brintos

brintos / llvm-project-archived public Read only

0
0
Text · 15.2 KiB · cf5182b Raw
400 lines · cpp
1//===--- FuzzyMatch.h - Approximate identifier matching  ---------*- C++-*-===//2//3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.4// See https://llvm.org/LICENSE.txt for license information.5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception6//7//===----------------------------------------------------------------------===//8//9// To check for a match between a Pattern ('u_p') and a Word ('unique_ptr'),10// we consider the possible partial match states:11//12//     u n i q u e _ p t r13//   +---------------------14//   |A . . . . . . . . . .15//  u|16//   |. . . . . . . . . . .17//  _|18//   |. . . . . . . O . . .19//  p|20//   |. . . . . . . . . . B21//22// Each dot represents some prefix of the pattern being matched against some23// prefix of the word.24//   - A is the initial state: '' matched against ''25//   - O is an intermediate state: 'u_' matched against 'unique_'26//   - B is the target state: 'u_p' matched against 'unique_ptr'27//28// We aim to find the best path from A->B.29//  - Moving right (consuming a word character)30//    Always legal: not all word characters must match.31//  - Moving diagonally (consuming both a word and pattern character)32//    Legal if the characters match.33//  - Moving down (consuming a pattern character) is never legal.34//    Never legal: all pattern characters must match something.35// Characters are matched case-insensitively.36// The first pattern character may only match the start of a word segment.37//38// The scoring is based on heuristics:39//  - when matching a character, apply a bonus or penalty depending on the40//    match quality (does case match, do word segments align, etc)41//  - when skipping a character, apply a penalty if it hurts the match42//    (it starts a word segment, or splits the matched region, etc)43//44// These heuristics require the ability to "look backward" one character, to45// see whether it was matched or not. Therefore the dynamic-programming matrix46// has an extra dimension (last character matched).47// Each entry also has an additional flag indicating whether the last-but-one48// character matched, which is needed to trace back through the scoring table49// and reconstruct the match.50//51// We treat strings as byte-sequences, so only ASCII has first-class support.52//53// This algorithm was inspired by VS code's client-side filtering, and aims54// to be mostly-compatible.55//56//===----------------------------------------------------------------------===//57 58#include "FuzzyMatch.h"59#include "llvm/Support/Format.h"60#include <optional>61 62namespace clang {63namespace clangd {64 65static char lower(char C) { return C >= 'A' && C <= 'Z' ? C + ('a' - 'A') : C; }66// A "negative infinity" score that won't overflow.67// We use this to mark unreachable states and forbidden solutions.68// Score field is 15 bits wide, min value is -2^14, we use half of that.69static constexpr int AwfulScore = -(1 << 13);70static bool isAwful(int S) { return S < AwfulScore / 2; }71static constexpr int PerfectBonus = 4; // Perfect per-pattern-char score.72 73FuzzyMatcher::FuzzyMatcher(llvm::StringRef Pattern)74    : PatN(std::min<int>(MaxPat, Pattern.size())),75      ScoreScale(PatN ? float{1} / (PerfectBonus * PatN) : 0), WordN(0) {76  std::copy(Pattern.begin(), Pattern.begin() + PatN, Pat);77  for (int I = 0; I < PatN; ++I)78    LowPat[I] = lower(Pat[I]);79  Scores[0][0][Miss] = {0, Miss};80  Scores[0][0][Match] = {AwfulScore, Miss};81  for (int P = 0; P <= PatN; ++P)82    for (int W = 0; W < P; ++W)83      for (Action A : {Miss, Match})84        Scores[P][W][A] = {AwfulScore, Miss};85  PatTypeSet = calculateRoles(llvm::StringRef(Pat, PatN),86                              llvm::MutableArrayRef(PatRole, PatN));87}88 89std::optional<float> FuzzyMatcher::match(llvm::StringRef Word) {90  if (!(WordContainsPattern = init(Word)))91    return std::nullopt;92  if (!PatN)93    return 1;94  buildGraph();95  auto Best = std::max(Scores[PatN][WordN][Miss].Score,96                       Scores[PatN][WordN][Match].Score);97  if (isAwful(Best))98    return std::nullopt;99  float Score =100      ScoreScale * std::min(PerfectBonus * PatN, std::max<int>(0, Best));101  // If the pattern is as long as the word, we have an exact string match,102  // since every pattern character must match something.103  if (WordN == PatN)104    Score *= 2; // May not be perfect 2 if case differs in a significant way.105  return Score;106}107 108// We get CharTypes from a lookup table. Each is 2 bits, 4 fit in each byte.109// The top 6 bits of the char select the byte, the bottom 2 select the offset.110// e.g. 'q' = 010100 01 = byte 28 (55), bits 3-2 (01) -> Lower.111constexpr static uint8_t CharTypes[] = {112    0x00, 0x00, 0x00, 0x00, // Control characters113    0x00, 0x00, 0x00, 0x00, // Control characters114    0xff, 0xff, 0xff, 0xff, // Punctuation115    0x55, 0x55, 0xf5, 0xff, // Numbers->Lower, more Punctuation.116    0xab, 0xaa, 0xaa, 0xaa, // @ and A-O117    0xaa, 0xaa, 0xea, 0xff, // P-Z, more Punctuation.118    0x57, 0x55, 0x55, 0x55, // ` and a-o119    0x55, 0x55, 0xd5, 0x3f, // p-z, Punctuation, DEL.120    0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, // Bytes over 127 -> Lower.121    0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, // (probably UTF-8).122    0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55,123    0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55, 0x55,124};125 126// The Role can be determined from the Type of a character and its neighbors:127//128//   Example  | Chars | Type | Role129//   ---------+--------------+-----130//   F(o)oBar | Foo   | Ull  | Tail131//   Foo(B)ar | oBa   | lUl  | Head132//   (f)oo    | ^fo   | Ell  | Head133//   H(T)TP   | HTT   | UUU  | Tail134//135// Our lookup table maps a 6 bit key (Prev, Curr, Next) to a 2-bit Role.136// A byte packs 4 Roles. (Prev, Curr) selects a byte, Next selects the offset.137// e.g. Lower, Upper, Lower -> 01 10 01 -> byte 6 (aa), bits 3-2 (10) -> Head.138constexpr static uint8_t CharRoles[] = {139    // clang-format off140    //         Curr= Empty Lower Upper Separ141    /* Prev=Empty */ 0x00, 0xaa, 0xaa, 0xff, // At start, Lower|Upper->Head142    /* Prev=Lower */ 0x00, 0x55, 0xaa, 0xff, // In word, Upper->Head;Lower->Tail143    /* Prev=Upper */ 0x00, 0x55, 0x59, 0xff, // Ditto, but U(U)U->Tail144    /* Prev=Separ */ 0x00, 0xaa, 0xaa, 0xff, // After separator, like at start145    // clang-format on146};147 148template <typename T> static T packedLookup(const uint8_t *Data, int I) {149  return static_cast<T>((Data[I >> 2] >> ((I & 3) * 2)) & 3);150}151CharTypeSet calculateRoles(llvm::StringRef Text,152                           llvm::MutableArrayRef<CharRole> Roles) {153  assert(Text.size() == Roles.size());154  if (Text.size() == 0)155    return 0;156  CharType Type = packedLookup<CharType>(CharTypes, Text[0]);157  CharTypeSet TypeSet = 1 << Type;158  // Types holds a sliding window of (Prev, Curr, Next) types.159  // Initial value is (Empty, Empty, type of Text[0]).160  int Types = Type;161  // Rotate slides in the type of the next character.162  auto Rotate = [&](CharType T) { Types = ((Types << 2) | T) & 0x3f; };163  for (unsigned I = 0; I < Text.size() - 1; ++I) {164    // For each character, rotate in the next, and look up the role.165    Type = packedLookup<CharType>(CharTypes, Text[I + 1]);166    TypeSet |= 1 << Type;167    Rotate(Type);168    Roles[I] = packedLookup<CharRole>(CharRoles, Types);169  }170  // For the last character, the "next character" is Empty.171  Rotate(Empty);172  Roles[Text.size() - 1] = packedLookup<CharRole>(CharRoles, Types);173  return TypeSet;174}175 176// Sets up the data structures matching Word.177// Returns false if we can cheaply determine that no match is possible.178bool FuzzyMatcher::init(llvm::StringRef NewWord) {179  WordN = std::min<int>(MaxWord, NewWord.size());180  if (PatN > WordN)181    return false;182  std::copy(NewWord.begin(), NewWord.begin() + WordN, Word);183  if (PatN == 0)184    return true;185  for (int I = 0; I < WordN; ++I)186    LowWord[I] = lower(Word[I]);187 188  // Cheap subsequence check.189  for (int W = 0, P = 0; P != PatN; ++W) {190    if (W == WordN)191      return false;192    if (LowWord[W] == LowPat[P])193      ++P;194  }195 196  // FIXME: some words are hard to tokenize algorithmically.197  // e.g. vsprintf is V S Print F, and should match [pri] but not [int].198  // We could add a tokenization dictionary for common stdlib names.199  WordTypeSet = calculateRoles(llvm::StringRef(Word, WordN),200                               llvm::MutableArrayRef(WordRole, WordN));201  return true;202}203 204// The forwards pass finds the mappings of Pattern onto Word.205// Score = best score achieved matching Word[..W] against Pat[..P].206// Unlike other tables, indices range from 0 to N *inclusive*207// Matched = whether we chose to match Word[W] with Pat[P] or not.208//209// Points are mostly assigned to matched characters, with 1 being a good score210// and 3 being a great one. So we treat the score range as [0, 3 * PatN].211// This range is not strict: we can apply larger bonuses/penalties, or penalize212// non-matched characters.213void FuzzyMatcher::buildGraph() {214  for (int W = 0; W < WordN; ++W) {215    Scores[0][W + 1][Miss] = {Scores[0][W][Miss].Score - skipPenalty(W, Miss),216                              Miss};217    Scores[0][W + 1][Match] = {AwfulScore, Miss};218  }219  for (int P = 0; P < PatN; ++P) {220    for (int W = P; W < WordN; ++W) {221      auto &Score = Scores[P + 1][W + 1], &PreMiss = Scores[P + 1][W];222 223      auto MatchMissScore = PreMiss[Match].Score;224      auto MissMissScore = PreMiss[Miss].Score;225      if (P < PatN - 1) { // Skipping trailing characters is always free.226        MatchMissScore -= skipPenalty(W, Match);227        MissMissScore -= skipPenalty(W, Miss);228      }229      Score[Miss] = (MatchMissScore > MissMissScore)230                        ? ScoreInfo{MatchMissScore, Match}231                        : ScoreInfo{MissMissScore, Miss};232 233      auto &PreMatch = Scores[P][W];234      auto MatchMatchScore =235          allowMatch(P, W, Match)236              ? PreMatch[Match].Score + matchBonus(P, W, Match)237              : AwfulScore;238      auto MissMatchScore = allowMatch(P, W, Miss)239                                ? PreMatch[Miss].Score + matchBonus(P, W, Miss)240                                : AwfulScore;241      Score[Match] = (MatchMatchScore > MissMatchScore)242                         ? ScoreInfo{MatchMatchScore, Match}243                         : ScoreInfo{MissMatchScore, Miss};244    }245  }246}247 248bool FuzzyMatcher::allowMatch(int P, int W, Action Last) const {249  if (LowPat[P] != LowWord[W])250    return false;251  // We require a "strong" match:252  // - for the first pattern character.  [foo] !~ "barefoot"253  // - after a gap.                      [pat] !~ "patnther"254  if (Last == Miss) {255    // We're banning matches outright, so conservatively accept some other cases256    // where our segmentation might be wrong:257    //  - allow matching B in ABCDef (but not in NDEBUG)258    //  - we'd like to accept print in sprintf, but too many false positives259    if (WordRole[W] == Tail &&260        (Word[W] == LowWord[W] || !(WordTypeSet & 1 << Lower)))261      return false;262  }263  return true;264}265 266int FuzzyMatcher::skipPenalty(int W, Action Last) const {267  if (W == 0) // Skipping the first character.268    return 3;269  if (WordRole[W] == Head) // Skipping a segment.270    return 1; // We want to keep this lower than a consecutive match bonus.271  // Instead of penalizing non-consecutive matches, we give a bonus to a272  // consecutive match in matchBonus. This produces a better score distribution273  // than penalties in case of small patterns, e.g. 'up' for 'unique_ptr'.274  return 0;275}276 277int FuzzyMatcher::matchBonus(int P, int W, Action Last) const {278  assert(LowPat[P] == LowWord[W]);279  int S = 1;280  bool IsPatSingleCase =281      (PatTypeSet == 1 << Lower) || (PatTypeSet == 1 << Upper);282  // Bonus: case matches, or a Head in the pattern aligns with one in the word.283  // Single-case patterns lack segmentation signals and we assume any character284  // can be a head of a segment.285  if (Pat[P] == Word[W] ||286      (WordRole[W] == Head && (IsPatSingleCase || PatRole[P] == Head)))287    ++S;288  // Bonus: a consecutive match. First character match also gets a bonus to289  // ensure prefix final match score normalizes to 1.0.290  if (W == 0 || Last == Match)291    S += 2;292  // Penalty: matching inside a segment (and previous char wasn't matched).293  if (WordRole[W] == Tail && P && Last == Miss)294    S -= 3;295  // Penalty: a Head in the pattern matches in the middle of a word segment.296  if (PatRole[P] == Head && WordRole[W] == Tail)297    --S;298  // Penalty: matching the first pattern character in the middle of a segment.299  if (P == 0 && WordRole[W] == Tail)300    S -= 4;301  assert(S <= PerfectBonus);302  return S;303}304 305llvm::SmallString<256> FuzzyMatcher::dumpLast(llvm::raw_ostream &OS) const {306  llvm::SmallString<256> Result;307  OS << "=== Match \"" << llvm::StringRef(Word, WordN) << "\" against ["308     << llvm::StringRef(Pat, PatN) << "] ===\n";309  if (PatN == 0) {310    OS << "Pattern is empty: perfect match.\n";311    return Result = llvm::StringRef(Word, WordN);312  }313  if (WordN == 0) {314    OS << "Word is empty: no match.\n";315    return Result;316  }317  if (!WordContainsPattern) {318    OS << "Substring check failed.\n";319    return Result;320  }321  if (isAwful(std::max(Scores[PatN][WordN][Match].Score,322                       Scores[PatN][WordN][Miss].Score))) {323    OS << "Substring check passed, but all matches are forbidden\n";324  }325  if (!(PatTypeSet & 1 << Upper))326    OS << "Lowercase query, so scoring ignores case\n";327 328  // Traverse Matched table backwards to reconstruct the Pattern/Word mapping.329  // The Score table has cumulative scores, subtracting along this path gives330  // us the per-letter scores.331  Action Last =332      (Scores[PatN][WordN][Match].Score > Scores[PatN][WordN][Miss].Score)333          ? Match334          : Miss;335  int S[MaxWord];336  Action A[MaxWord];337  for (int W = WordN - 1, P = PatN - 1; W >= 0; --W) {338    A[W] = Last;339    const auto &Cell = Scores[P + 1][W + 1][Last];340    if (Last == Match)341      --P;342    const auto &Prev = Scores[P + 1][W][Cell.Prev];343    S[W] = Cell.Score - Prev.Score;344    Last = Cell.Prev;345  }346  for (int I = 0; I < WordN; ++I) {347    if (A[I] == Match && (I == 0 || A[I - 1] == Miss))348      Result.push_back('[');349    if (A[I] == Miss && I > 0 && A[I - 1] == Match)350      Result.push_back(']');351    Result.push_back(Word[I]);352  }353  if (A[WordN - 1] == Match)354    Result.push_back(']');355 356  for (char C : llvm::StringRef(Word, WordN))357    OS << " " << C << " ";358  OS << "\n";359  for (int I = 0, J = 0; I < WordN; I++)360    OS << " " << (A[I] == Match ? Pat[J++] : ' ') << " ";361  OS << "\n";362  for (int I = 0; I < WordN; I++)363    OS << llvm::format("%2d ", S[I]);364  OS << "\n";365 366  OS << "\nSegmentation:";367  OS << "\n'" << llvm::StringRef(Word, WordN) << "'\n ";368  for (int I = 0; I < WordN; ++I)369    OS << "?-+ "[static_cast<int>(WordRole[I])];370  OS << "\n[" << llvm::StringRef(Pat, PatN) << "]\n ";371  for (int I = 0; I < PatN; ++I)372    OS << "?-+ "[static_cast<int>(PatRole[I])];373  OS << "\n";374 375  OS << "\nScoring table (last-Miss, last-Match):\n";376  OS << " |    ";377  for (char C : llvm::StringRef(Word, WordN))378    OS << "  " << C << " ";379  OS << "\n";380  OS << "-+----" << std::string(WordN * 4, '-') << "\n";381  for (int I = 0; I <= PatN; ++I) {382    for (Action A : {Miss, Match}) {383      OS << ((I && A == Miss) ? Pat[I - 1] : ' ') << "|";384      for (int J = 0; J <= WordN; ++J) {385        if (!isAwful(Scores[I][J][A].Score))386          OS << llvm::format("%3d%c", Scores[I][J][A].Score,387                             Scores[I][J][A].Prev == Match ? '*' : ' ');388        else389          OS << "    ";390      }391      OS << "\n";392    }393  }394 395  return Result;396}397 398} // namespace clangd399} // namespace clang400