Add support for custom _forw functions

This commit adds support for custom (user-provided) `_forw` functions.
A `_forw` function, if available, is called in place of the actual
function.

For example, if the primal code contains:

```cpp
someFn(u, v, w);
```

and user has defined a custom `_forw` function for `someFn` as follows:

```cpp
namespace clad {
  namespace custom_derivatives {
    void someFn_forw(double u, double v, double w, double *d_u,
      double *d_v, double *dw) {
      // ...
      // ...
    }
  }
}
```

Then clad will generate the derivative function as follows:

```cpp
// forward-pass
clad::custom_derivatives::someFn_forw(u, v, w, d_u, d_v, d_w);
// ...

// reverse-pass; no change in reverse-pass
someFn_pullback(u, v, w, d_u, d_v, d_w);
// ...
```

But more importantly, why do we need such a functionality? Two reasons:

- Supporting reference/pointer return types in the reverse-mode. This
  has been discussed at great length here:
#425 (#425)

- Supporting types whose elements grows dynamically, such as
  `std::vector` and `std::map`. The issue is that we correctly
  need to update the size/property of the adjoint variable when a
  function call updates the size/property of the corresponding primal
  variable. For example: a call to `vec.push_back(...)` should update
  the size of `_d_vec` as well. However, the actual function call does
  not modify the adjoint variable in any way. Here comes `_forw` functions
  to the rescue. `_forw` functions makes it possible to adjust the adjoint
  variable size/properties along with executing the actual function call.

Please note that `_forw` function signature takes adjoint variables as
arguments and return `clad::ValueAndAdjoint<U, V>` to support the
reference/pointer return type.

include/clad/Differentiator/BuiltinDerivatives.h

@@ -30,6 +30,11 @@ template <typename T, typename U> struct ValueAndPushforward {
   }
 };
 
+template <typename T, typename U> struct ValueAndAdjoint {
+  T value;
+  U adjoint;
+};
+
 /// It is used to identify constructor custom pushforwards. For
 /// constructor custom pushforward functions, we cannot use the same
 /// strategy which we use for custom pushforward for member

include/clad/Differentiator/CladUtils.h

@@ -321,6 +321,9 @@ namespace clad {
 
     bool IsMemoryFunction(const clang::FunctionDecl* FD);
     bool IsMemoryDeallocationFunction(const clang::FunctionDecl* FD);
+
+    /// Returns true if QT is a non-const reference type.
+    bool isNonConstReferenceType(clang::QualType QT);
     } // namespace utils
     } // namespace clad
 

include/clad/Differentiator/Differentiator.h

@@ -22,12 +22,8 @@
 #include <cstring>
 
 namespace clad {
-template <typename T, typename U> struct ValueAndAdjoint {
-  T value;
-  U adjoint;
-};
 
-  /// \returns the size of a c-style string
+/// \returns the size of a c-style string
 inline CUDA_HOST_DEVICE unsigned int GetLength(const char* code) {
   const char* code_copy = code;
 #ifdef __CUDACC__
@@ -507,7 +503,7 @@ inline CUDA_HOST_DEVICE unsigned int GetLength(const char* code) {
 
   // Gradient Structure for Reverse Mode Enzyme
   template <unsigned N> struct EnzymeGradient { double d_arr[N]; };
-}
+  } // namespace clad
 #endif // CLAD_DIFFERENTIATOR
 
 // Enable clad after the header was included.

include/clad/Differentiator/ReverseModeVisitor.h

@@ -97,6 +97,11 @@ namespace clad {
     // Function to Differentiate with Enzyme as Backend
     void DifferentiateWithEnzyme();
 
+    /// Tries to find and build call to user-provided `_forw` function.
+    clang::Expr* BuildCallToCustomForwPassFn(
+        const clang::FunctionDecl* FD, llvm::ArrayRef<clang::Expr*> primalArgs,
+        llvm::ArrayRef<clang::Expr*> derivedArgs, clang::Expr* baseExpr);
+
   public:
     using direction = rmv::direction;
     clang::Expr* dfdx() {

include/clad/Differentiator/STLBuiltins.h

@@ -202,6 +202,35 @@ void fill_pushforward(::std::array<T, N>* a, const T& u,
   d_a->fill(d_u);
 }
 
+template <typename T, typename U>
+void push_back_forw(::std::vector<T>* v, U val, ::std::vector<T>* d_v,
+                    U* d_val) {
+  v->push_back(val);
+  d_v->push_back(0);
+}
+
+template <typename T, typename U>
+void push_back_pullback(::std::vector<T>* v, U val, ::std::vector<T>* d_v,
+                        U* d_val) {
+  *d_val += d_v->back();
+  d_v->pop_back();
+}
+
+template <typename T>
+clad::ValueAndAdjoint<T&, T&> operator_subscript_forw(
+    ::std::vector<T>* vec, typename ::std::vector<T>::size_type idx,
+    ::std::vector<T>* d_vec, typename ::std::vector<T>::size_type* d_idx) {
+  return {(*vec)[idx], (*d_vec)[idx]};
+}
+
+template <typename T, typename P>
+void operator_subscript_pullback(::std::vector<T>* vec,
+                                 typename ::std::vector<T>::size_type idx,
+                                 P d_y, ::std::vector<T>* d_vec,
+                                 typename ::std::vector<T>::size_type* d_idx) {
+  (*d_vec)[idx] += d_y;
+}
+
 } // namespace class_functions
 } // namespace custom_derivatives
 } // namespace clad

lib/Differentiator/CladUtils.cpp

@@ -705,5 +705,10 @@ namespace clad {
       return FD->getNameAsString() == "free";
 #endif
     }
+
+    bool isNonConstReferenceType(clang::QualType QT) {
+      return QT->isReferenceType() &&
+             !QT.getNonReferenceType().isConstQualified();
+    }
   } // namespace utils
 } // namespace clad

lib/Differentiator/ReverseModeVisitor.cpp

@@ -1799,6 +1799,7 @@ Expr* getArraySizeExpr(const ArrayType* AT, ASTContext& context,
 
     // Stores differentiation result of implicit `this` object, if any.
     StmtDiff baseDiff;
+    Expr* baseExpr = nullptr;
     // If it has more args or f_darg0 was not found, we look for its pullback
     // function.
     const auto* MD = dyn_cast<CXXMethodDecl>(FD);
@@ -1822,6 +1823,7 @@ Expr* getArraySizeExpr(const ArrayType* AT, ASTContext& context,
             baseOriginalE = OCE->getArg(0);
 
           baseDiff = Visit(baseOriginalE);
+          baseExpr = baseDiff.getExpr();
           Expr* baseDiffStore = GlobalStoreAndRef(baseDiff.getExpr());
           baseDiff.updateStmt(baseDiffStore);
           Expr* baseDerivative = baseDiff.getExpr_dx();
@@ -2007,8 +2009,18 @@ Expr* getArraySizeExpr(const ArrayType* AT, ASTContext& context,
     Expr* call = nullptr;
 
     QualType returnType = FD->getReturnType();
-    if (returnType->isReferenceType() &&
-        !returnType.getNonReferenceType().isConstQualified()) {
+    if (Expr* customForwardPassCE = BuildCallToCustomForwPassFn(
+            FD, CallArgs, DerivedCallOutputArgs, baseExpr)) {
+      if (!utils::isNonConstReferenceType(returnType))
+        return StmtDiff{customForwardPassCE};
+      auto* callRes = StoreAndRef(customForwardPassCE);
+      auto* resValue =
+          utils::BuildMemberExpr(m_Sema, getCurrentScope(), callRes, "value");
+      auto* resAdjoint =
+          utils::BuildMemberExpr(m_Sema, getCurrentScope(), callRes, "adjoint");
+      return StmtDiff(resValue, nullptr, resAdjoint);
+    }
+    if (utils::isNonConstReferenceType(returnType)) {
       DiffRequest calleeFnForwPassReq;
       calleeFnForwPassReq.Function = FD;
       calleeFnForwPassReq.Mode = DiffMode::reverse_mode_forward_pass;
@@ -4260,4 +4272,24 @@ Expr* getArraySizeExpr(const ArrayType* AT, ASTContext& context,
                                diffParams.end());
     return params;
   }
+
+  Expr* ReverseModeVisitor::BuildCallToCustomForwPassFn(
+      const FunctionDecl* FD, llvm::ArrayRef<Expr*> primalArgs,
+      llvm::ArrayRef<clang::Expr*> derivedArgs, Expr* baseExpr) {
+    std::string forwPassFnName =
+        clad::utils::ComputeEffectiveFnName(FD) + "_forw";
+    llvm::SmallVector<Expr*, 4> args;
+    if (baseExpr) {
+      baseExpr = BuildOp(UnaryOperatorKind::UO_AddrOf, baseExpr,
+                         m_DiffReq->getLocation());
+      args.push_back(baseExpr);
+    }
+    args.append(primalArgs.begin(), primalArgs.end());
+    args.append(derivedArgs.begin(), derivedArgs.end());
+    Expr* customForwPassCE =
+        m_Builder.BuildCallToCustomDerivativeOrNumericalDiff(
+            forwPassFnName, args, getCurrentScope(),
+            const_cast<DeclContext*>(FD->getDeclContext()));
+    return customForwPassCE;
+  }
 } // end namespace clad

lib/Differentiator/TBRAnalyzer.cpp

@@ -203,9 +203,13 @@ TBRAnalyzer::VarData::VarData(QualType QT, bool forceNonRefType) {
       elemType = pointerType->getPointeeType().getTypePtrOrNull();
     else
       elemType = QT->getArrayElementTypeNoTypeQual();
-    ProfileID nonConstIdxID;
-    auto& idxData = (*m_Val.m_ArrData)[nonConstIdxID];
-    idxData = VarData(QualType::getFromOpaquePtr(elemType));
+    // FIXME: In some cases for Mac, 'elemType' is nullptr for std::vector<T>
+    // internal members.
+    if (elemType) {
+      ProfileID nonConstIdxID;
+      auto& idxData = (*m_Val.m_ArrData)[nonConstIdxID];
+      idxData = VarData(QualType::getFromOpaquePtr(elemType));
+    }
   } else if (QT->isBuiltinType()) {
     m_Type = VarData::FUND_TYPE;
     m_Val.m_FundData = false;

test/Gradient/FunctionCalls.C

@@ -260,10 +260,6 @@ double fn7(double i, double j) {
 
 // CHECK: void custom_identity_pullback(double &i, double _d_y, double *_d_i);
 
-// CHECK: clad::ValueAndAdjoint<double &, double &> custom_identity_forw(double &i, double *d_i) {
-// CHECK-NEXT:     return {i, *d_i};
-// CHECK-NEXT: }
-
 // CHECK: void fn7_grad(double i, double j, double *_d_i, double *_d_j) {
 // CHECK-NEXT:     double _t0 = i;
 // CHECK-NEXT:     clad::ValueAndAdjoint<double &, double &> _t1 = identity_forw(i, &*_d_i);
@@ -274,7 +270,7 @@ double fn7(double i, double j) {
 // CHECK-NEXT:     double &_d_l = _t3.adjoint;
 // CHECK-NEXT:     double &l = _t3.value;
 // CHECK-NEXT:     double _t4 = i;
-// CHECK-NEXT:     clad::ValueAndAdjoint<double &, double &> _t5 = custom_identity_forw(i, &*_d_i);
+// CHECK-NEXT:     clad::ValueAndAdjoint<double &, double &> _t5 = {{.*}}custom_derivatives::custom_identity_forw(i, &*_d_i);
 // CHECK-NEXT:     double &_d_temp = _t5.adjoint;
 // CHECK-NEXT:     double &temp = _t5.value;
 // CHECK-NEXT:     double _t6 = k;