = 0; i < singleBitVectors.Length; i++)

src/libraries/System.Text.RegularExpressions/src/System/Text/RegularExpressions/Symbolic/MatchReversal.cs

@@ -0,0 +1,39 @@
+// Licensed to the .NET Foundation under one or more agreements.
+// The .NET Foundation licenses this file to you under the MIT license.
+
+using System.Diagnostics;
+
+namespace System.Text.RegularExpressions.Symbolic
+{
+    /// <summary>Provides details on how a match may be processed in reverse to find the beginning of a match once a match's existence has been confirmed.</summary>
+    internal readonly struct MatchReversalInfo<TSet> where TSet : IComparable<TSet>, IEquatable<TSet>
+    {
+        /// <summary>Initializes the match reversal details.</summary>
+        internal MatchReversalInfo(MatchReversalKind kind, int fixedLength, MatchingState<TSet>? adjustedStartState = null)
+        {
+            Debug.Assert(kind is MatchReversalKind.MatchStart or MatchReversalKind.FixedLength or MatchReversalKind.PartialFixedLength);
+            Debug.Assert(fixedLength >= 0);
+            Debug.Assert((adjustedStartState is not null) == (kind is MatchReversalKind.PartialFixedLength));
+
+            Kind = kind;
+            FixedLength = fixedLength;
+            AdjustedStartState = adjustedStartState;
+        }
+
+        /// <summary>Gets the kind of the match reversal processing required.</summary>
+        internal MatchReversalKind Kind { get; }
+
+        /// <summary>Gets the fixed length of the match, if one is known.</summary>
+        /// <remarks>
+        /// For <see cref="MatchReversalKind.MatchStart"/>, this is ignored.
+        /// For <see cref="MatchReversalKind.FixedLength"/>, this is the full length of the match. The beginning may be found simply
+        /// by subtracting this length from the end.
+        /// For <see cref="MatchReversalKind.PartialFixedLength"/>, this is the length of fixed portion of the match.
+        /// </remarks>
+        internal int FixedLength { get; }
+
+        /// <summary>Gets the adjusted start state to use for partial fixed-length matches.</summary>
+        /// <remarks>This will be non-null iff <see cref="Kind"/> is <see cref="MatchReversalKind.PartialFixedLength"/>.</remarks>
+        internal MatchingState<TSet>? AdjustedStartState { get; }
+    }
+}

src/libraries/System.Text.RegularExpressions/src/System/Text/RegularExpressions/Symbolic/MatchReversalKind.cs

@@ -0,0 +1,26 @@
+// Licensed to the .NET Foundation under one or more agreements.
+// The .NET Foundation licenses this file to you under the MIT license.
+
+namespace System.Text.RegularExpressions.Symbolic
+{
+    /// <summary>Specifies the kind of a <see cref="MatchReversalInfo{TSet}"/>.</summary>
+    internal enum MatchReversalKind
+    {
+        /// <summary>The regex should be run in reverse to find beginning of the match.</summary>
+        MatchStart,
+
+        /// <summary>The end of the pattern is of a fixed length and can be skipped as part of running a regex in reverse to find the beginning of the match.</summary>
+        /// <remarks>
+        /// Reverse execution is not necessary for a subset of the match.
+        /// <see cref="MatchReversalInfo{TSet}.FixedLength"/> will contain the length of the fixed portion.
+        /// </remarks>
+        PartialFixedLength,
+
+        /// <summary>The entire pattern is of a fixed length.</summary>
+        /// <remarks>
+        /// Reverse execution is not necessary to find the beginning of the match.
+        /// <see cref="MatchReversalInfo{TSet}.FixedLength"/> will contain the length of the match.
+        /// </remarks>
+        FixedLength
+    }
+}

src/libraries/System.Text.RegularExpressions/src/System/Text/RegularExpressions/Symbolic/MatchingState.cs

@@ -14,6 +14,7 @@ internal MatchingState(SymbolicRegexNode<TSet> node, uint prevCharKind)
         {
             Node = node;
             PrevCharKind = prevCharKind;
+            NullabilityInfo = BuildNullabilityInfo();
         }
 
         /// <summary>The regular expression that labels this state and gives it its semantics.</summary>
@@ -95,21 +96,37 @@ internal SymbolicRegexNode<TSet> Next(SymbolicRegexBuilder<TSet> builder, TSet m
             return Node.CreateNfaDerivativeWithEffects(builder, minterm, context);
         }
 
+        /// <summary>Determines whether the node is nullable for the given context.</summary>
+        /// <remarks>
+        /// This is functionally equivalent to <see cref="SymbolicRegexNode{TSet}.IsNullableFor(uint)"/>, but using cached
+        /// answers stored in <see cref="NullabilityInfo"/>.
+        /// </remarks>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         internal bool IsNullableFor(uint nextCharKind)
         {
-            Debug.Assert(CharKind.IsValidCharKind(nextCharKind));
-            uint context = CharKind.Context(PrevCharKind, nextCharKind);
-            return Node.IsNullableFor(context);
+            Debug.Assert(nextCharKind is >= 0 and < CharKind.CharKindCount);
+            return (NullabilityInfo & (1 << (int)nextCharKind)) != 0;
         }
 
+        /// <summary>Gets the nullability info for the matching state.</summary>
+        /// <remarks>
+        /// <list>
+        /// <item>00000 -> node cannot be nullable</item>
+        /// <item>00001 -> nullable for General</item>
+        /// <item>00010 -> nullable for BeginningEnd</item>
+        /// <item>00100 -> nullable for NewLine</item>
+        /// <item>01000 -> nullable for NewLineS</item>
+        /// <item>10000 -> nullable for WordLetter</item>
+        /// </list>
+        /// </remarks>
+        internal int NullabilityInfo { get; }
+
         /// <summary>
         /// Builds a <see cref="StateFlags"/> with the relevant flags set.
         /// </summary>
-        /// <param name="solver">a solver for <typeparamref name="TSet"/></param>
         /// <param name="isInitial">whether this state is an initial state</param>
         /// <returns>the flags for this matching state</returns>
-        internal StateFlags BuildStateFlags(ISolver<TSet> solver, bool isInitial)
+        internal StateFlags BuildStateFlags(bool isInitial)
         {
             StateFlags info = 0;
 
@@ -118,11 +135,6 @@ internal StateFlags BuildStateFlags(ISolver<TSet> solver, bool isInitial)
                 info |= StateFlags.IsInitialFlag;
             }
 
-            if (IsDeadend(solver))
-            {
-                info |= StateFlags.IsDeadendFlag;
-            }
-
             if (Node.CanBeNullable)
             {
                 info |= StateFlags.CanBeNullableFlag;
@@ -140,6 +152,22 @@ internal StateFlags BuildStateFlags(ISolver<TSet> solver, bool isInitial)
             return info;
         }
 
+        /// <summary>Builds the nullability information for the matching state.</summary>
+        /// <remarks>Nullability for each context is encoded in a bit. See <see cref="NullabilityInfo"/>.</remarks>
+        private byte BuildNullabilityInfo()
+        {
+            byte nullabilityInfo = 0;
+            if (Node.CanBeNullable)
+            {
+                for (uint charKind = 0; charKind < CharKind.CharKindCount; charKind++)
+                {
+                    nullabilityInfo |= (byte)(Node.IsNullableFor(CharKind.Context(PrevCharKind, charKind)) ? 1 << (int)charKind : 0);
+                }
+            }
+
+            return nullabilityInfo;
+        }
+
         public override bool Equals(object? obj) =>
             obj is MatchingState<TSet> s && PrevCharKind == s.PrevCharKind && Node.Equals(s.Node);
 

src/libraries/System.Text.RegularExpressions/src/System/Text/RegularExpressions/Symbolic/MintermClassifier.cs

@@ -1,7 +1,9 @@
 // Licensed to the .NET Foundation under one or more agreements.
 // The .NET Foundation licenses this file to you under the MIT license.
 
+using System.Buffers;
 using System.Diagnostics;
+using System.Numerics;
 using System.Runtime.CompilerServices;
 
 namespace System.Text.RegularExpressions.Symbolic
@@ -20,81 +22,104 @@ namespace System.Text.RegularExpressions.Symbolic
     /// </remarks>
     internal sealed class MintermClassifier
     {
-        /// <summary>An array used when there's a single minterm, in order to map every ASCII character to it trivially.</summary>
-        private static readonly int[] AllAsciiIsZeroMintermArray = new int[128];
+        /// <summary>Mapping for characters to minterms, used in the vast majority case when there are less than 256 minterms.</summary>
+        /// <remarks>_lookup[char] provides the minterm ID. If char &gt;= _lookup.Length, its minterm is 0.</remarks>
+        private readonly byte[]? _lookup;
 
-        /// <summary>Array providing fast mapping from an ASCII character (the array index) to its corresponding minterm ID.</summary>
-        private readonly int[] _ascii;
-        /// <summary>A multi-terminal BDD for mapping any non-ASCII character to its associated minterm ID.</summary>
-        /// <remarks>
-        /// The use of a multi-terminal BDD here is an implementation detail.  Should we decide its important to optimize non-ASCII inputs further,
-        /// or to consolidate the mechanism with the other engines, an alternatie lookup algorithm / data structure could be employed.
-        /// </remarks>
-        private readonly BDD _nonAscii;
+        /// <summary>Mapping for characters to minterms, used when there are at least 256 minterms. This is rarely used.</summary>
+        /// <remarks>_intLookup[char] provides the minterm ID. If char &gt;= _intLookup.Length, its minterm is 0.</remarks>
+        private readonly int[]? _intLookup;
 
         /// <summary>Create a classifier that maps a character to the ID of its associated minterm.</summary>
         /// <param name="minterms">A BDD for classifying all characters (ASCII and non-ASCII) to their corresponding minterm IDs.</param>
-        /// <param name="solver">The character set solver to use.</param>
-        public MintermClassifier(BDD[] minterms, CharSetSolver solver)
+        public MintermClassifier(BDD[] minterms)
         {
             Debug.Assert(minterms.Length > 0, "Requires at least");
 
             if (minterms.Length == 1)
             {
                 // With only a single minterm, the mapping is trivial: everything maps to it (ID 0).
-                // For ASCII, use an array containing all zeros.  For non-ASCII, use a BDD that maps everything to 0.
-                _ascii = AllAsciiIsZeroMintermArray;
-                _nonAscii = solver.ReplaceTrue(BDD.True, 0);
+                _lookup = [];
                 return;
             }
 
-            // Create a multi-terminal BDD for mapping any character to its associated minterm.
-            BDD anyCharacterToMintermId = BDD.False;
-            for (int i = 0; i < minterms.Length; i++)
-            {
-                // Each supplied minterm BDD decides whether a given character maps to it or not.
-                // We need to combine all of those into a multi-terminal BDD that decides which
-                // minterm a character maps to.  To do that, we take each minterm BDD and replace
-                // its True result with the ID of the minterm, such that a character that would
-                // have returned True for that BDD now returns the minterm ID.
-                BDD charToTargetMintermId = solver.ReplaceTrue(minterms[i], i);
+            // Compute all minterm ranges. We do this here in order to determine the maximum character value
+            // in order to size the lookup array to minimize steady-state memory consumption of the potentially
+            // large lookup array. We prefer to use the byte[] _lookup when possible, in order to keep memory
+            // consumption to a minimum; doing so accomodates up to 255 minterms, which is the vast majority case.
+            // However, when there are more than 255 minterms, we need to use int[] _intLookup.
+            (uint, uint)[][] charRangesPerMinterm = ArrayPool<(uint, uint)[]>.Shared.Rent(minterms.Length);
 
-                // Now union this BDD with the multi-terminal BDD we've built up thus far. Unioning
-                // is valid because every character belongs to exactly one minterm and thus will
-                // only map to an ID instead of False in exactly one of the input BDDs.
-                anyCharacterToMintermId = solver.Or(anyCharacterToMintermId, charToTargetMintermId);
+            int maxChar = -1;
+            for (int mintermId = 1; mintermId < minterms.Length; mintermId++)
+            {
+                (uint, uint)[] ranges = BDDRangeConverter.ToRanges(minterms[mintermId]);
+                charRangesPerMinterm[mintermId] = ranges;
+                maxChar = Math.Max(maxChar, (int)ranges[^1].Item2);
             }
 
-            // Now that we have our mapping that supports any input character, we want to optimize for
-            // ASCII inputs.  Rather than forcing every input ASCII character to consult the BDD at match
-            // time, we precompute a lookup table, where each ASCII character can be used to index into the
-            // array to determine the ID for its corresponding minterm.
-            var ascii = new int[128];
-            for (int i = 0; i < ascii.Length; i++)
+            // It's incredibly rare for a regex to use more than a couple hundred minterms,
+            // but we need a fallback just in case. (Over 128 unique sets also means it's never ASCII only.)
+            if (minterms.Length > 255)
+            {
+                _intLookup = CreateLookup<int>(minterms, charRangesPerMinterm, maxChar);
+            }
+            else
             {
-                ascii[i] = anyCharacterToMintermId.Find(i);
+                _lookup = CreateLookup<byte>(minterms, charRangesPerMinterm, maxChar);
             }
-            _ascii = ascii;
 
-            // We can also further optimize the BDD in two ways:
-            // 1. We can now remove the ASCII characters from it, as we'll always consult the lookup table first
-            //    for ASCII inputs and thus will never use the BDD for them.  While optional (skipping this step will not
-            //    affect correctness), removing the ASCII values from the BDD reduces the size of the multi-terminal BDD.
-            // 2. We can check if every character now maps to the same minterm ID (the same terminal in the
-            //    multi-terminal BDD).  This can be relatively common after (1) above is applied, as many
-            //    patterns don't distinguish between any non-ASCII characters (e.g. "[0-9]*").  If every character
-            //    in the BDD now maps to the same minterm, we can replace the BDD with a much simpler/faster/smaller one.
-            BDD nonAsciiBDD = solver.And(anyCharacterToMintermId, solver.NonAscii);
-            nonAsciiBDD = nonAsciiBDD.IsEssentiallyBoolean(out BDD? singleTerminalBDD) ? singleTerminalBDD : nonAsciiBDD;
-            _nonAscii = nonAsciiBDD;
+            // Return the rented array. We clear it before returning it in order to avoid all the ranges arrays being kept alive.
+            Array.Clear(charRangesPerMinterm, 0, minterms.Length);
+            ArrayPool<(uint, uint)[]>.Shared.Return(charRangesPerMinterm);
+
+            // Creates the lookup array.
+            static T[] CreateLookup<T>(BDD[] minterms, ReadOnlySpan<(uint, uint)[]> charRangesPerMinterm, int _maxChar) where T : IBinaryInteger<T>
+            {
+                T[] lookup = new T[_maxChar + 1];
+                for (int mintermId = 1; mintermId < minterms.Length; mintermId++)
+                {
+                    // Each minterm maps to a range of characters. Set each of the characters in those ranges to the corresponding minterm.
+                    foreach ((uint start, uint end) in charRangesPerMinterm[mintermId])
+                    {
+                        lookup.AsSpan((int)start, (int)(end + 1 - start)).Fill(T.CreateTruncating(mintermId));
+                    }
+                }
+
+                return lookup;
+            }
         }
 
-        /// <summary>Gets the ID of the minterm associated with the specified character.</summary>
+        /// <summary>Gets the ID of the minterm associated with the specified character. </summary>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public int GetMintermID(int c)
         {
-            int[] ascii = _ascii;
-            return (uint)c < (uint)ascii.Length ? ascii[c] : _nonAscii.Find(c);
+            if (_lookup is not null)
+            {
+                byte[] lookup = _lookup;
+                return (uint)c < (uint)lookup.Length ? lookup[c] : 0;
+            }
+            else
+            {
+                int[] lookup = _intLookup!;
+                return (uint)c < (uint)lookup.Length ? lookup[c] : 0;
+            }
         }
+        /// <summary>
+        /// Gets a quick mapping from char to minterm for the common case when there are &lt;= 255 minterms.
+        /// Null if there are greater than 255 minterms.
+        /// </summary>
+        public byte[]? ByteLookup => _lookup;
+
+        /// <summary>
+        /// Gets a mapping from char to minterm for the rare case when there are &gt;= 255 minterms.
+        /// Null in the common case where there are fewer than 255 minterms.
+        /// </summary>
+        public int[]? IntLookup => _intLookup;
+
+        /// <summary>
+        /// Maximum ordinal character for a non-0 minterm, used to conserve memory
+        /// </summary>
+        public int MaxChar => (_lookup?.Length ?? _intLookup!.Length) - 1;
     }
 }