Merge branch 'more_clang_fixes' into 'master'

address more problems with clang (and follow its optimization hints)

See merge request !35

pocketfft_hdronly.h

@@ -110,8 +110,9 @@ constexpr bool FORWARD  = true,
 #if defined(__INTEL_COMPILER)
 // do nothing. This is necessary because this compiler also sets __GNUC__.
 #elif defined(__clang__)
-#ifdef __APPLE__
-#  if (__clang_major__ > 9) || (__clang_major__ == 9 && __clang_minor__ >= 3)
+// AppleClang has their own version numbering
+#ifdef __apple_build_version__
+#  if (__clang_major__ > 9) || (__clang_major__ == 9 && __clang_minor__ >= 1)
 #     undef POCKETFFT_NO_VECTORS
 #  endif
 #elif __clang_major__ >= 5
@@ -216,8 +217,6 @@ template<typename T> struct cmplx {
   cmplx(T r_, T i_) : r(r_), i(i_) {}
   void Set(T r_, T i_) { r=r_; i=i_; }
   void Set(T r_) { r=r_; i=T(0); }
-  void Split(T &r_, T &i_) const { r_=r; i_=i; }
-  void SplitConj(T &r_, T &i_) const { r_=r; i_=-i; }
   cmplx &operator+= (const cmplx &other)
     { r+=other.r; i+=other.i; return *this; }
   template<typename T2>cmplx &operator*= (T2 other)
@@ -3203,11 +3202,17 @@ template<typename T> POCKETFFT_NOINLINE void general_c2r(
           if (forward)
             for (; i<len-1; i+=2, ++ii)
               for (size_t j=0; j<vlen; ++j)
-                in[it.iofs(j,ii)].SplitConj(tdatav[i][j], tdatav[i+1][j]);
+                {
+                tdatav[i  ][j] =  in[it.iofs(j,ii)].r;
+                tdatav[i+1][j] = -in[it.iofs(j,ii)].i;
+                }
           else
             for (; i<len-1; i+=2, ++ii)
               for (size_t j=0; j<vlen; ++j)
-                in[it.iofs(j,ii)].Split(tdatav[i][j], tdatav[i+1][j]);
+                {
+                tdatav[i  ][j] = in[it.iofs(j,ii)].r;
+                tdatav[i+1][j] = in[it.iofs(j,ii)].i;
+                }
           if (i<len)
             for (size_t j=0; j<vlen; ++j)
               tdatav[i][j] = in[it.iofs(j,ii)].r;
@@ -3225,10 +3230,16 @@ template<typename T> POCKETFFT_NOINLINE void general_c2r(
         size_t i=1, ii=1;
         if (forward)
           for (; i<len-1; i+=2, ++ii)
-            in[it.iofs(ii)].SplitConj(tdata[i], tdata[i+1]);
+            {
+            tdata[i  ] =  in[it.iofs(ii)].r;
+            tdata[i+1] = -in[it.iofs(ii)].i;
+            }
         else
           for (; i<len-1; i+=2, ++ii)
-            in[it.iofs(ii)].Split(tdata[i], tdata[i+1]);
+            {
+            tdata[i  ] = in[it.iofs(ii)].r;
+            tdata[i+1] = in[it.iofs(ii)].i;
+            }
         if (i<len)
           tdata[i] = in[it.iofs(ii)].r;
         }

pypocketfft.cc

@@ -130,7 +130,7 @@ template<typename T> py::array c2c_internal(const py::array &in,
   T fct = norm_fct<T>(inorm, dims, axes);
   pocketfft::c2c(dims, s_in, s_out, axes, forward, d_in, d_out, fct, nthreads);
   }
-  return res;
+  return move(res);
   }
 
 template<typename T> py::array c2c_sym_internal(const py::array &in,
@@ -159,7 +159,7 @@ template<typename T> py::array c2c_sym_internal(const py::array &in,
     iter.advance();
     }
   }
-  return res;
+  return move(res);
   }
 
 py::array c2c(const py::array &a, const py::object &axes_, bool forward,