[pseudo] Add crude heuristics to choose taken preprocessor branches.

[sammccall]

Summary

In files where different preprocessing paths are possible, our goal is to
choose a preprocessed token sequence which we can parse that pins down as much
of the grammatical structure as possible.
This forms the "primary parse", and the not-taken branches get parsed later,
and are constrained to be compatible with the primary parse.

Concretely:

int x =
  #ifdef // TAKEN
    2 + 2 + 2 // determined during primary parse to be an expression
  #else
    2 // constrained to be an expression during a secondary parse
  #endif
  ;

[hokein]

clang/include/clang/Tooling/Syntax/Pseudo/DirectiveMap.h

123

off-topic: a random idea on the name of DirectiveMap came up to my mind when reading the code, how about DirectiveTree, the comment of it says the structure is a tree, and the BranchChooser is actually a tree-visiting pattern.

clang/include/clang/Tooling/Syntax/Pseudo/Token.h

84

nit: the name doesn't match the LLVM code-style.

clang/lib/Tooling/Syntax/Pseudo/DirectiveMap.cpp

210

We evaluate each conditional directive *independently*, I wonder whether it is important to use any low-hanging tricks to ensure consistent choices are made among different conditional directives. (my gut feeling is that it is nice to have, but we should not worry too much about it at the moment).

For example,

#ifdef __cplusplus
extern "C" {
#endif
....
#ifdef __cplusplus
}
#endif

If we enable the first #if, and the second one should be enabled as well. (this example is a trivial and common pattern, it has been handled by the existing code).

218

nit: add a comment saying higher is better.

225

Maybe (-Errors, Tokens, Directives) > (-Other.Errors, ...)

295

nit: !C.Taken => !C.Token.hasValue(), which is clearer.

305

nit: instead of using trivially, I think confidently fits better what the method does.

[sammccall]

clang/lib/Tooling/Syntax/Pseudo/DirectiveMap.cpp

210

Added a comment describing this case, I don't think we should tackle this now.

305

You're right this is a bad name, but also because returning false doesn't mean it's not trivially/confidently taken, but rather that it's trivially/confidently *not* taken.
I think removing the adverb actually makes this clearer.

Also this case is confusing

#if foo
#else // is this "always taken"?
#bar

Renamed to isTakenWhenReached, wdyt?

[sammccall]

Rebase and address comments

[hokein]

Thanks, looks good!

clang-tools-extra/pseudo/include/clang-pseudo/DirectiveMap.h

95 ↗

(On Diff #419142)

it is done in this patch, can be removed now.

clang-tools-extra/pseudo/lib/DirectiveMap.cpp

329 ↗

(On Diff #419142)

I thought the first token of the Directive should always be tok::hash, isn't it?

[sammccall]

clang-tools-extra/pseudo/lib/DirectiveMap.cpp

329 ↗

(On Diff #419142)

Yeah, this was defensive against parsing recovery, but we're sufficiently close to the parser that we don't need to be.

clang/include/clang/Tooling/Syntax/Pseudo/DirectiveMap.h

struct DirectiveMap {
   struct Conditional {
     /// The sequence of directives that introduce top-level alternative parses.
     ///
     /// The first branch will have an #if type directive.
     /// Subsequent branches will have #else type directives.
     std::vector<std::pair<Directive, DirectiveMap>> Branches;
     /// The directive terminating the conditional, should be #endif.
     Directive End;
+    /// The index of the conditional branch we chose as active.
+    /// None indicates no branch was taken (e.g. #if 0 ... #endif).
+    /// The initial map from of `parse()` has no branches marked as taken.
+    /// See `chooseConditionalBranches()`.
+    llvm::Optional<unsigned> Taken;
   };
 
   /// Some piece of the file. {One of Code, Directive, Conditional}.
   class Chunk; // Defined below.
   std::vector<Chunk> Chunks;
 
   /// Extract preprocessor structure by examining the raw tokens.
   static DirectiveMap parse(const TokenStream &);
 
   // FIXME: add heuristically selection of conditional branches.
   // FIXME: allow deriving a preprocessed stream
 };
 llvm::raw_ostream &operator<<(llvm::raw_ostream &, const DirectiveMap &);
 llvm::raw_ostream &operator<<(llvm::raw_ostream &, const DirectiveMap::Chunk &);
 llvm::raw_ostream &operator<<(llvm::raw_ostream &, const DirectiveMap::Code &);
 llvm::raw_ostream &operator<<(llvm::raw_ostream &,
                               const DirectiveMap::Directive &);
 llvm::raw_ostream &operator<<(llvm::raw_ostream &,
                               const DirectiveMap::Conditional &);
 
+/// Selects a "taken" branch for each conditional directive in the file.
+///
+/// The choice is somewhat arbitrary, but aims to produce a useful parse:
+///  - idioms like `#if 0` are respected
+///  - we avoid paths that reach `#error`
+///  - we try to maximize the amount of code seen
+/// The choice may also be "no branch taken".
+///
+/// Choices are also made for conditionals themselves inside not-taken branches:
+///   #if 1 // taken!
+///   #else // not taken
+///      #if 1 // taken!
+///      #endif
+///   #endif
+///
+/// The choices are stored in Conditional::Taken nodes.
+void chooseConditionalBranches(DirectiveMap &, const TokenStream &Code);
+
 // FIXME: This approximates std::variant<Code, Directive, Conditional>.
 //         Switch once we can use C++17.
 class DirectiveMap::Chunk {
 public:
   enum Kind { K_Empty, K_Code, K_Directive, K_Conditional };
   Kind kind() const {
     return CodeVariant          ? K_Code
            : DirectiveVariant   ? K_Directive

clang/include/clang/Tooling/Syntax/Pseudo/Token.h

struct Token {
   // Helpers to get/set Flags based on `enum class`.
   template <class T> bool flag(T Mask) const {
     return Flags & uint8_t{static_cast<std::underlying_type_t<T>>(Mask)};
   }
   template <class T> void setFlag(T Mask) {
     Flags |= uint8_t{static_cast<std::underlying_type_t<T>>(Mask)};
   }
 
+  /// Returns the next token in the stream. this may not be a sentinel.
+  const Token &next() const {
+    assert(Kind != tok::eof);
+    return *(this + 1);
+  }
+  /// Returns the next token in the stream, skipping over comments.
+  const Token &next_nc() const {
+    const Token *T = this;
+    do
+      T = &T->next();
+    while (T->Kind == tok::comment);
+    return *T;
+  }
+
   /// The type of token as determined by clang's lexer.
   clang::tok::TokenKind Kind = clang::tok::unknown;
 };
 static_assert(sizeof(Token) <= sizeof(char *) + 16, "Careful with layout!");
 llvm::raw_ostream &operator<<(llvm::raw_ostream &, const Token &);
 
 /// A half-open range of tokens within a stream.
 struct Token::Range {

clang/lib/Tooling/Syntax/Pseudo/DirectiveMap.cpp

 
 DirectiveMap DirectiveMap::parse(const TokenStream &Code) {
   DirectiveMap Result;
   DirectiveParser(Code).parse(&Result);
   return Result;
 }
 
 static void dump(llvm::raw_ostream &OS, const DirectiveMap &, unsigned Indent);
-static void dump(llvm::raw_ostream &OS, const DirectiveMap::Directive &Directive,
-                 unsigned Indent) {
-  OS.indent(Indent) << llvm::formatv("#{0} ({1} tokens)\n",
-                                     tok::getPPKeywordSpelling(Directive.Kind),
-                                     Directive.Tokens.size());
+static void dump(llvm::raw_ostream &OS,
+                 const DirectiveMap::Directive &Directive, unsigned Indent,
+                 bool Taken = false) {
+  OS.indent(Indent) << llvm::formatv(
+      "#{0} ({1} tokens){2}\n", tok::getPPKeywordSpelling(Directive.Kind),
+      Directive.Tokens.size(), Taken ? " TAKEN" : "");
 }
 static void dump(llvm::raw_ostream &OS, const DirectiveMap::Code &Code,
                  unsigned Indent) {
   OS.indent(Indent) << llvm::formatv("code ({0} tokens)\n", Code.Tokens.size());
 }
 static void dump(llvm::raw_ostream &OS,
                  const DirectiveMap::Conditional &Conditional,
                  unsigned Indent) {
-  for (const auto &Branch : Conditional.Branches) {
-    dump(OS, Branch.first, Indent);
+  for (unsigned I = 0; I < Conditional.Branches.size(); ++I) {
+    const auto &Branch = Conditional.Branches[I];
+    dump(OS, Branch.first, Indent, Conditional.Taken == I);
     dump(OS, Branch.second, Indent + 2);
   }
   dump(OS, Conditional.End, Indent);
 }
 
 static void dump(llvm::raw_ostream &OS, const DirectiveMap::Chunk &Chunk,
                  unsigned Indent) {
   switch (Chunk.kind()) {
#define OSTREAM_DUMP(Type)                                                     \
   }
 OSTREAM_DUMP(DirectiveMap)
 OSTREAM_DUMP(DirectiveMap::Chunk)
 OSTREAM_DUMP(DirectiveMap::Directive)
 OSTREAM_DUMP(DirectiveMap::Conditional)
 OSTREAM_DUMP(DirectiveMap::Code)
 #undef OSTREAM_DUMP
 
+namespace {
+class BranchChooser {
+public:
+  BranchChooser(const TokenStream &Code) : Code(Code) {}
+
+  void choose(DirectiveMap &M) { walk(M); }
+
+private:
+  // Describes the code seen by making particular branch choices.
+  struct Score {
+    unsigned Tokens = 0; // excluding comments and directives
+    unsigned Directives = 0;
+    unsigned Errors = 0; // #error directives
+
+    bool operator>(const Score &Other) const {
+      // Seeing errors is bad, other things are good.
+      return std::tie(Other.Errors, Tokens, Directives) >
+             std::tie(Errors, Other.Tokens, Other.Directives);
+    }
+
+    Score &operator+=(const Score &Other) {
+      Tokens += Other.Tokens;
+      Directives += Other.Directives;
+      Errors += Other.Errors;
+      return *this;
+    }
+  };
+
+  Score walk(DirectiveMap::Code &C) {
+    Score S;
+    for (const Token &T : Code.tokens(C.Tokens))
+      if (T.Kind != tok::comment)
+        ++S.Tokens;
+    return S;
+  }
+
+  Score walk(DirectiveMap::Directive &D) {
+    Score S;
+    S.Directives = 1;
+    S.Errors = D.Kind == tok::pp_error;
+    return S;
+  }
+
+  Score walk(DirectiveMap::Chunk &C) {
+    switch (C.kind()) {
+    case DirectiveMap::Chunk::K_Code:
+      return walk((DirectiveMap::Code &)C);
+    case DirectiveMap::Chunk::K_Directive:
+      return walk((DirectiveMap::Directive &)C);
+    case DirectiveMap::Chunk::K_Conditional:
+      return walk((DirectiveMap::Conditional &)C);
+    case DirectiveMap::Chunk::K_Empty:
+      break;
+    }
+    llvm_unreachable("bad chunk kind");
+  }
+
+  Score walk(DirectiveMap &M) {
+    Score S;
+    for (DirectiveMap::Chunk &C : M.Chunks)
+      S += walk(C);
+    return S;
+  }
+
+  Score walk(DirectiveMap::Conditional &C) {
+    Score Best;
+    bool MayTakeTrivial = true;
+    bool TookTrivial = false;
+
+    for (unsigned I = 0; I < C.Branches.size(); ++I) {
+      // Walk the branch to make its nested choices in any case.
+      Score BranchScore = walk(C.Branches[I].second);
+      // If we already took an #if 1, don't consider any other branches.
+      if (TookTrivial)
+        continue;
+      // Is this a trivial #if 0 or #if 1?
+      if (auto TriviallyTaken = isTriviallyTaken(C.Branches[I].first)) {
+        if (!*TriviallyTaken)
+          continue; // Don't consider #if 0 even if it scores well.
+        if (MayTakeTrivial)
+          TookTrivial = true;
+      } else {
+        // After a nontrivial condition, #elif 1 isn't guaranteed taken.
+        MayTakeTrivial = false;
+      }
+      // Is this the best branch so far? (Including if it's #if 1).
+      if (TookTrivial || !C.Taken || BranchScore > Best) {
+        Best = BranchScore;
+        C.Taken = I;
+      }
+    }
+    return Best;
+  }
+
+  // Return true if the directive starts an always-taken conditional branch,
+  // false if the branch is never taken, and None otherwise.
+  llvm::Optional<bool> isTriviallyTaken(const DirectiveMap::Directive &Dir) {
+    switch (Dir.Kind) {
+    case clang::tok::pp_if:
+    case clang::tok::pp_elif:
+      break; // handled below
+    case clang::tok::pp_else:
+      return true;
+    default: // #ifdef etc
+      return llvm::None;
+    }
+
+    const auto &Tokens = Code.tokens(Dir.Tokens);
+    if (Tokens.empty() || Tokens.front().Kind != tok::hash)
+      return llvm::None;
+    const Token &Name = Tokens.front().next_nc();
+    const Token &Value = Name.next_nc();
+    // Does the condition consist of exactly one token?
+    if (&Value >= Tokens.end() || &Value.next_nc() < Tokens.end())
+      return llvm::None;
+    return llvm::StringSwitch<llvm::Optional<bool>>(Value.text())
+        .Cases("true", "1", true)
+        .Cases("false", "0", false)
+        .Default(llvm::None);
+  }
+
+  const TokenStream &Code;
+};
+
+} // namespace
+
+void chooseConditionalBranches(DirectiveMap &Map, const TokenStream &Code) {
+  BranchChooser{Code}.choose(Map);
+}
+
 } // namespace pseudo
 } // namespace syntax
 } // namespace clang

clang/test/Syntax/lex.c

 TOKEN-NEXT: numeric_constant 4:2 "0"
 TOKEN-NEXT: semi             4:2 ";"
 TOKEN-NEXT: hash             5:0 "#" flags=1
 TOKEN-NEXT: raw_identifier   5:0 "endif"
 TOKEN-NEXT: r_brace          6:0 "}" flags=1
 
 RUN: clang-pseudo -source %s -print-directive-map | FileCheck %s -check-prefix=PPS --strict-whitespace
      PPS: code (5 tokens)
-PPS-NEXT: #ifndef (3 tokens)
+PPS-NEXT: #ifndef (3 tokens) TAKEN
 PPS-NEXT:   code (4 tokens)
 PPS-NEXT: #else (2 tokens)
 PPS-NEXT:   code (3 tokens)
 PPS-NEXT: #endif (2 tokens)
 PPS-NEXT: code (2 tokens)
                 ^ including this block comment
 
 *******************************************************************************/

clang/tools/clang-pseudo/ClangPseudo.cpp

if (Grammar.getNumOccurrences()) {
     return 0;
   }
 
   if (Source.getNumOccurrences()) {
     std::string Text = readOrDie(Source);
     clang::LangOptions LangOpts; // FIXME: use real options.
     auto Stream = clang::syntax::pseudo::lex(Text, LangOpts);
     auto Structure = clang::syntax::pseudo::DirectiveMap::parse(Stream);
+    clang::syntax::pseudo::chooseConditionalBranches(Structure, Stream);
 
     if (PrintDirectiveMap)
       llvm::outs() << Structure;
     if (PrintSource)
       Stream.print(llvm::outs());
     if (PrintTokens)
       llvm::outs() << Stream;
   }
 
   return 0;
 }

clang/unittests/Tooling/Syntax/Pseudo/DirectiveMapTest.cpp

)cpp";
   EXPECT_EQ(tok::pp_if, X.Branches.front().first.Kind);
   EXPECT_THAT(X.Branches.front().second.Chunks,
               ElementsAre(chunkKind(Chunk::K_Code)));
   // The missing terminating directive is marked as pp_not_keyword.
   EXPECT_EQ(tok::pp_not_keyword, X.End.Kind);
   EXPECT_EQ(0u, X.End.Tokens.size());
 }
 
+TEST(DirectiveMap, ChooseBranches) {
+  LangOptions Opts;
+  const std::string Cases[] = {
+      R"cpp(
+        // Branches with no alternatives are taken
+        #if COND // TAKEN
+        int x;
+        #endif
+      )cpp",
+
+      R"cpp(
+        // Empty branches are better than nothing
+        #if COND // TAKEN
+        #endif
+      )cpp",
+
+      R"cpp(
+        // Trivially false branches are not taken, even with no alternatives.
+        #if 0
+        int x;
+        #endif
+      )cpp",
+
+      R"cpp(
+        // Longer branches are preferred over shorter branches
+        #if COND // TAKEN
+        int x = 1;
+        #else
+        int x;
+        #endif
+
+        #if COND
+        int x;
+        #else // TAKEN
+        int x = 1;
+        #endif
+      )cpp",
+
+      R"cpp(
+        // Trivially true branches are taken if previous branches are trivial.
+        #if 1 // TAKEN
+        #else
+          int x = 1;
+        #endif
+
+        #if 0
+          int x = 1;
+        #elif 0
+          int x = 2;
+        #elif 1 // TAKEN
+          int x;
+        #endif
+
+        #if 0
+          int x = 1;
+        #elif FOO // TAKEN
+          int x = 2;
+        #elif 1
+          int x;
+        #endif
+      )cpp",
+
+      R"cpp(
+        // #else is a trivially true branch
+        #if 0
+          int x = 1;
+        #elif 0
+          int x = 2;
+        #else // TAKEN
+          int x;
+        #endif
+      )cpp",
+
+      R"cpp(
+        // Directives break ties, but nondirective text is more important.
+        #if FOO
+          #define A 1 2 3
+        #else // TAKEN
+          #define B 4 5 6
+          #define C 7 8 9
+        #endif
+
+        #if FOO // TAKEN
+          ;
+          #define A 1 2 3
+        #else
+          #define B 4 5 6
+          #define C 7 8 9
+        #endif
+      )cpp",
+
+      R"cpp(
+        // Avoid #error directives.
+        #if FOO
+          int x = 42;
+          #error This branch is no good
+        #else // TAKEN
+        #endif
+
+        #if FOO
+          // All paths here lead to errors.
+          int x = 42;
+          #if 1 // TAKEN
+            #if COND // TAKEN
+              #error This branch is no good
+            #else
+              #error This one is no good either
+            #endif
+          #endif
+        #else // TAKEN
+        #endif
+      )cpp",
+
+      R"cpp(
+        // Populate taken branches recursively.
+        #if FOO // TAKEN
+          int x = 42;
+          #if BAR
+            ;
+          #else // TAKEN
+            int y = 43;
+          #endif
+        #else
+          int x;
+          #if BAR // TAKEN
+            int y;
+          #else
+            ;
+          #endif
+        #endif
+      )cpp",
+  };
+  for (const auto &Code : Cases) {
+    TokenStream S = cook(lex(Code, Opts), Opts);
+
+    std::function<void(const DirectiveMap &)> Verify =
+        [&](const DirectiveMap &M) {
+          for (const auto &C : M.Chunks) {
+            if (C.kind() != DirectiveMap::Chunk::K_Conditional)
+              continue;
+            const DirectiveMap::Conditional &Cond(C);
+            for (unsigned I = 0; I < Cond.Branches.size(); ++I) {
+              auto Directive = S.tokens(Cond.Branches[I].first.Tokens);
+              EXPECT_EQ(I == Cond.Taken, Directive.back().text() == "// TAKEN")
+                  << "At line " << Directive.front().Line << " of: " << Code;
+              Verify(Cond.Branches[I].second);
+            }
+          }
+        };
+
+    DirectiveMap Map = DirectiveMap::parse(S);
+    chooseConditionalBranches(Map, S);
+    Verify(Map);
+  }
+}
+
 } // namespace
 } // namespace pseudo
 } // namespace syntax
 } // namespace clang