[analyzer][HTMLRewriter] Cache partial rewrite results.

clang/include/clang/Rewrite/Core/HTMLRewrite.h

@@ -24,6 +24,15 @@ class RewriteBuffer;
 class Preprocessor;
 
 namespace html {
+  struct RelexRewriteCache;
+  using RelexRewriteCacheRef = std::shared_ptr<RelexRewriteCache>;
+
+  /// If you need to rewrite the same file multiple times, you can instantiate
+  /// a RelexRewriteCache and refer functions such as SyntaxHighlight()
+  /// and HighlightMacros() to it so that to avoid re-lexing the file each time.
+  /// The cache may outlive the rewriter as long as cached FileIDs and source
+  /// locations continue to make sense for the translation unit as a whole.
+  RelexRewriteCacheRef instantiateRelexRewriteCache();
 
   /// HighlightRange - Highlight a range in the source code with the specified
   /// start/end tags.  B/E must be in the same file.  This ensures that
@@ -67,13 +76,15 @@ namespace html {
 
   /// SyntaxHighlight - Relex the specified FileID and annotate the HTML with
   /// information about keywords, comments, etc.
-  void SyntaxHighlight(Rewriter &R, FileID FID, const Preprocessor &PP);
+  void SyntaxHighlight(Rewriter &R, FileID FID, const Preprocessor &PP,
+                       RelexRewriteCacheRef Cache = nullptr);
 
   /// HighlightMacros - This uses the macro table state from the end of the
   /// file, to reexpand macros and insert (into the HTML) information about the
   /// macro expansions.  This won't be perfectly perfect, but it will be
   /// reasonably close.
-  void HighlightMacros(Rewriter &R, FileID FID, const Preprocessor &PP);
+  void HighlightMacros(Rewriter &R, FileID FID, const Preprocessor &PP,
+                       RelexRewriteCacheRef Cache = nullptr);
 
 } // end html namespace
 } // end clang namespace

clang/lib/Rewrite/HTMLRewrite.cpp

@@ -21,8 +21,10 @@
 #include "llvm/Support/MemoryBuffer.h"
 #include "llvm/Support/raw_ostream.h"
 #include <memory>
-using namespace clang;
 
+using namespace clang;
+using namespace llvm;
+using namespace html;
 
 /// HighlightRange - Highlight a range in the source code with the specified
 /// start/end tags.  B/E must be in the same file.  This ensures that
@@ -104,6 +106,32 @@ void html::HighlightRange(RewriteBuffer &RB, unsigned B, unsigned E,
   }
 }
 
+namespace clang::html {
+struct RelexRewriteCache {
+  // These structs mimic input arguments of HighlightRange().
+  struct Highlight {
+    SourceLocation B, E;
+    std::string StartTag, EndTag;
+    bool IsTokenRange;
+  };
+  struct RawHighlight {
+    unsigned B, E;
+    std::string StartTag, EndTag;
+  };
+
+  // SmallVector isn't appropriate because these vectors are almost never small.
+  using HighlightList = std::vector<Highlight>;
+  using RawHighlightList = std::vector<RawHighlight>;
+
+  DenseMap<FileID, RawHighlightList> SyntaxHighlights;
+  DenseMap<FileID, HighlightList> MacroHighlights;
+};
+} // namespace clang::html
+
+html::RelexRewriteCacheRef html::instantiateRelexRewriteCache() {
+  return std::make_shared<RelexRewriteCache>();
+}
+
 void html::EscapeText(Rewriter &R, FileID FID,
                       bool EscapeSpaces, bool ReplaceTabs) {
 
@@ -442,13 +470,18 @@ input.spoilerhider:checked + label + .spoiler{
 /// information about keywords, macro expansions etc.  This uses the macro
 /// table state from the end of the file, so it won't be perfectly perfect,
 /// but it will be reasonably close.
-void html::SyntaxHighlight(Rewriter &R, FileID FID, const Preprocessor &PP) {
-  RewriteBuffer &RB = R.getEditBuffer(FID);
+static void SyntaxHighlightImpl(
+    Rewriter &R, FileID FID, const Preprocessor &PP,
+    llvm::function_ref<void(RewriteBuffer &, unsigned, unsigned, const char *,
+                            const char *, const char *)>
+        HighlightRangeCallback) {
 
+  RewriteBuffer &RB = R.getEditBuffer(FID);
   const SourceManager &SM = PP.getSourceManager();
   llvm::MemoryBufferRef FromFile = SM.getBufferOrFake(FID);
+  const char *BufferStart = FromFile.getBuffer().data();
+
   Lexer L(FID, FromFile, SM, PP.getLangOpts());
-  const char *BufferStart = L.getBuffer().data();
 
   // Inform the preprocessor that we want to retain comments as tokens, so we
   // can highlight them.
@@ -475,13 +508,13 @@ void html::SyntaxHighlight(Rewriter &R, FileID FID, const Preprocessor &PP) {
 
       // If this is a pp-identifier, for a keyword, highlight it as such.
       if (Tok.isNot(tok::identifier))
-        HighlightRange(RB, TokOffs, TokOffs+TokLen, BufferStart,
-                       "<span class='keyword'>", "</span>");
+        HighlightRangeCallback(RB, TokOffs, TokOffs + TokLen, BufferStart,
+                               "<span class='keyword'>", "</span>");
       break;
     }
     case tok::comment:
-      HighlightRange(RB, TokOffs, TokOffs+TokLen, BufferStart,
-                     "<span class='comment'>", "</span>");
+      HighlightRangeCallback(RB, TokOffs, TokOffs + TokLen, BufferStart,
+                             "<span class='comment'>", "</span>");
       break;
     case tok::utf8_string_literal:
       // Chop off the u part of u8 prefix
@@ -498,8 +531,8 @@ void html::SyntaxHighlight(Rewriter &R, FileID FID, const Preprocessor &PP) {
       [[fallthrough]];
     case tok::string_literal:
       // FIXME: Exclude the optional ud-suffix from the highlighted range.
-      HighlightRange(RB, TokOffs, TokOffs+TokLen, BufferStart,
-                     "<span class='string_literal'>", "</span>");
+      HighlightRangeCallback(RB, TokOffs, TokOffs + TokLen, BufferStart,
+                             "<span class='string_literal'>", "</span>");
       break;
     case tok::hash: {
       // If this is a preprocessor directive, all tokens to end of line are too.
@@ -516,8 +549,8 @@ void html::SyntaxHighlight(Rewriter &R, FileID FID, const Preprocessor &PP) {
       }
 
       // Find end of line.  This is a hack.
-      HighlightRange(RB, TokOffs, TokEnd, BufferStart,
-                     "<span class='directive'>", "</span>");
+      HighlightRangeCallback(RB, TokOffs, TokEnd, BufferStart,
+                             "<span class='directive'>", "</span>");
 
       // Don't skip the next token.
       continue;
@@ -527,12 +560,43 @@ void html::SyntaxHighlight(Rewriter &R, FileID FID, const Preprocessor &PP) {
     L.LexFromRawLexer(Tok);
   }
 }
+void html::SyntaxHighlight(Rewriter &R, FileID FID, const Preprocessor &PP,
+                           RelexRewriteCacheRef Cache) {
+  RewriteBuffer &RB = R.getEditBuffer(FID);
+  const SourceManager &SM = PP.getSourceManager();
+  llvm::MemoryBufferRef FromFile = SM.getBufferOrFake(FID);
+  const char *BufferStart = FromFile.getBuffer().data();
+
+  if (Cache) {
+    auto CacheIt = Cache->SyntaxHighlights.find(FID);
+    if (CacheIt != Cache->SyntaxHighlights.end()) {
+      for (const RelexRewriteCache::RawHighlight &H : CacheIt->second) {
+        HighlightRange(RB, H.B, H.E, BufferStart, H.StartTag.data(),
+                       H.EndTag.data());
+      }
+      return;
+    }
+  }
+
+  // "Every time you would call HighlightRange, cache the inputs as well."
+  auto HighlightRangeCallback = [&](RewriteBuffer &RB, unsigned B, unsigned E,
+                                    const char *BufferStart,
+                                    const char *StartTag, const char *EndTag) {
+    HighlightRange(RB, B, E, BufferStart, StartTag, EndTag);
+
+    if (Cache)
+      Cache->SyntaxHighlights[FID].push_back({B, E, StartTag, EndTag});
+  };
+
+  SyntaxHighlightImpl(R, FID, PP, HighlightRangeCallback);
+}
+
+static void HighlightMacrosImpl(
+    Rewriter &R, FileID FID, const Preprocessor &PP,
+    llvm::function_ref<void(Rewriter &, SourceLocation, SourceLocation,
+                            const char *, const char *, bool)>
+        HighlightRangeCallback) {
 
-/// HighlightMacros - This uses the macro table state from the end of the
-/// file, to re-expand macros and insert (into the HTML) information about the
-/// macro expansions.  This won't be perfectly perfect, but it will be
-/// reasonably close.
-void html::HighlightMacros(Rewriter &R, FileID FID, const Preprocessor& PP) {
   // Re-lex the raw token stream into a token buffer.
   const SourceManager &SM = PP.getSourceManager();
   std::vector<Token> TokenStream;
@@ -659,11 +723,44 @@ void html::HighlightMacros(Rewriter &R, FileID FID, const Preprocessor& PP) {
     // get highlighted.
     Expansion = "<span class='macro_popup'>" + Expansion + "</span></span>";
 
-    HighlightRange(R, LLoc.getBegin(), LLoc.getEnd(), "<span class='macro'>",
-                   Expansion.c_str(), LLoc.isTokenRange());
+    HighlightRangeCallback(R, LLoc.getBegin(), LLoc.getEnd(),
+                           "<span class='macro'>", Expansion.c_str(),
+                           LLoc.isTokenRange());
   }
 
   // Restore the preprocessor's old state.
   TmpPP.setDiagnostics(*OldDiags);
   TmpPP.setPragmasEnabled(PragmasPreviouslyEnabled);
 }
+
+/// HighlightMacros - This uses the macro table state from the end of the
+/// file, to re-expand macros and insert (into the HTML) information about the
+/// macro expansions.  This won't be perfectly perfect, but it will be
+/// reasonably close.
+void html::HighlightMacros(Rewriter &R, FileID FID, const Preprocessor &PP,
+                           RelexRewriteCacheRef Cache) {
+  if (Cache) {
+    auto CacheIt = Cache->MacroHighlights.find(FID);
+    if (CacheIt != Cache->MacroHighlights.end()) {
+      for (const RelexRewriteCache::Highlight &H : CacheIt->second) {
+        HighlightRange(R, H.B, H.E, H.StartTag.data(), H.EndTag.data(),
+                       H.IsTokenRange);
+      }
+      return;
+    }
+  }
+
+  // "Every time you would call HighlightRange, cache the inputs as well."
+  auto HighlightRangeCallback = [&](Rewriter &R, SourceLocation B,
+                                    SourceLocation E, const char *StartTag,
+                                    const char *EndTag, bool isTokenRange) {
+    HighlightRange(R, B, E, StartTag, EndTag, isTokenRange);
+
+    if (Cache) {
+      Cache->MacroHighlights[FID].push_back(
+          {B, E, StartTag, EndTag, isTokenRange});
+    }
+  };
+
+  HighlightMacrosImpl(R, FID, PP, HighlightRangeCallback);
+}

clang/lib/StaticAnalyzer/Core/HTMLDiagnostics.cpp

@@ -69,6 +69,8 @@ class HTMLDiagnostics : public PathDiagnosticConsumer {
   const Preprocessor &PP;
   const bool SupportsCrossFileDiagnostics;
   llvm::StringSet<> EmittedHashes;
+  html::RelexRewriteCacheRef RewriterCache =
+      html::instantiateRelexRewriteCache();
 
 public:
   HTMLDiagnostics(PathDiagnosticConsumerOptions DiagOpts,
@@ -309,10 +311,6 @@ void HTMLDiagnostics::ReportDiag(const PathDiagnostic& D,
     return;
   }
 
-  // FIXME: This causes each file to be re-parsed and syntax-highlighted
-  // and macro-expanded separately for each report. We could cache such rewrites
-  // across all reports and only re-do the part that's actually different:
-  // the warning/note bubbles.
   std::string report = GenerateHTML(D, R, SMgr, path, declName.c_str());
   if (report.empty()) {
     llvm::errs() << "warning: no diagnostics generated for main file.\n";
@@ -882,8 +880,8 @@ void HTMLDiagnostics::RewriteFile(Rewriter &R, const PathPieces &path,
   // If we have a preprocessor, relex the file and syntax highlight.
   // We might not have a preprocessor if we come from a deserialized AST file,
   // for example.
-  html::SyntaxHighlight(R, FID, PP);
-  html::HighlightMacros(R, FID, PP);
+  html::SyntaxHighlight(R, FID, PP, RewriterCache);
+  html::HighlightMacros(R, FID, PP, RewriterCache);
 }
 
 void HTMLDiagnostics::HandlePiece(Rewriter &R, FileID BugFileID,

clang/test/Analysis/html_diagnostics/counter.c

@@ -0,0 +1,29 @@
+// RUN: rm -fR %t
+// RUN: mkdir %t
+// RUN: %clang_analyze_cc1 -analyzer-checker=core \
+// RUN:                    -analyzer-output=html -o %t -verify %s
+// RUN: grep -v CHECK %t/report-*.html | FileCheck %s
+
+
+void foo() {
+  int *x = 0;
+  *x = __COUNTER__; // expected-warning{{Dereference of null pointer (loaded from variable 'x')}}
+}
+
+void bar() {
+  int *y;
+  *y = __COUNTER__; // expected-warning{{Dereference of undefined pointer value (loaded from variable 'y')}}
+}
+
+// The checks below confirm that both reports have the same values for __COUNTER__.
+//
+// FIXME: The correct values are (0, 1, 0, 1). Because we re-lex the file in order
+// to detect macro expansions for HTML report purposes, they turn into (2, 3, 2, 3)
+// by the time we emit HTML. But at least it's better than (2, 3, 4, 5)
+// which would have been the case if we re-lexed the file *each* time we
+// emitted an HTML report.
+
+// CHECK: <span class='macro'>__COUNTER__<span class='macro_popup'>2</span></span>
+// CHECK: <span class='macro'>__COUNTER__<span class='macro_popup'>3</span></span>
+// CHECK: <span class='macro'>__COUNTER__<span class='macro_popup'>2</span></span>
+// CHECK: <span class='macro'>__COUNTER__<span class='macro_popup'>3</span></span>