[flang][runtime] Enable more code for offload device builds. (#67489)

I extended the "closure" of the device code containing the initial
transformational.cpp. The device side of the library should not be
complete at least for some APIs. For example, I tested with C OpenMP
code calling BesselJnX0 with a nullptr descriptor that failed with
a runtime error when executing on a GPU.

I added `--expt-relaxed-constexpr` for NVCC compiler to avoid multiple
warnings about missing `__attribute__((device))` on constexpr methods
coming from C++ header files.

Author: vzakhari

flang/include/flang/Runtime/api-attrs.h

@@ -42,6 +42,18 @@
 #endif
 #endif /* !defined(RT_EXT_API_GROUP_END) */
 
+/*
+ * RT_OFFLOAD_API_GROUP_BEGIN/END pair is placed around definitions
+ * of functions that can be referenced in other modules of Flang
+ * runtime. For OpenMP offload these functions are made "declare target"
+ * making sure they are compiled for the target even though direct
+ * references to them from other "declare target" functions may not
+ * be seen. Host-only functions should not be put in between these
+ * two macros.
+ */
+#define RT_OFFLOAD_API_GROUP_BEGIN RT_EXT_API_GROUP_BEGIN
+#define RT_OFFLOAD_API_GROUP_END RT_EXT_API_GROUP_END
+
 /*
  * RT_VAR_GROUP_BEGIN/END pair is placed around definitions
  * of module scope variables referenced by Flang runtime (directly
@@ -88,4 +100,16 @@
 #endif
 #endif /* !defined(RT_CONST_VAR_ATTRS) */
 
+/*
+ * RT_DEVICE_COMPILATION is defined for any device compilation.
+ * Note that it can only be used reliably with compilers that perform
+ * separate host and device compilations.
+ */
+#if ((defined(__CUDACC__) || defined(__CUDA__)) && defined(__CUDA_ARCH__)) || \
+    (defined(_OPENMP) && (defined(__AMDGCN__) || defined(__NVPTX__)))
+#define RT_DEVICE_COMPILATION 1
+#else
+#undef RT_DEVICE_COMPILATION
+#endif
+
 #endif /* !FORTRAN_RUNTIME_API_ATTRS_H_ */

flang/include/flang/Runtime/descriptor.h

@@ -181,20 +181,21 @@ class Descriptor {
       ISO::CFI_attribute_t attribute = CFI_attribute_other);
 
   // CUDA_TODO: Clang does not support unique_ptr on device.
-  static OwningPtr<Descriptor> Create(TypeCode t, std::size_t elementBytes,
-      void *p = nullptr, int rank = maxRank,
+  static RT_API_ATTRS OwningPtr<Descriptor> Create(TypeCode t,
+      std::size_t elementBytes, void *p = nullptr, int rank = maxRank,
       const SubscriptValue *extent = nullptr,
       ISO::CFI_attribute_t attribute = CFI_attribute_other,
       int derivedTypeLenParameters = 0);
-  static OwningPtr<Descriptor> Create(TypeCategory, int kind, void *p = nullptr,
-      int rank = maxRank, const SubscriptValue *extent = nullptr,
+  static RT_API_ATTRS OwningPtr<Descriptor> Create(TypeCategory, int kind,
+      void *p = nullptr, int rank = maxRank,
+      const SubscriptValue *extent = nullptr,
       ISO::CFI_attribute_t attribute = CFI_attribute_other);
-  static OwningPtr<Descriptor> Create(int characterKind,
+  static RT_API_ATTRS OwningPtr<Descriptor> Create(int characterKind,
       SubscriptValue characters, void *p = nullptr, int rank = maxRank,
       const SubscriptValue *extent = nullptr,
       ISO::CFI_attribute_t attribute = CFI_attribute_other);
-  static OwningPtr<Descriptor> Create(const typeInfo::DerivedType &dt,
-      void *p = nullptr, int rank = maxRank,
+  static RT_API_ATTRS OwningPtr<Descriptor> Create(
+      const typeInfo::DerivedType &dt, void *p = nullptr, int rank = maxRank,
       const SubscriptValue *extent = nullptr,
       ISO::CFI_attribute_t attribute = CFI_attribute_other);
 

flang/include/flang/Runtime/memory.h

@@ -12,31 +12,116 @@
 #ifndef FORTRAN_RUNTIME_MEMORY_H_
 #define FORTRAN_RUNTIME_MEMORY_H_
 
+#include "flang/Runtime/api-attrs.h"
+#include <cassert>
 #include <memory>
+#include <type_traits>
 
 namespace Fortran::runtime {
 
 class Terminator;
 
-[[nodiscard]] void *AllocateMemoryOrCrash(
+[[nodiscard]] RT_API_ATTRS void *AllocateMemoryOrCrash(
     const Terminator &, std::size_t bytes);
 template <typename A> [[nodiscard]] A &AllocateOrCrash(const Terminator &t) {
   return *reinterpret_cast<A *>(AllocateMemoryOrCrash(t, sizeof(A)));
 }
-void FreeMemory(void *);
-template <typename A> void FreeMemory(A *p) {
+RT_API_ATTRS void FreeMemory(void *);
+template <typename A> RT_API_ATTRS void FreeMemory(A *p) {
   FreeMemory(reinterpret_cast<void *>(p));
 }
 template <typename A> void FreeMemoryAndNullify(A *&p) {
   FreeMemory(p);
   p = nullptr;
 }
 
-template <typename A> struct OwningPtrDeleter {
-  void operator()(A *p) { FreeMemory(p); }
+// Very basic implementation mimicking std::unique_ptr.
+// It should work for any offload device compiler.
+// It uses a fixed memory deleter based on FreeMemory(),
+// and does not support array objects with runtime length.
+template <typename A> class OwningPtr {
+public:
+  using pointer_type = A *;
+
+  OwningPtr() = default;
+  RT_API_ATTRS explicit OwningPtr(pointer_type p) : ptr_(p) {}
+  RT_API_ATTRS OwningPtr(const OwningPtr &) = delete;
+  RT_API_ATTRS OwningPtr &operator=(const OwningPtr &) = delete;
+  RT_API_ATTRS OwningPtr(OwningPtr &&other) {
+    ptr_ = other.ptr_;
+    other.ptr_ = pointer_type{};
+  }
+  RT_API_ATTRS OwningPtr &operator=(OwningPtr &&other) {
+    if (this != &other) {
+      delete_ptr(ptr_);
+      ptr_ = other.ptr_;
+      other.ptr_ = pointer_type{};
+    }
+    return *this;
+  }
+  constexpr RT_API_ATTRS OwningPtr(std::nullptr_t) : OwningPtr() {}
+
+  // Delete the pointer, if owns one.
+  RT_API_ATTRS ~OwningPtr() {
+    if (ptr_ != pointer_type{}) {
+      delete_ptr(ptr_);
+      ptr_ = pointer_type{};
+    }
+  }
+
+  // Release the ownership.
+  RT_API_ATTRS pointer_type release() {
+    pointer_type p = ptr_;
+    ptr_ = pointer_type{};
+    return p;
+  }
+
+  // Replace the pointer.
+  RT_API_ATTRS void reset(pointer_type p = pointer_type{}) {
+    std::swap(ptr_, p);
+    if (p != pointer_type{}) {
+      // Delete the owned pointer.
+      delete_ptr(p);
+    }
+  }
+
+  // Exchange the pointer with another object.
+  RT_API_ATTRS void swap(OwningPtr &other) { std::swap(ptr_, other.ptr_); }
+
+  // Get the stored pointer.
+  RT_API_ATTRS pointer_type get() const { return ptr_; }
+
+  RT_API_ATTRS explicit operator bool() const {
+    return get() != pointer_type{};
+  }
+
+  RT_API_ATTRS typename std::add_lvalue_reference<A>::type operator*() const {
+    assert(get() != pointer_type{});
+    return *get();
+  }
+
+  RT_API_ATTRS pointer_type operator->() const { return get(); }
+
+private:
+  RT_API_ATTRS void delete_ptr(pointer_type p) { FreeMemory(p); }
+  pointer_type ptr_{};
 };
 
-template <typename A> using OwningPtr = std::unique_ptr<A, OwningPtrDeleter<A>>;
+template <typename X, typename Y>
+inline RT_API_ATTRS bool operator!=(
+    const OwningPtr<X> &x, const OwningPtr<Y> &y) {
+  return x.get() != y.get();
+}
+
+template <typename X>
+inline RT_API_ATTRS bool operator!=(const OwningPtr<X> &x, std::nullptr_t) {
+  return (bool)x;
+}
+
+template <typename X>
+inline RT_API_ATTRS bool operator!=(std::nullptr_t, const OwningPtr<X> &x) {
+  return (bool)x;
+}
 
 template <typename A> class SizedNew {
 public:

flang/include/flang/Runtime/type-code.h

@@ -26,29 +26,33 @@ class TypeCode {
 
   RT_API_ATTRS int raw() const { return raw_; }
 
-  constexpr bool IsValid() const {
+  constexpr RT_API_ATTRS bool IsValid() const {
     return raw_ >= CFI_type_signed_char && raw_ <= CFI_TYPE_LAST;
   }
-  constexpr bool IsInteger() const {
+  constexpr RT_API_ATTRS bool IsInteger() const {
     return raw_ >= CFI_type_signed_char && raw_ <= CFI_type_ptrdiff_t;
   }
-  constexpr bool IsReal() const {
+  constexpr RT_API_ATTRS bool IsReal() const {
     return raw_ >= CFI_type_half_float && raw_ <= CFI_type_float128;
   }
-  constexpr bool IsComplex() const {
+  constexpr RT_API_ATTRS bool IsComplex() const {
     return raw_ >= CFI_type_half_float_Complex &&
         raw_ <= CFI_type_float128_Complex;
   }
-  constexpr bool IsCharacter() const {
+  constexpr RT_API_ATTRS bool IsCharacter() const {
     return raw_ == CFI_type_char || raw_ == CFI_type_char16_t ||
         raw_ == CFI_type_char32_t;
   }
-  constexpr bool IsLogical() const {
+  constexpr RT_API_ATTRS bool IsLogical() const {
     return raw_ == CFI_type_Bool ||
         (raw_ >= CFI_type_int_least8_t && raw_ <= CFI_type_int_least64_t);
   }
-  constexpr bool IsDerived() const { return raw_ == CFI_type_struct; }
-  constexpr bool IsIntrinsic() const { return IsValid() && !IsDerived(); }
+  constexpr RT_API_ATTRS bool IsDerived() const {
+    return raw_ == CFI_type_struct;
+  }
+  constexpr RT_API_ATTRS bool IsIntrinsic() const {
+    return IsValid() && !IsDerived();
+  }
 
   RT_API_ATTRS std::optional<std::pair<TypeCategory, int>>
   GetCategoryAndKind() const;
@@ -65,7 +69,7 @@ class TypeCode {
       return thisCK && thatCK && *thisCK == *thatCK;
     }
   }
-  bool operator!=(TypeCode that) const { return !(*this == that); }
+  RT_API_ATTRS bool operator!=(TypeCode that) const { return !(*this == that); }
 
 private:
   ISO::CFI_type_t raw_{CFI_type_other};

flang/runtime/CMakeLists.txt

@@ -150,7 +150,10 @@ option(FLANG_EXPERIMENTAL_CUDA_RUNTIME
 
 # List of files that are buildable for all devices.
 set(supported_files
+  descriptor.cpp
+  terminator.cpp
   transformational.cpp
+  type-code.cpp
   )
 
 if (FLANG_EXPERIMENTAL_CUDA_RUNTIME)
@@ -175,6 +178,11 @@ if (FLANG_EXPERIMENTAL_CUDA_RUNTIME)
       -Xclang -fcuda-allow-variadic-functions
       )
   endif()
+  if ("${CMAKE_CUDA_COMPILER_ID}" MATCHES "NVIDIA")
+    set(CUDA_COMPILE_OPTIONS
+      --expt-relaxed-constexpr
+      )
+  endif()
   set_source_files_properties(${supported_files} PROPERTIES COMPILE_OPTIONS
     "${CUDA_COMPILE_OPTIONS}"
     )

flang/runtime/ISO_Fortran_util.h

@@ -18,15 +18,15 @@
 #include <cstdlib>
 
 namespace Fortran::ISO {
-static inline constexpr bool IsCharacterType(CFI_type_t ty) {
+static inline constexpr RT_API_ATTRS bool IsCharacterType(CFI_type_t ty) {
   return ty == CFI_type_char || ty == CFI_type_char16_t ||
       ty == CFI_type_char32_t;
 }
-static inline constexpr bool IsAssumedSize(const CFI_cdesc_t *dv) {
+static inline constexpr RT_API_ATTRS bool IsAssumedSize(const CFI_cdesc_t *dv) {
   return dv->rank > 0 && dv->dim[dv->rank - 1].extent == -1;
 }
 
-static inline std::size_t MinElemLen(CFI_type_t type) {
+static inline RT_API_ATTRS std::size_t MinElemLen(CFI_type_t type) {
   auto typeParams{Fortran::runtime::TypeCode{type}.GetCategoryAndKind()};
   if (!typeParams) {
     Fortran::runtime::Terminator terminator{__FILE__, __LINE__};
@@ -38,10 +38,10 @@ static inline std::size_t MinElemLen(CFI_type_t type) {
       typeParams->first, typeParams->second);
 }
 
-static inline int VerifyEstablishParameters(CFI_cdesc_t *descriptor,
-    void *base_addr, CFI_attribute_t attribute, CFI_type_t type,
-    std::size_t elem_len, CFI_rank_t rank, const CFI_index_t extents[],
-    bool external) {
+static inline RT_API_ATTRS int VerifyEstablishParameters(
+    CFI_cdesc_t *descriptor, void *base_addr, CFI_attribute_t attribute,
+    CFI_type_t type, std::size_t elem_len, CFI_rank_t rank,
+    const CFI_index_t extents[], bool external) {
   if (attribute != CFI_attribute_other && attribute != CFI_attribute_pointer &&
       attribute != CFI_attribute_allocatable) {
     return CFI_INVALID_ATTRIBUTE;
@@ -77,9 +77,9 @@ static inline int VerifyEstablishParameters(CFI_cdesc_t *descriptor,
   return CFI_SUCCESS;
 }
 
-static inline void EstablishDescriptor(CFI_cdesc_t *descriptor, void *base_addr,
-    CFI_attribute_t attribute, CFI_type_t type, std::size_t elem_len,
-    CFI_rank_t rank, const CFI_index_t extents[]) {
+static inline RT_API_ATTRS void EstablishDescriptor(CFI_cdesc_t *descriptor,
+    void *base_addr, CFI_attribute_t attribute, CFI_type_t type,
+    std::size_t elem_len, CFI_rank_t rank, const CFI_index_t extents[]) {
   descriptor->base_addr = base_addr;
   descriptor->elem_len = elem_len;
   descriptor->version = CFI_VERSION;

flang/runtime/derived.h

@@ -11,6 +11,8 @@
 #ifndef FORTRAN_RUNTIME_DERIVED_H_
 #define FORTRAN_RUNTIME_DERIVED_H_
 
+#include "flang/Runtime/api-attrs.h"
+
 namespace Fortran::runtime::typeInfo {
 class DerivedType;
 }
@@ -21,21 +23,21 @@ class Terminator;
 
 // Perform default component initialization, allocate automatic components.
 // Returns a STAT= code (0 when all's well).
-int Initialize(const Descriptor &, const typeInfo::DerivedType &, Terminator &,
-    bool hasStat = false, const Descriptor *errMsg = nullptr);
+RT_API_ATTRS int Initialize(const Descriptor &, const typeInfo::DerivedType &,
+    Terminator &, bool hasStat = false, const Descriptor *errMsg = nullptr);
 
 // Call FINAL subroutines, if any
-void Finalize(
+RT_API_ATTRS void Finalize(
     const Descriptor &, const typeInfo::DerivedType &derived, Terminator *);
 
 // Call FINAL subroutines, deallocate allocatable & automatic components.
 // Does not deallocate the original descriptor.
-void Destroy(const Descriptor &, bool finalize, const typeInfo::DerivedType &,
-    Terminator *);
+RT_API_ATTRS void Destroy(const Descriptor &, bool finalize,
+    const typeInfo::DerivedType &, Terminator *);
 
 // Return true if the passed descriptor is for a derived type
 // entity that has a dynamic (allocatable, automatic) component.
-bool HasDynamicComponent(const Descriptor &);
+RT_API_ATTRS bool HasDynamicComponent(const Descriptor &);
 
 } // namespace Fortran::runtime
 #endif // FORTRAN_RUNTIME_DERIVED_H_