Correct handling of failed lazy_ptr constructors.

Contructor function must not be attempted twice even if the construction
attempt returns nullptr.

Change-Id: I75353e5e511769a96e4332f7f60887f6559c1cd5

runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

@@ -137,7 +137,7 @@ GpuAgent::GpuAgent(HSAuint32 node, const HsaNodeProperties& node_props)
 
 GpuAgent::~GpuAgent() {
   for (auto& blit : blits_) {
-    if (blit.created()) {
+    if (!blit.empty()) {
       hsa_status_t status = blit->Destroy(*this);
       assert(status == HSA_STATUS_SUCCESS);
     }
@@ -680,7 +680,7 @@ hsa_status_t GpuAgent::DmaFill(void* ptr, uint32_t value, size_t count) {
 
 hsa_status_t GpuAgent::EnableDmaProfiling(bool enable) {
   for (auto& blit : blits_) {
-    if (blit.created()) {
+    if (!blit.empty()) {
       const hsa_status_t stat = blit->EnableProfiling(enable);
       if (stat != HSA_STATUS_SUCCESS) {
         return stat;

runtime/hsa-runtime/core/util/lazy_ptr.h

@@ -87,6 +87,7 @@ template <typename T> class lazy_ptr {
 
   const std::unique_ptr<T>& operator->() const {
     make(true);
+    assert(obj != nullptr && "Null dereference through lazy_ptr.");
     return obj;
   }
 
@@ -107,7 +108,17 @@ template <typename T> class lazy_ptr {
   void touch() const { make(false); }
 
   // Tells if the lazy object has been constructed or not.
-  bool created() const { return obj != nullptr; }
+  // Construction may fail silently (return nullptr).
+  bool created() const {
+    std::atomic_thread_fence(std::memory_order_acquire);
+    return func == nullptr;
+  }
+
+  // Tells if the lazy object exists or not.
+  bool empty() const {
+    std::atomic_thread_fence(std::memory_order_acquire);
+    return obj == nullptr;
+  }
 
  private:
   mutable std::unique_ptr<T> obj;
@@ -122,16 +133,15 @@ template <typename T> class lazy_ptr {
       return;
     }
     MAKE_SCOPE_GUARD([&]() { lock.Release(); });
-    if (obj != nullptr) return;
+    if (func == nullptr) return;
     T* ptr = func();
-    std::atomic_thread_fence(std::memory_order_release);
     obj.reset(ptr);
+    std::atomic_thread_fence(std::memory_order_release);
     func = nullptr;
   }
 
   __forceinline void make(bool block) const {
-    std::atomic_thread_fence(std::memory_order_acquire);
-    if (obj == nullptr) {
+    if (!created()) {
       make_body(block);
     }
   }