[clang] nonblocking/nonallocating attributes: 2nd pass caller/callee analysis (PR #99656)

[via cfe-commits]

================
@@ -2397,6 +2397,1262 @@ class UnsafeBufferUsageReporter : public UnsafeBufferUsageHandler {
 };
 } // namespace
 
+// =============================================================================
+
+namespace FXAnalysis {
+
+enum class DiagnosticID : uint8_t {
+  None = 0, // sentinel for an empty Diagnostic
+  Throws,
+  Catches,
+  CallsObjC,
+  AllocatesMemory,
+  HasStaticLocal,
+  AccessesThreadLocal,
+
+  // These only apply to callees, where the analysis stops at the Decl
+  DeclDisallowsInference,
+
+  CallsDeclWithoutEffect,
+  CallsExprWithoutEffect,
+};
+
+// Holds an effect diagnosis, potentially for the entire duration of the
+// analysis phase, in order to refer to it when explaining why a caller has been
+// made unsafe by a callee.
+struct Diagnostic {
+  FunctionEffect Effect;
+  DiagnosticID ID = DiagnosticID::None;
+  SourceLocation Loc;
+  const Decl *Callee = nullptr; // only valid for Calls*
+
+  Diagnostic() = default;
+
+  Diagnostic(const FunctionEffect &Effect, DiagnosticID ID, SourceLocation Loc,
+             const Decl *Callee = nullptr)
+      : Effect(Effect), ID(ID), Loc(Loc), Callee(Callee) {}
+};
+
+enum class SpecialFuncType : uint8_t { None, OperatorNew, OperatorDelete };
+enum class CallType {
+  // unknown: probably function pointer
+  Unknown,
+  Function,
+  Virtual,
+  Block
+};
+
+// Return whether a function's effects CAN be verified.
+// The question of whether it SHOULD be verified is independent.
+static bool functionIsVerifiable(const FunctionDecl *FD) {
+  if (!(FD->hasBody() || FD->isInlined())) {
+    // externally defined; we couldn't verify if we wanted to.
+    return false;
+  }
+  if (FD->isTrivial()) {
+    // Otherwise `struct x { int a; };` would have an unverifiable default
+    // constructor.
+    return true;
+  }
+  return true;
+}
+
+/// A mutable set of FunctionEffect, for use in places where any conditions
+/// have been resolved or can be ignored.
+class EffectSet {
+  // This implementation optimizes footprint, since we hold one of these for
+  // every function visited, which, due to inference, can be many more functions
+  // than have declared effects.
+
+  template <typename T, typename SizeT, SizeT Capacity> struct FixedVector {
+    SizeT Count = 0;
+    T Items[Capacity] = {};
+
+    using value_type = T;
+
+    using iterator = T *;
+    using const_iterator = const T *;
+    iterator begin() { return &Items[0]; }
+    iterator end() { return &Items[Count]; }
+    const_iterator begin() const { return &Items[0]; }
+    const_iterator end() const { return &Items[Count]; }
+    const_iterator cbegin() const { return &Items[0]; }
+    const_iterator cend() const { return &Items[Count]; }
+
+    void insert(iterator I, const T &Value) {
+      assert(Count < Capacity);
+      iterator E = end();
+      if (I != E)
+        std::copy_backward(I, E, E + 1);
+      *I = Value;
+      ++Count;
+    }
+
+    void push_back(const T &Value) {
+      assert(Count < Capacity);
+      Items[Count++] = Value;
+    }
+  };
+
+  // As long as FunctionEffect is only 1 byte, and there are only 2 verifiable
+  // effects, this fixed-size vector with a capacity of 7 is more than
+  // sufficient and is only 8 bytes.
+  FixedVector<FunctionEffect, uint8_t, 7> Impl;
+
+public:
+  EffectSet() = default;
+  explicit EffectSet(FunctionEffectsRef FX) { insert(FX); }
+
+  operator ArrayRef<FunctionEffect>() const {
+    return ArrayRef(Impl.cbegin(), Impl.cend());
+  }
+
+  using iterator = const FunctionEffect *;
+  iterator begin() const { return Impl.cbegin(); }
+  iterator end() const { return Impl.cend(); }
+
+  void insert(const FunctionEffect &Effect) {
+    FunctionEffect *Iter = Impl.begin();
+    FunctionEffect *End = Impl.end();
+    // linear search; lower_bound is overkill for a tiny vector like this
+    for (; Iter != End; ++Iter) {
+      if (*Iter == Effect)
+        return;
+      if (Effect < *Iter)
+        break;
+    }
+    Impl.insert(Iter, Effect);
+  }
+  void insert(const EffectSet &Set) {
+    for (const FunctionEffect &Item : Set) {
+      // push_back because set is already sorted
+      Impl.push_back(Item);
+    }
+  }
+  void insert(FunctionEffectsRef FX) {
+    for (const FunctionEffectWithCondition &EC : FX) {
+      assert(EC.Cond.getCondition() ==
+             nullptr); // should be resolved by now, right?
+      // push_back because set is already sorted
+      Impl.push_back(EC.Effect);
+    }
+  }
+  bool contains(const FunctionEffect::Kind EK) const {
+    for (const FunctionEffect &E : Impl)
+      if (E.kind() == EK)
+        return true;
+    return false;
+  }
+
+  void dump(llvm::raw_ostream &OS) const;
+
+  static EffectSet difference(ArrayRef<FunctionEffect> LHS,
+                              ArrayRef<FunctionEffect> RHS) {
+    EffectSet Result;
+    std::set_difference(LHS.begin(), LHS.end(), RHS.begin(), RHS.end(),
+                        std::back_inserter(Result.Impl));
+    return Result;
+  }
+};
+
+LLVM_DUMP_METHOD void EffectSet::dump(llvm::raw_ostream &OS) const {
+  OS << "Effects{";
+  bool First = true;
+  for (const FunctionEffect &Effect : *this) {
+    if (!First)
+      OS << ", ";
+    else
+      First = false;
+    OS << Effect.name();
+  }
+  OS << "}";
+}
+
+// Transitory, more extended information about a callable, which can be a
+// function, block, function pointer, etc.
+struct CallableInfo {
+  // CDecl holds the function's definition, if any.
+  // FunctionDecl if CallType::Function or Virtual
+  // BlockDecl if CallType::Block
+  const Decl *CDecl;
+  mutable std::optional<std::string> MaybeName;
----------------
Sirraide wrote:

I don’t think we need to bother storing this if it’s only used for dumping and emitting diagnostics, neither of which really need to be optimised for speed.

https://github.com/llvm/llvm-project/pull/99656