carray 3.0.1 → 3.0.2

data/ext/carray_kernels_reduce_extreme.c

@@ -137,6 +137,7 @@ min_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (int8_t) (__li_buf_acc[__j]);
             }
@@ -216,6 +217,7 @@ min_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (int8_t) (__li_buf_acc[__j]);
             }
@@ -341,6 +343,7 @@ min_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int8_t) (__li_buf_acc[__j]);
               }
@@ -501,6 +504,7 @@ min_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int8_t) (__li_buf_acc[__j]);
               }
@@ -656,6 +660,7 @@ min_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint8_t) (__li_buf_acc[__j]);
             }
@@ -735,6 +740,7 @@ min_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint8_t) (__li_buf_acc[__j]);
             }
@@ -860,6 +866,7 @@ min_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint8_t) (__li_buf_acc[__j]);
               }
@@ -1020,6 +1027,7 @@ min_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint8_t) (__li_buf_acc[__j]);
               }
@@ -1175,6 +1183,7 @@ min_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (int16_t) (__li_buf_acc[__j]);
             }
@@ -1254,6 +1263,7 @@ min_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (int16_t) (__li_buf_acc[__j]);
             }
@@ -1379,6 +1389,7 @@ min_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int16_t) (__li_buf_acc[__j]);
               }
@@ -1539,6 +1550,7 @@ min_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int16_t) (__li_buf_acc[__j]);
               }
@@ -1694,6 +1706,7 @@ min_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint16_t) (__li_buf_acc[__j]);
             }
@@ -1773,6 +1786,7 @@ min_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint16_t) (__li_buf_acc[__j]);
             }
@@ -1898,6 +1912,7 @@ min_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint16_t) (__li_buf_acc[__j]);
               }
@@ -2058,6 +2073,7 @@ min_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint16_t) (__li_buf_acc[__j]);
               }
@@ -2213,6 +2229,7 @@ min_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (int32_t) (__li_buf_acc[__j]);
             }
@@ -2292,6 +2309,7 @@ min_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (int32_t) (__li_buf_acc[__j]);
             }
@@ -2417,6 +2435,7 @@ min_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int32_t) (__li_buf_acc[__j]);
               }
@@ -2577,6 +2596,7 @@ min_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int32_t) (__li_buf_acc[__j]);
               }
@@ -2732,6 +2752,7 @@ min_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint32_t) (__li_buf_acc[__j]);
             }
@@ -2811,6 +2832,7 @@ min_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint32_t) (__li_buf_acc[__j]);
             }
@@ -2936,6 +2958,7 @@ min_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint32_t) (__li_buf_acc[__j]);
               }
@@ -3096,6 +3119,7 @@ min_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint32_t) (__li_buf_acc[__j]);
               }
@@ -3251,6 +3275,7 @@ min_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (int64_t) (__li_buf_acc[__j]);
             }
@@ -3330,6 +3355,7 @@ min_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (int64_t) (__li_buf_acc[__j]);
             }
@@ -3455,6 +3481,7 @@ min_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int64_t) (__li_buf_acc[__j]);
               }
@@ -3615,6 +3642,7 @@ min_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int64_t) (__li_buf_acc[__j]);
               }
@@ -3770,6 +3798,7 @@ min_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
             }
@@ -3849,6 +3878,7 @@ min_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
             }
@@ -3974,6 +4004,7 @@ min_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
               }
@@ -4134,6 +4165,7 @@ min_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
               }
@@ -4217,6 +4249,7 @@ min_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
       ca_size_t __outer_off = 0;
       float acc;
       acc = (INFINITY);
+      int64_t __anf_seen = 0;
       ca_size_t masked_cnt = 0;  /* unused on streaming (no mask) */
       (void) masked_cnt;
       ca_lazy_arena_enter();
@@ -4231,11 +4264,13 @@ min_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
         for ( __i = 0; __i < __n; __i++ ) {
           float v = __chunk[__i];
           acc = (v < acc) ? v : acc;
+          __anf_seen |= (v == v);
         }
         __outer_off += __r;
       }
       ca_lazy_arena_release(__chunk);
       ca_lazy_arena_exit();
+      if ( ! __anf_seen && ca->elements > 0 ) acc = (float) NAN;
       return rb_float_new((float) (acc));
     }
     /* L.1 / L.7: loop-interchange fast path with inner-tiling
@@ -4289,6 +4324,25 @@ min_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+{
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (float) (INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const float *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (float) NAN;
+    }
+  }
+}
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (float) (__li_buf_acc[__j]);
             }
@@ -4368,6 +4422,25 @@ min_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+{
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (float) (INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const float *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (float) NAN;
+    }
+  }
+}
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (float) (__li_buf_acc[__j]);
             }
@@ -4493,6 +4566,25 @@ min_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    {
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (float) (INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const float *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (float) NAN;
+    }
+  }
+}
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (float) (__li_buf_acc[__j]);
               }
@@ -4653,6 +4745,25 @@ min_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    {
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (float) (INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const float *__li_row = __li_plane + __li_i * __li_M_stride + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (float) NAN;
+    }
+  }
+}
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (float) (__li_buf_acc[__j]);
               }
@@ -4690,6 +4801,16 @@ ca_size_t   out_i = 0;
 #define __min_f32_EXPR(__x) ((float)(__x))
       CA_SLAB_REDUCE_MIN8_EX(float, float, st, p, m, acc, INFINITY, __min_f32_EXPR, masked_cnt);
 #undef __min_f32_EXPR
+      /* all-NaN fix-up (all_nan_result: :nan). */
+      if ( (acc == (INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(float, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          acc = (float) NAN;
+        }
+      }
       if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
@@ -4736,6 +4857,7 @@ min_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
       ca_size_t __outer_off = 0;
       double acc;
       acc = (INFINITY);
+      int64_t __anf_seen = 0;
       ca_size_t masked_cnt = 0;  /* unused on streaming (no mask) */
       (void) masked_cnt;
       ca_lazy_arena_enter();
@@ -4750,11 +4872,13 @@ min_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
         for ( __i = 0; __i < __n; __i++ ) {
           double v = __chunk[__i];
           acc = (v < acc) ? v : acc;
+          __anf_seen |= (v == v);
         }
         __outer_off += __r;
       }
       ca_lazy_arena_release(__chunk);
       ca_lazy_arena_exit();
+      if ( ! __anf_seen && ca->elements > 0 ) acc = (double) NAN;
       return rb_float_new((double) (acc));
     }
     /* L.1 / L.7: loop-interchange fast path with inner-tiling
@@ -4808,6 +4932,25 @@ min_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+{
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (double) (INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const double *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (double) NAN;
+    }
+  }
+}
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (double) (__li_buf_acc[__j]);
             }
@@ -4887,6 +5030,25 @@ min_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+{
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (double) (INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const double *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (double) NAN;
+    }
+  }
+}
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (double) (__li_buf_acc[__j]);
             }
@@ -5012,6 +5174,25 @@ min_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    {
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (double) (INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const double *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (double) NAN;
+    }
+  }
+}
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (double) (__li_buf_acc[__j]);
               }
@@ -5172,6 +5353,25 @@ min_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    {
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (double) (INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const double *__li_row = __li_plane + __li_i * __li_M_stride + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (double) NAN;
+    }
+  }
+}
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (double) (__li_buf_acc[__j]);
               }
@@ -5209,6 +5409,16 @@ ca_size_t   out_i = 0;
 #define __min_f64_EXPR(__x) ((double)(__x))
       CA_SLAB_REDUCE_MIN8_EX(double, double, st, p, m, acc, INFINITY, __min_f64_EXPR, masked_cnt);
 #undef __min_f64_EXPR
+      /* all-NaN fix-up (all_nan_result: :nan). */
+      if ( (acc == (INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(double, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          acc = (double) NAN;
+        }
+      }
       if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
@@ -5327,6 +5537,7 @@ min_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
             }
@@ -5406,6 +5617,7 @@ min_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
             }
@@ -5531,6 +5743,7 @@ min_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
               }
@@ -5691,6 +5904,7 @@ min_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
               }
@@ -5846,6 +6060,7 @@ min_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (VALUE) (__li_buf_acc[__j]);
             }
@@ -5925,6 +6140,7 @@ min_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (VALUE) (__li_buf_acc[__j]);
             }
@@ -6050,6 +6266,7 @@ min_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (VALUE) (__li_buf_acc[__j]);
               }
@@ -6210,6 +6427,7 @@ min_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (VALUE) (__li_buf_acc[__j]);
               }
@@ -6573,6 +6791,7 @@ max_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (int8_t) (__li_buf_acc[__j]);
             }
@@ -6652,6 +6871,7 @@ max_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (int8_t) (__li_buf_acc[__j]);
             }
@@ -6777,6 +6997,7 @@ max_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int8_t) (__li_buf_acc[__j]);
               }
@@ -6937,6 +7158,7 @@ max_ki_native_i8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int8_t) (__li_buf_acc[__j]);
               }
@@ -7092,6 +7314,7 @@ max_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint8_t) (__li_buf_acc[__j]);
             }
@@ -7171,6 +7394,7 @@ max_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint8_t) (__li_buf_acc[__j]);
             }
@@ -7296,6 +7520,7 @@ max_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint8_t) (__li_buf_acc[__j]);
               }
@@ -7456,6 +7681,7 @@ max_ki_native_u8 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int k
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint8_t) (__li_buf_acc[__j]);
               }
@@ -7611,6 +7837,7 @@ max_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (int16_t) (__li_buf_acc[__j]);
             }
@@ -7690,6 +7917,7 @@ max_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (int16_t) (__li_buf_acc[__j]);
             }
@@ -7815,6 +8043,7 @@ max_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int16_t) (__li_buf_acc[__j]);
               }
@@ -7975,6 +8204,7 @@ max_ki_native_i16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int16_t) (__li_buf_acc[__j]);
               }
@@ -8130,6 +8360,7 @@ max_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint16_t) (__li_buf_acc[__j]);
             }
@@ -8209,6 +8440,7 @@ max_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint16_t) (__li_buf_acc[__j]);
             }
@@ -8334,6 +8566,7 @@ max_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint16_t) (__li_buf_acc[__j]);
               }
@@ -8494,6 +8727,7 @@ max_ki_native_u16 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint16_t) (__li_buf_acc[__j]);
               }
@@ -8649,6 +8883,7 @@ max_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (int32_t) (__li_buf_acc[__j]);
             }
@@ -8728,6 +8963,7 @@ max_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (int32_t) (__li_buf_acc[__j]);
             }
@@ -8853,6 +9089,7 @@ max_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int32_t) (__li_buf_acc[__j]);
               }
@@ -9013,6 +9250,7 @@ max_ki_native_i32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int32_t) (__li_buf_acc[__j]);
               }
@@ -9168,6 +9406,7 @@ max_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint32_t) (__li_buf_acc[__j]);
             }
@@ -9247,6 +9486,7 @@ max_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint32_t) (__li_buf_acc[__j]);
             }
@@ -9372,6 +9612,7 @@ max_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint32_t) (__li_buf_acc[__j]);
               }
@@ -9532,6 +9773,7 @@ max_ki_native_u32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint32_t) (__li_buf_acc[__j]);
               }
@@ -9687,6 +9929,7 @@ max_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (int64_t) (__li_buf_acc[__j]);
             }
@@ -9766,6 +10009,7 @@ max_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (int64_t) (__li_buf_acc[__j]);
             }
@@ -9891,6 +10135,7 @@ max_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int64_t) (__li_buf_acc[__j]);
               }
@@ -10051,6 +10296,7 @@ max_ki_native_i64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (int64_t) (__li_buf_acc[__j]);
               }
@@ -10206,6 +10452,7 @@ max_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
             }
@@ -10285,6 +10532,7 @@ max_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
             }
@@ -10410,6 +10658,7 @@ max_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
               }
@@ -10570,6 +10819,7 @@ max_ki_native_u64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
               }
@@ -10653,6 +10903,7 @@ max_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
       ca_size_t __outer_off = 0;
       float acc;
       acc = (-INFINITY);
+      int64_t __anf_seen = 0;
       ca_size_t masked_cnt = 0;  /* unused on streaming (no mask) */
       (void) masked_cnt;
       ca_lazy_arena_enter();
@@ -10667,11 +10918,13 @@ max_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
         for ( __i = 0; __i < __n; __i++ ) {
           float v = __chunk[__i];
           acc = (v > acc) ? v : acc;
+          __anf_seen |= (v == v);
         }
         __outer_off += __r;
       }
       ca_lazy_arena_release(__chunk);
       ca_lazy_arena_exit();
+      if ( ! __anf_seen && ca->elements > 0 ) acc = (float) NAN;
       return rb_float_new((float) (acc));
     }
     /* L.1 / L.7: loop-interchange fast path with inner-tiling
@@ -10725,6 +10978,25 @@ max_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+{
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (float) (-INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const float *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (float) NAN;
+    }
+  }
+}
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (float) (__li_buf_acc[__j]);
             }
@@ -10804,6 +11076,25 @@ max_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+{
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (float) (-INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const float *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (float) NAN;
+    }
+  }
+}
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (float) (__li_buf_acc[__j]);
             }
@@ -10929,6 +11220,25 @@ max_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    {
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (float) (-INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const float *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (float) NAN;
+    }
+  }
+}
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (float) (__li_buf_acc[__j]);
               }
@@ -11089,6 +11399,25 @@ max_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    {
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (float) (-INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const float *__li_row = __li_plane + __li_i * __li_M_stride + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (float) NAN;
+    }
+  }
+}
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (float) (__li_buf_acc[__j]);
               }
@@ -11126,6 +11455,16 @@ ca_size_t   out_i = 0;
 #define __max_f32_EXPR(__x) ((float)(__x))
       CA_SLAB_REDUCE_MAX8_EX(float, float, st, p, m, acc, -INFINITY, __max_f32_EXPR, masked_cnt);
 #undef __max_f32_EXPR
+      /* all-NaN fix-up (all_nan_result: :nan). */
+      if ( (acc == (-INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(float, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          acc = (float) NAN;
+        }
+      }
       if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
@@ -11172,6 +11511,7 @@ max_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
       ca_size_t __outer_off = 0;
       double acc;
       acc = (-INFINITY);
+      int64_t __anf_seen = 0;
       ca_size_t masked_cnt = 0;  /* unused on streaming (no mask) */
       (void) masked_cnt;
       ca_lazy_arena_enter();
@@ -11186,11 +11526,13 @@ max_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
         for ( __i = 0; __i < __n; __i++ ) {
           double v = __chunk[__i];
           acc = (v > acc) ? v : acc;
+          __anf_seen |= (v == v);
         }
         __outer_off += __r;
       }
       ca_lazy_arena_release(__chunk);
       ca_lazy_arena_exit();
+      if ( ! __anf_seen && ca->elements > 0 ) acc = (double) NAN;
       return rb_float_new((double) (acc));
     }
     /* L.1 / L.7: loop-interchange fast path with inner-tiling
@@ -11244,6 +11586,25 @@ max_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+{
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (double) (-INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const double *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (double) NAN;
+    }
+  }
+}
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (double) (__li_buf_acc[__j]);
             }
@@ -11323,6 +11684,25 @@ max_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+{
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (double) (-INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const double *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (double) NAN;
+    }
+  }
+}
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (double) (__li_buf_acc[__j]);
             }
@@ -11448,6 +11828,25 @@ max_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    {
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (double) (-INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const double *__li_row = __li_plane + __li_i * __li_INNER + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (double) NAN;
+    }
+  }
+}
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (double) (__li_buf_acc[__j]);
               }
@@ -11608,6 +12007,25 @@ max_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    {
+  int __li_susp = 0;
+  for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+    if ( __li_buf_acc[__j] == (double) (-INFINITY) ) { __li_susp = 1; break; }
+  }
+  if ( __li_susp ) {
+    int8_t __li_buf_seen[512];
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) __li_buf_seen[__j] = 0;
+    for ( ca_size_t __li_i = 0; __li_i < __li_M; __li_i++ ) {
+      const double *__li_row = __li_plane + __li_i * __li_M_stride + __li_tile;
+      for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+        __li_buf_seen[__j] |= (__li_row[__j] == __li_row[__j]);
+      }
+    }
+    for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
+      if ( ! __li_buf_seen[__j] ) __li_buf_acc[__j] = (double) NAN;
+    }
+  }
+}
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (double) (__li_buf_acc[__j]);
               }
@@ -11645,6 +12063,16 @@ ca_size_t   out_i = 0;
 #define __max_f64_EXPR(__x) ((double)(__x))
       CA_SLAB_REDUCE_MAX8_EX(double, double, st, p, m, acc, -INFINITY, __max_f64_EXPR, masked_cnt);
 #undef __max_f64_EXPR
+      /* all-NaN fix-up (all_nan_result: :nan). */
+      if ( (acc == (-INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(double, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          acc = (double) NAN;
+        }
+      }
       if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
@@ -11763,6 +12191,7 @@ max_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
             }
@@ -11842,6 +12271,7 @@ max_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
             }
@@ -11967,6 +12397,7 @@ max_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
               }
@@ -12127,6 +12558,7 @@ max_ki_native_bool (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, int
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (uint64_t) (__li_buf_acc[__j]);
               }
@@ -12282,6 +12714,7 @@ max_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               op[__li_o * __li_INNER + __li_tile + __j] = (VALUE) (__li_buf_acc[__j]);
             }
@@ -12361,6 +12794,7 @@ max_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                 (void) v;
               }
             }
+
             for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
               __li_op_k[__li_o * __li_INNER + __li_tile + __j] = (VALUE) (__li_buf_acc[__j]);
             }
@@ -12486,6 +12920,7 @@ max_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (VALUE) (__li_buf_acc[__j]);
               }
@@ -12646,6 +13081,7 @@ max_ki_native_object (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
                   (void) v;
                 }
               }
+    
               for ( ca_size_t __j = 0; __j < __li_tile_len; __j++ ) {
                 __li_op_k[__li_tile + __j] = (VALUE) (__li_buf_acc[__j]);
               }
@@ -13377,6 +13813,17 @@ minmax_ki_native_f32 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       float hi = -INFINITY;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(float, st, p, m, lo, INFINITY, if (v < lo) lo = v; if (v > hi) hi = v;, masked_cnt);
+      /* all-NaN fix-up (all_nan_result: :nan). */
+      if ( (lo == (INFINITY) && hi == (-INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(float, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          lo = (float) NAN;
+          hi = (float) NAN;
+        }
+      }
       if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
         if ( ! op_mask_a ) {
           ca_create_mask(co_a);
@@ -13432,6 +13879,17 @@ minmax_ki_native_f64 (VALUE self, CArray *ca, int8_t *slab_axes, int8_t naxes, i
       double hi = -INFINITY;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(double, st, p, m, lo, INFINITY, if (v < lo) lo = v; if (v > hi) hi = v;, masked_cnt);
+      /* all-NaN fix-up (all_nan_result: :nan). */
+      if ( (lo == (INFINITY) && hi == (-INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(double, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          lo = (double) NAN;
+          hi = (double) NAN;
+        }
+      }
       if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
         if ( ! op_mask_a ) {
           ca_create_mask(co_a);
@@ -14021,7 +14479,18 @@ ca_size_t   out_i = 0;
       int64_t best_i = 0;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(float, st, p, m, best_v, INFINITY, if (v < best_v) { best_v = v; best_i = idx; }, masked_cnt);
-      if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
+      int __anf_all_nan = 0;
+      /* all-NaN fix-up (all_nan_result: :undef). */
+      if ( (best_v == (INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(float, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          __anf_all_nan = 1;
+        }
+      }
+      if ( ((min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count)) || __anf_all_nan ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
           op_mask = (boolean8_t *) co->mask->ptr;
@@ -14063,7 +14532,18 @@ ca_size_t   out_i = 0;
       int64_t best_i = 0;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(double, st, p, m, best_v, INFINITY, if (v < best_v) { best_v = v; best_i = idx; }, masked_cnt);
-      if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
+      int __anf_all_nan = 0;
+      /* all-NaN fix-up (all_nan_result: :undef). */
+      if ( (best_v == (INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(double, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          __anf_all_nan = 1;
+        }
+      }
+      if ( ((min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count)) || __anf_all_nan ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
           op_mask = (boolean8_t *) co->mask->ptr;
@@ -14708,7 +15188,18 @@ ca_size_t   out_i = 0;
       int64_t best_i = 0;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(float, st, p, m, best_v, -INFINITY, if (v > best_v) { best_v = v; best_i = idx; }, masked_cnt);
-      if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
+      int __anf_all_nan = 0;
+      /* all-NaN fix-up (all_nan_result: :undef). */
+      if ( (best_v == (-INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(float, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          __anf_all_nan = 1;
+        }
+      }
+      if ( ((min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count)) || __anf_all_nan ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
           op_mask = (boolean8_t *) co->mask->ptr;
@@ -14750,7 +15241,18 @@ ca_size_t   out_i = 0;
       int64_t best_i = 0;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(double, st, p, m, best_v, -INFINITY, if (v > best_v) { best_v = v; best_i = idx; }, masked_cnt);
-      if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
+      int __anf_all_nan = 0;
+      /* all-NaN fix-up (all_nan_result: :undef). */
+      if ( (best_v == (-INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(double, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          __anf_all_nan = 1;
+        }
+      }
+      if ( ((min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count)) || __anf_all_nan ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
           op_mask = (boolean8_t *) co->mask->ptr;
@@ -15641,7 +16143,18 @@ ca_size_t   out_i = 0;
       int64_t best_i = 0;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(float, st, p, m, best_v, INFINITY, if (v < best_v) { best_v = v; best_i = idx; }, masked_cnt);
-      if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
+      int __anf_all_nan = 0;
+      /* all-NaN fix-up (all_nan_result: :undef). */
+      if ( (best_v == (INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(float, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          __anf_all_nan = 1;
+        }
+      }
+      if ( ((min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count)) || __anf_all_nan ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
           op_mask = (boolean8_t *) co->mask->ptr;
@@ -15711,7 +16224,18 @@ ca_size_t   out_i = 0;
       int64_t best_i = 0;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(double, st, p, m, best_v, INFINITY, if (v < best_v) { best_v = v; best_i = idx; }, masked_cnt);
-      if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
+      int __anf_all_nan = 0;
+      /* all-NaN fix-up (all_nan_result: :undef). */
+      if ( (best_v == (INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(double, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          __anf_all_nan = 1;
+        }
+      }
+      if ( ((min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count)) || __anf_all_nan ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
           op_mask = (boolean8_t *) co->mask->ptr;
@@ -16588,7 +17112,18 @@ ca_size_t   out_i = 0;
       int64_t best_i = 0;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(float, st, p, m, best_v, -INFINITY, if (v > best_v) { best_v = v; best_i = idx; }, masked_cnt);
-      if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
+      int __anf_all_nan = 0;
+      /* all-NaN fix-up (all_nan_result: :undef). */
+      if ( (best_v == (-INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(float, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          __anf_all_nan = 1;
+        }
+      }
+      if ( ((min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count)) || __anf_all_nan ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
           op_mask = (boolean8_t *) co->mask->ptr;
@@ -16658,7 +17193,18 @@ ca_size_t   out_i = 0;
       int64_t best_i = 0;
       ca_size_t masked_cnt = 0;
       CA_SLAB_REDUCE_T_EX(double, st, p, m, best_v, -INFINITY, if (v > best_v) { best_v = v; best_i = idx; }, masked_cnt);
-      if ( (min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count) ) {
+      int __anf_all_nan = 0;
+      /* all-NaN fix-up (all_nan_result: :undef). */
+      if ( (best_v == (-INFINITY)) && masked_cnt < st.slab_elements ) {
+        int64_t   __anf_seen = 0;
+        ca_size_t __anf_mc   = 0;
+        CA_SLAB_REDUCE_T_EX(double, st, p, m, __anf_seen, 0, __anf_seen |= (v == v), __anf_mc);
+        (void) __anf_mc;
+        if ( ! __anf_seen ) {
+          __anf_all_nan = 1;
+        }
+      }
+      if ( ((min_count < 0 ? masked_cnt == st.slab_elements : st.slab_elements - masked_cnt < min_count)) || __anf_all_nan ) {
         if ( ! op_mask ) {
           ca_create_mask(co);
           op_mask = (boolean8_t *) co->mask->ptr;