backport fix for FFTW3 threading issue

include/armadillo_bits/fft_engine_fftw3.hpp

@@ -50,7 +50,8 @@ class fft_engine_fftw3
   podarray<cx_type> Y_work;  // for storing output
   
   inline
-  ~fft_engine_fftw3()
+  void
+  finish()
     {
     arma_debug_sigprint();
     
@@ -59,6 +60,33 @@ class fft_engine_fftw3
     // fftw3::cleanup<cx_type>();  // NOTE: this also removes any wisdom acquired by FFTW3 
     }
   
+  inline
+  ~fft_engine_fftw3()
+    {
+    arma_debug_sigprint();
+    
+    #if defined(ARMA_USE_OPENMP)
+      {
+      #pragma omp critical (arma_fft_engine_fftw3)
+        {
+        (*this).finish();
+        }
+      }
+    #elif defined(ARMA_USE_STD_MUTEX)
+      {
+      std::mutex& plan_mutex = fft_engine_fftw3_aux::get_plan_mutex();
+      
+      const std::lock_guard<std::mutex> lock(plan_mutex);
+      
+      (*this).finish();
+      }
+    #else
+      {
+      (*this).finish();
+      }
+    #endif
+    }
+  
   inline
   fft_engine_fftw3(const uword in_N)
     : N         (in_N   )