Merge branch 'avoid_globals' into 'master'

move global variables into static functions

See merge request !33

pocketfft_hdronly.h

@@ -497,9 +497,8 @@ namespace threading {
 
 #ifdef POCKETFFT_NO_MULTITHREADING
 
-constexpr size_t thread_id = 0;
-constexpr size_t num_threads = 1;
-constexpr size_t max_threads = 1;
+constexpr inline size_t thread_id() { return 0; }
+constexpr inline size_t num_threads() { return 1; }
 
 template <typename Func>
 void thread_map(size_t /* nthreads */, Func f)
@@ -507,8 +506,16 @@ void thread_map(size_t /* nthreads */, Func f)
 
 #else
 
-thread_local size_t thread_id = 0;
-thread_local size_t num_threads = 1;
+inline size_t &thread_id()
+  {
+  static thread_local size_t thread_id_=0;
+  return thread_id_;
+  }
+inline size_t &num_threads()
+  {
+  static thread_local size_t num_threads_=1;
+  return num_threads_;
+  }
 static const size_t max_threads = std::max(1u, std::thread::hardware_concurrency());
 
 class latch
@@ -674,8 +681,8 @@ void thread_map(size_t nthreads, Func f)
     {
     pool.submit(
       [&f, &counter, &ex, &ex_mut, i, nthreads] {
-      thread_id = i;
-      num_threads = nthreads;
+      thread_id() = i;
+      num_threads() = nthreads;
       try { f(); }
       catch (...)
         {
@@ -2765,10 +2772,10 @@ template<size_t N> class multi_iter
         str_i(iarr.stride(idim_)), p_oi(0), str_o(oarr.stride(idim_)),
         idim(idim_), rem(iarr.size()/iarr.shape(idim))
       {
-      auto nshares = threading::num_threads;
+      auto nshares = threading::num_threads();
       if (nshares==1) return;
       if (nshares==0) throw std::runtime_error("can't run with zero threads");
-      auto myshare = threading::thread_id;
+      auto myshare = threading::thread_id();
       if (myshare>=nshares) throw std::runtime_error("impossible share requested");
       size_t nbase = rem/nshares;
       size_t additional = rem%nshares;

pypocketfft.cc

@@ -374,7 +374,7 @@ py::array genuine_hartley(const py::array &in, const py::object &axes_,
   }
 
 // Export good_size in raw C-API to reduce overhead (~4x faster)
-PyObject * good_size(PyObject * self, PyObject * args)
+PyObject * good_size(PyObject * /*self*/, PyObject * args)
   {
   Py_ssize_t n_ = -1;
   int real = false;