[workspace/symbol] Sort candidates with a fuzzy matching algorithm (#182)

src/config.h

@@ -24,7 +24,7 @@ struct Config {
   bool logSkippedPathsForIndex = false;
 
   // Maximum workspace search results.
-  int maxWorkspaceSearchResults = 1000;
+  int maxWorkspaceSearchResults = 200;
 
   // Force a certain number of indexer threads. If less than 1 a default value
   // should be used.

src/messages/workspace_symbol.cc

@@ -1,3 +1,8 @@
+#include <ctype.h>
+#include <limits.h>
+#include <algorithm>
+#include <functional>
+
 #include "lex_utils.h"
 #include "message_handler.h"
 #include "query_utils.h"
@@ -7,29 +12,30 @@
 namespace {
 
 // Lookup |symbol| in |db| and insert the value into |result|.
-void InsertSymbolIntoResult(QueryDatabase* db,
+bool InsertSymbolIntoResult(QueryDatabase* db,
                             WorkingFiles* working_files,
                             SymbolIdx symbol,
                             std::vector<lsSymbolInformation>* result) {
   optional<lsSymbolInformation> info =
       GetSymbolInfo(db, working_files, symbol, false /*use_short_name*/);
   if (!info)
-    return;
+    return false;
 
   optional<QueryLocation> location = GetDefinitionExtentOfSymbol(db, symbol);
   if (!location) {
     auto decls = GetDeclarationsOfSymbolForGotoDefinition(db, symbol);
     if (decls.empty())
-      return;
+      return false;
     location = decls[0];
   }
 
   optional<lsLocation> ls_location =
       GetLsLocation(db, working_files, *location);
   if (!ls_location)
-    return;
+    return false;
   info->location = *ls_location;
   result->push_back(*info);
+  return true;
 }
 
 struct lsWorkspaceSymbolParams {
@@ -51,12 +57,119 @@ struct Out_WorkspaceSymbol : public lsOutMessage<Out_WorkspaceSymbol> {
 };
 MAKE_REFLECT_STRUCT(Out_WorkspaceSymbol, jsonrpc, id, result);
 
+
+///// Fuzzy matching
+
+// Negative but far from INT_MIN so that intermediate results are hard to
+// overflow
+constexpr int kMinScore = INT_MIN / 2;
+// Penalty of dropping a leading character in str
+constexpr int kLeadingGapScore = -4;
+// Penalty of dropping a non-leading character in str
+constexpr int kGapScore = -5;
+// Bonus of aligning with an initial character of a word in pattern. Must be
+// greater than 1
+constexpr int kPatternStartMultiplier = 2;
+
+constexpr int kWordStartScore = 100;
+constexpr int kNonWordScore = 90;
+
+// Less than kWordStartScore
+constexpr int kConsecutiveScore = kWordStartScore + kGapScore;
+// Slightly less than kConsecutiveScore
+constexpr int kCamelScore = kWordStartScore + kGapScore - 1;
+
+enum class CharClass { Lower, Upper, Digit, NonWord };
+
+static enum CharClass getCharClass(int c) {
+  if (islower(c)) return CharClass::Lower;
+  if (isupper(c)) return CharClass::Upper;
+  if (isdigit(c)) return CharClass::Digit;
+  return CharClass::NonWord;
+}
+
+static int getScoreFor(CharClass prev, CharClass curr) {
+  if (prev == CharClass::NonWord && curr != CharClass::NonWord)
+    return kWordStartScore;
+  if ((prev == CharClass::Lower && curr == CharClass::Upper) ||
+      (prev != CharClass::Digit && curr == CharClass::Digit))
+    return kCamelScore;
+  if (curr == CharClass::NonWord)
+    return kNonWordScore;
+  return 0;
+}
+
+/*
+fuzzyEvaluate implements a global sequence alignment algorithm to find the maximum accumulated score by aligning `pattern` to `str`. It applies when `pattern` is a subsequence of `str`.
+
+Scoring criteria
+- Prefer matches at the start of a word, or the start of subwords in CamelCase/camelCase/camel123 words. See kWordStartScore/kCamelScore
+- Non-word characters matter. See kNonWordScore
+- The first characters of words of `pattern` receive bonus because they usually have more significance than the rest. See kPatternStartMultiplier
+- Superfluous characters in `str` will reduce the score (gap penalty). See kGapScore
+- Prefer early occurrence of the first character. See kLeadingGapScore/kGapScore
+
+The recurrence of the dynamic programming:
+dp[i][j]: maximum accumulated score by aligning pattern[0..i] to str[0..j]
+dp[0][j] = leading_gap_penalty(0, j) + score[j]
+dp[i][j] = max(dp[i-1][j-1] + CONSECUTIVE_SCORE, max(dp[i-1][k] + gap_penalty(k+1, j) + score[j] : k < j))
+The first dimension can be suppressed since we do not need a matching scheme, which reduces the space complexity from O(N*M) to O(M)
+*/
+int fuzzyEvaluate(const std::string& pattern,
+                  const std::string& str,
+                  std::vector<int>& score,
+                  std::vector<int>& dp) {
+  bool pfirst = true,  // aligning the first character of pattern
+      pstart = true;   // whether we are aligning the start of a word in pattern
+  int uleft = 0,       // value of the upper left cell
+      ulefts = 0,      // maximum value of uleft and cells on the left
+      left, lefts;     // similar to uleft/ulefts, but for the next row
+
+  // Calculate position score for each character in str.
+  CharClass prev = CharClass::NonWord;
+  for (int i = 0; i < int(str.size()); i++) {
+    CharClass cur = getCharClass(str[i]);
+    score[i] = getScoreFor(prev, cur);
+    prev = cur;
+  }
+  std::fill_n(dp.begin(), str.size(), kMinScore);
+
+  // Align each character of pattern.
+  for (unsigned char pc: pattern) {
+    if (isspace(pc)) {
+      pstart = true;
+      continue;
+    }
+    lefts = kMinScore;
+    // Enumerate the character in str to be aligned with pc.
+    for (int i = 0; i < int(str.size()); i++) {
+      left = dp[i];
+      lefts = std::max(lefts + kGapScore, left);
+      if (tolower(pc) == tolower(str[i])) {
+        int t = score[i] * (pstart ? kPatternStartMultiplier : 1);
+        dp[i] = pfirst ? kLeadingGapScore * i + t
+                       : std::max(uleft + kConsecutiveScore, ulefts + t);
+      } else
+        dp[i] = kMinScore;
+      uleft = left;
+      ulefts = lefts;
+    }
+    pfirst = pstart = false;
+  }
+
+  // Enumerate the end position of the match in str.
+  lefts = kMinScore;
+  for (int i = 0; i < int(str.size()); i++)
+    // For function types, db->detailed_names may have trailing characters for
+    // parameters. We do not want to penalize them.
+    // If we use `short_name` instead of `detailed_name` for fuzzy matching, the
+    // penulty kGapScore can be used.
+    lefts = std::max(lefts /*+ kGapScore */, dp[i]);
+  return lefts;
+}
+
 struct WorkspaceSymbolHandler : BaseMessageHandler<Ipc_WorkspaceSymbol> {
   void Run(Ipc_WorkspaceSymbol* request) override {
-    // TODO: implement fuzzy search, see
-    // https://github.com/junegunn/fzf/blob/master/src/matcher.go for
-    // inspiration
-
     Out_WorkspaceSymbol out;
     out.id = request->id;
 
@@ -66,7 +179,10 @@ struct WorkspaceSymbolHandler : BaseMessageHandler<Ipc_WorkspaceSymbol> {
     std::string query = request->params.query;
 
     std::unordered_set<std::string> inserted_results;
+    // db->detailed_names indices of each lsSymbolInformation in out.result
+    std::vector<int> result_indices;
     inserted_results.reserve(config->maxWorkspaceSearchResults);
+    result_indices.reserve(config->maxWorkspaceSearchResults);
 
     for (int i = 0; i < db->detailed_names.size(); ++i) {
       if (db->detailed_names[i].find(query) != std::string::npos) {
@@ -74,26 +190,56 @@ struct WorkspaceSymbolHandler : BaseMessageHandler<Ipc_WorkspaceSymbol> {
         if (!inserted_results.insert(db->detailed_names[i]).second)
           continue;
 
-        InsertSymbolIntoResult(db, working_files, db->symbols[i], &out.result);
+        if (InsertSymbolIntoResult(db, working_files, db->symbols[i], &out.result)) {
+          result_indices.push_back(i);
           if (out.result.size() >= config->maxWorkspaceSearchResults)
             break;
         }
       }
+    }
 
     if (out.result.size() < config->maxWorkspaceSearchResults) {
+      std::string query_without_space;
+      query_without_space.reserve(query.size());
+      for (char c: query)
+        if (!isspace(c))
+          query_without_space += c;
+
       for (int i = 0; i < db->detailed_names.size(); ++i) {
-        if (SubstringMatch(query, db->detailed_names[i])) {
+        if (SubstringMatch(query_without_space, db->detailed_names[i])) {
           // Do not show the same entry twice.
           if (!inserted_results.insert(db->detailed_names[i]).second)
             continue;
 
-          InsertSymbolIntoResult(db, working_files, db->symbols[i],
-                                 &out.result);
+          if (InsertSymbolIntoResult(db, working_files, db->symbols[i], &out.result)) {
+            result_indices.push_back(i);
             if (out.result.size() >= config->maxWorkspaceSearchResults)
               break;
           }
         }
       }
+    }
+
+    if (out.result.size() < config->maxWorkspaceSearchResults) {
+      int longest = 0;
+      for (int i: result_indices)
+        longest = std::max(longest, int(db->detailed_names[i].size()));
+
+      std::vector<int> score(longest), // score for each position
+          dp(longest); // dp[i]: maximum value by aligning pattern[0..pi] to str[0..si]
+      std::vector<std::pair<int, int>> permutation(result_indices.size());
+      for (int i = 0; i < int(result_indices.size()); i++) {
+        permutation[i] = {
+            fuzzyEvaluate(query, db->detailed_names[result_indices[i]], score,
+                          dp),
+            i};
+      }
+      std::sort(permutation.begin(), permutation.end(),
+                std::greater<std::pair<int, int>>());
+      for (int i = 0; i < int(result_indices.size()); i++)
+        if (i != permutation[i].second)
+          std::swap(out.result[i], out.result[permutation[i].second]);
+    }
 
     LOG_S(INFO) << "[querydb] Found " << out.result.size()
                 << " results for query " << query;