Add bitmasked-option for tbr analysis

include/clad/Differentiator/CladConfig.h

@@ -24,20 +24,25 @@ enum order {
 enum opts {
   use_enzyme = 1 << ORDER_BITS,
   vector_mode = 1 << (ORDER_BITS + 1),
+
+  // Storing two bits for tbr analysis.
+  // 00 - default, 01 - enable, 10 - disable, 11 - not used / invalid
+  enable_tbr = 1 << (ORDER_BITS + 2),
+  disable_tbr = 1 << (ORDER_BITS + 3),
 }; // enum opts
 
-constexpr unsigned GetDerivativeOrder(unsigned const bitmasked_opts) {
+constexpr unsigned GetDerivativeOrder(const unsigned bitmasked_opts) {
   return bitmasked_opts & ORDER_MASK;
 }
 
-constexpr bool HasOption(unsigned const bitmasked_opts, unsigned const option) {
+constexpr bool HasOption(const unsigned bitmasked_opts, const unsigned option) {
   return (bitmasked_opts & option) == option;
 }
 
 constexpr unsigned GetBitmaskedOpts() { return 0; }
-constexpr unsigned GetBitmaskedOpts(unsigned const first) { return first; }
+constexpr unsigned GetBitmaskedOpts(const unsigned first) { return first; }
 template <typename... Opts>
-constexpr unsigned GetBitmaskedOpts(unsigned const first, Opts... opts) {
+constexpr unsigned GetBitmaskedOpts(const unsigned first, Opts... opts) {
   return first | GetBitmaskedOpts(opts...);
 }
 

include/clad/Differentiator/DiffPlanner.h

@@ -86,6 +86,12 @@ namespace clad {
   using DiffSchedule = llvm::SmallVector<DiffRequest, 16>;
   using DiffInterval = std::vector<clang::SourceRange>;
 
+  struct RequestOptions {
+    /// This is a flag to indicate the default behaviour to enable/disable
+    /// TBR analysis during reverse-mode differentiation.
+    bool EnableTBRAnalysis = false;
+  };
+
   class DiffCollector: public clang::RecursiveASTVisitor<DiffCollector> {
     /// The source interval where clad was activated.
     ///
@@ -101,9 +107,11 @@ namespace clad {
     const clang::FunctionDecl* m_TopMostFD = nullptr;
     clang::Sema& m_Sema;
 
+    RequestOptions& m_Options;
+
   public:
     DiffCollector(clang::DeclGroupRef DGR, DiffInterval& Interval,
-                  DiffSchedule& plans, clang::Sema& S);
+                  DiffSchedule& plans, clang::Sema& S, RequestOptions& opts);
     bool VisitCallExpr(clang::CallExpr* E);
 
   private:

include/clad/Differentiator/Differentiator.h

@@ -358,9 +358,8 @@ inline CUDA_HOST_DEVICE unsigned int GetLength(const char* code) {
   /// \param[in] args independent parameters information
   /// \returns `CladFunction` object to access the corresponding derived
   /// function.
-  template <unsigned... BitMaskedOpts /*To check for enzyme*/,
-            typename ArgSpec = const char*, typename F,
-            typename DerivedFnType = GradientDerivedFnTraits_t<F>,
+  template <unsigned... BitMaskedOpts, typename ArgSpec = const char*,
+            typename F, typename DerivedFnType = GradientDerivedFnTraits_t<F>,
             typename = typename std::enable_if<
                 !std::is_class<remove_reference_and_pointer_t<F>>::value>::type>
   CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>, true> __attribute__((
@@ -376,9 +375,8 @@ inline CUDA_HOST_DEVICE unsigned int GetLength(const char* code) {
   /// Specialization for differentiating functors.
   /// The specialization is needed because objects have to be passed
   /// by reference whereas functions have to be passed by value.
-  template <unsigned... BitMaskedOpts /*To check for enzyme*/,
-            typename ArgSpec = const char*, typename F,
-            typename DerivedFnType = GradientDerivedFnTraits_t<F>,
+  template <unsigned... BitMaskedOpts, typename ArgSpec = const char*,
+            typename F, typename DerivedFnType = GradientDerivedFnTraits_t<F>,
             typename = typename std::enable_if<
                 std::is_class<remove_reference_and_pointer_t<F>>::value>::type>
   CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>, true> __attribute__((
@@ -397,38 +395,34 @@ inline CUDA_HOST_DEVICE unsigned int GetLength(const char* code) {
   /// \param[in] args independent parameters information
   /// \returns `CladFunction` object to access the corresponding derived
   /// function.
-  template <typename ArgSpec = const char*, typename F,
-            typename DerivedFnType = HessianDerivedFnTraits_t<F>,
+  template <unsigned... BitMaskedOpts, typename ArgSpec = const char*,
+            typename F, typename DerivedFnType = HessianDerivedFnTraits_t<F>,
             typename = typename std::enable_if<
                 !std::is_class<remove_reference_and_pointer_t<F>>::value>::type>
   CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>> __attribute__((
       annotate("H")))
   hessian(F f, ArgSpec args = "",
           DerivedFnType derivedFn = static_cast<DerivedFnType>(nullptr),
           const char* code = "") {
-    assert(f && "Must pass in a non-0 argument");
-    return CladFunction<
-        DerivedFnType,
-        ExtractFunctorTraits_t<F>>(derivedFn /* will be replaced by hessian*/,
-                                   code);
+      assert(f && "Must pass in a non-0 argument");
+      return CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>>(
+          derivedFn /* will be replaced by hessian*/, code);
   }
 
   /// Specialization for differentiating functors.
   /// The specialization is needed because objects have to be passed
   /// by reference whereas functions have to be passed by value.
-  template <typename ArgSpec = const char*, typename F,
-            typename DerivedFnType = HessianDerivedFnTraits_t<F>,
+  template <unsigned... BitMaskedOpts, typename ArgSpec = const char*,
+            typename F, typename DerivedFnType = HessianDerivedFnTraits_t<F>,
             typename = typename std::enable_if<
                 std::is_class<remove_reference_and_pointer_t<F>>::value>::type>
   CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>> __attribute__((
       annotate("H")))
   hessian(F&& f, ArgSpec args = "",
           DerivedFnType derivedFn = static_cast<DerivedFnType>(nullptr),
           const char* code = "") {
-    return CladFunction<
-        DerivedFnType,
-        ExtractFunctorTraits_t<F>>(derivedFn /* will be replaced by hessian*/,
-                                   code, f);
+      return CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>>(
+          derivedFn /* will be replaced by hessian*/, code, f);
   }
 
   /// Generates function which computes jacobian matrix of the given function
@@ -438,38 +432,34 @@ inline CUDA_HOST_DEVICE unsigned int GetLength(const char* code) {
   /// \param[in] args independent parameters information
   /// \returns `CladFunction` object to access the corresponding derived
   /// function.
-  template <typename ArgSpec = const char*, typename F,
-            typename DerivedFnType = JacobianDerivedFnTraits_t<F>,
+  template <unsigned... BitMaskedOpts, typename ArgSpec = const char*,
+            typename F, typename DerivedFnType = JacobianDerivedFnTraits_t<F>,
             typename = typename std::enable_if<
                 !std::is_class<remove_reference_and_pointer_t<F>>::value>::type>
   CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>> __attribute__((
       annotate("J")))
   jacobian(F f, ArgSpec args = "",
            DerivedFnType derivedFn = static_cast<DerivedFnType>(nullptr),
            const char* code = "") {
-    assert(f && "Must pass in a non-0 argument");
-    return CladFunction<
-        DerivedFnType,
-        ExtractFunctorTraits_t<F>>(derivedFn /* will be replaced by Jacobian*/,
-                                   code);
+      assert(f && "Must pass in a non-0 argument");
+      return CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>>(
+          derivedFn /* will be replaced by Jacobian*/, code);
   }
 
   /// Specialization for differentiating functors.
   /// The specialization is needed because objects have to be passed
   /// by reference whereas functions have to be passed by value.
-  template <typename ArgSpec = const char*, typename F,
-            typename DerivedFnType = JacobianDerivedFnTraits_t<F>,
+  template <unsigned... BitMaskedOpts, typename ArgSpec = const char*,
+            typename F, typename DerivedFnType = JacobianDerivedFnTraits_t<F>,
             typename = typename std::enable_if<
                 std::is_class<remove_reference_and_pointer_t<F>>::value>::type>
   CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>> __attribute__((
       annotate("J")))
   jacobian(F&& f, ArgSpec args = "",
            DerivedFnType derivedFn = static_cast<DerivedFnType>(nullptr),
            const char* code = "") {
-    return CladFunction<
-        DerivedFnType,
-        ExtractFunctorTraits_t<F>>(derivedFn /* will be replaced by Jacobian*/,
-                                   code, f);
+      return CladFunction<DerivedFnType, ExtractFunctorTraits_t<F>>(
+          derivedFn /* will be replaced by Jacobian*/, code, f);
   }
 
   template <typename ArgSpec = const char*, typename F,

lib/Differentiator/DiffPlanner.cpp

@@ -232,9 +232,10 @@ namespace clad {
   }
 
   DiffCollector::DiffCollector(DeclGroupRef DGR, DiffInterval& Interval,
-                               DiffSchedule& plans, clang::Sema& S)
+                               DiffSchedule& plans, clang::Sema& S,
+                               RequestOptions& opts)
       : m_Interval(Interval), m_DiffPlans(plans), m_TopMostFD(nullptr),
-        m_Sema(S) {
+        m_Sema(S), m_Options(opts) {
 
     if (Interval.empty())
       return;
@@ -556,27 +557,52 @@ namespace clad {
         return true;
       DiffRequest request{};
 
-      if (A->getAnnotation().equals("D")) {
-        request.Mode = DiffMode::forward;
-
-        // bitmask_opts is a template pack of unsigned integers, so we need to
-        // do bitwise or of all the values to get the final value.
-        unsigned bitmasked_opts_value = 0;
+      // bitmask_opts is a template pack of unsigned integers, so we need to
+      // do bitwise or of all the values to get the final value.
+      unsigned bitmasked_opts_value = 0;
+      bool enable_tbr_in_req = false;
+      bool disable_tbr_in_req = false;
+      if (!A->getAnnotation().equals("E") && FD->getTemplateSpecializationArgs()) {
         const auto template_arg = FD->getTemplateSpecializationArgs()->get(0);
         if (template_arg.getKind() == TemplateArgument::Pack)
           for (const auto& arg :
                FD->getTemplateSpecializationArgs()->get(0).pack_elements())
             bitmasked_opts_value |= arg.getAsIntegral().getExtValue();
         else
           bitmasked_opts_value = template_arg.getAsIntegral().getExtValue();
+
+        // Set option for TBR analysis.
+        enable_tbr_in_req =
+            clad::HasOption(bitmasked_opts_value, clad::opts::enable_tbr);
+        disable_tbr_in_req =
+            clad::HasOption(bitmasked_opts_value, clad::opts::disable_tbr);
+        if (enable_tbr_in_req && disable_tbr_in_req) {
+          utils::EmitDiag(m_Sema, DiagnosticsEngine::Error, endLoc,
+                          "Both enable and disable TBR options are specified.");
+          return true;
+        }
+        if (enable_tbr_in_req || disable_tbr_in_req) {
+          // override the default value of TBR analysis.
+          request.EnableTBRAnalysis = enable_tbr_in_req && !disable_tbr_in_req;
+        } else {
+          request.EnableTBRAnalysis = m_Options.EnableTBRAnalysis;
+        }
+      }
+
+      if (A->getAnnotation().equals("D")) {
+        request.Mode = DiffMode::forward;
         unsigned derivative_order =
             clad::GetDerivativeOrder(bitmasked_opts_value);
         if (derivative_order == 0) {
           derivative_order = 1; // default to first order derivative.
         }
         request.RequestedDerivativeOrder = derivative_order;
-        if (clad::HasOption(bitmasked_opts_value, clad::opts::use_enzyme)) {
+        if (clad::HasOption(bitmasked_opts_value, clad::opts::use_enzyme))
           request.use_enzyme = true;
+        if (enable_tbr_in_req) {
+          utils::EmitDiag(m_Sema, DiagnosticsEngine::Error, endLoc,
+                          "TBR analysis is not yet supported in forward mode.");
+          return true;
         }
         if (clad::HasOption(bitmasked_opts_value, clad::opts::vector_mode)) {
           request.Mode = DiffMode::vector_forward_mode;
@@ -601,17 +627,6 @@ namespace clad {
         request.Mode = DiffMode::jacobian;
       } else if (A->getAnnotation().equals("G")) {
         request.Mode = DiffMode::reverse;
-
-        // bitmask_opts is a template pack of unsigned integers, so we need to
-        // do bitwise or of all the values to get the final value.
-        unsigned bitmasked_opts_value = 0;
-        const auto template_arg = FD->getTemplateSpecializationArgs()->get(0);
-        if (template_arg.getKind() == TemplateArgument::Pack)
-          for (const auto& arg :
-               FD->getTemplateSpecializationArgs()->get(0).pack_elements())
-            bitmasked_opts_value |= arg.getAsIntegral().getExtValue();
-        else
-          bitmasked_opts_value = template_arg.getAsIntegral().getExtValue();
         if (clad::HasOption(bitmasked_opts_value, clad::opts::use_enzyme))
           request.use_enzyme = true;
         // reverse vector mode is not yet supported.

test/Analyses/TBR.cpp

@@ -1,4 +1,4 @@
-// RUN: %cladclang -mllvm -debug-only=clad-tbr -Xclang -plugin-arg-clad -Xclang -enable-tbr %s -I%S/../../include -oReverseLoops.out 2>&1 | FileCheck %s
+// RUN: %cladclang -mllvm -debug-only=clad-tbr %s -I%S/../../include -oReverseLoops.out 2>&1 | FileCheck %s
 // REQUIRES: asserts
 //CHECK-NOT: {{.*error|warning|note:.*}}
 
@@ -13,7 +13,7 @@ double f1(double x) {
 
 #define TEST(F, x) { \
   result[0] = 0; \
-  auto F##grad = clad::gradient(F);\
+  auto F##grad = clad::gradient<clad::opts::enable_tbr>(F);\
   F##grad.execute(x, result);\
   printf("{%.2f}\n", result[0]); \
 }

test/Gradient/Pointers.C

@@ -617,7 +617,7 @@ int main() {
   d_structPointer.execute(5, &d_x);
   printf("%.2f\n", d_x); // CHECK-EXEC: 1.00
 
-  auto d_cStyleMemoryAlloc = clad::gradient(cStyleMemoryAlloc, "x");
+  auto d_cStyleMemoryAlloc = clad::gradient<clad::opts::disable_tbr>(cStyleMemoryAlloc, "x");
   d_x = 0;
   d_cStyleMemoryAlloc.execute(5, 7, &d_x);
   printf("%.2f\n", d_x); // CHECK-EXEC: 4.00

test/Misc/Args.C

@@ -5,6 +5,9 @@
 // CHECK_HELP-NEXT: -fdump-derived-fn
 // CHECK_HELP-NEXT: -fdump-derived-fn-ast
 // CHECK_HELP-NEXT: -fgenerate-source-file
+// CHECK_HELP-NEXT: -fno-validate-clang-version
+// CHECK_HELP-NEXT: -enable-tbr
+// CHECK_HELP-NEXT: -disable-tbr
 // CHECK_HELP-NEXT: -fcustom-estimation-model
 // CHECK_HELP-NEXT: -fprint-num-diff-errors
 // CHECK_HELP-NEXT: -help
@@ -23,3 +26,7 @@
 // RUN: -Xclang %t.so %S/../../demos/ErrorEstimation/CustomModel/test.cpp \
 // RUN: -I%S/../../include 2>&1 | FileCheck --check-prefix=CHECK_SO_INVALID %s
 // CHECK_SO_INVALID: Failed to load '{{.*.so}}', {{.*}}. Aborting.
+
+// RUN: clang -fsyntax-only -fplugin=%cladlib -Xclang -plugin-arg-clad -Xclang -enable-tbr \
+// RUN:  -Xclang -plugin-arg-clad -Xclang -disable-tbr %s 2>&1 | FileCheck --check-prefix=CHECK_TBR %s
+// CHECK_TBR: -enable-tbr and -disable-tbr cannot be used together

tools/ClangPlugin.cpp

@@ -107,18 +107,15 @@ namespace clad {
       if (m_HandleTopLevelDeclInternal)
         return true;
 
+      RequestOptions opts{};
+      SetRequestOptions(opts);
       DiffSchedule requests{};
-      DiffCollector collector(DGR, CladEnabledRange, requests, m_CI.getSema());
+      DiffCollector collector(DGR, CladEnabledRange, requests, m_CI.getSema(),
+                              opts);
 
       if (requests.empty())
         return true;
 
-      // FIXME: flags have to be set manually since DiffCollector's constructor
-      // does not have access to m_DO.
-      if (m_DO.EnableTBRAnalysis)
-        for (DiffRequest& request : requests)
-          request.EnableTBRAnalysis = true;
-
       // FIXME: Remove the PerformPendingInstantiations altogether. We should
       // somehow make the relevant functions referenced.
       // Instantiate all pending for instantiations templates, because we will
@@ -318,6 +315,19 @@ namespace clad {
       m_HasRuntime = !R.empty();
       return m_HasRuntime;
     }
+
+    void CladPlugin::SetRequestOptions(RequestOptions& opts) {
+      /// ---------- Set TBR analysis flag ----------
+      // If user has explicitly specified the mode for TBR analysis, use it.
+      if (m_DO.EnableTBRAnalysis || m_DO.DisableTbrAnalysis) {
+        opts.EnableTBRAnalysis =
+            m_DO.EnableTBRAnalysis && !m_DO.DisableTbrAnalysis;
+      } else {
+        // If user has not specified the mode for TBR analysis, use the default
+        // mode.
+        opts.EnableTBRAnalysis = false; // Default mode.
+      }
+    }
   } // end namespace plugin
 
   clad::CladTimerGroup::CladTimerGroup()