compiler: Tweak Dereference

Author: FabioLuporini

devito/ir/iet/nodes.py

@@ -1047,9 +1047,8 @@ class Dereference(ExprStmt, Node):
     A node encapsulating a dereference from a `pointer` to a `pointee`.
     The following cases are supported:
 
-        * `pointer` is a PointerArray or TempFunction, and `pointee` is an Array.
-        * `pointer` is an ArrayObject representing a pointer to a C struct, and
-          `pointee` is a field in `pointer`.
+        * `pointer` is an AbstractFunction, and `pointee` is an Array.
+        * `pointer` is an AbstractObject, and `pointee` is an Array.
         * `pointer` is a Symbol with its _C_ctype deriving from ct._Pointer, and
           `pointee` is a Symbol representing the dereferenced value.
     """
@@ -1075,20 +1074,23 @@ def expr_symbols(self):
             assert issubclass(self.pointer._C_ctype, ctypes._Pointer), \
                    "Scalar dereference must have a pointer ctype"
             ret.extend([self.pointer._C_symbol, self.pointee._C_symbol])
-        elif self.pointer.is_PointerArray or self.pointer.is_TempFunction:
+        elif self.pointer.is_AbstractFunction:
             ret.extend([self.pointer.indexed, self.pointee.indexed])
             ret.extend(flatten(i.free_symbols
                                for i in self.pointee.symbolic_shape[1:]))
             ret.extend(self.pointer.free_symbols)
+        elif self.pointer.is_AbstractObject:
+            ret.extend([self.pointer, self.pointee.indexed])
+            ret.extend(flatten(i.free_symbols
+                               for i in self.pointee.symbolic_shape[1:]))
         else:
-            ret.extend([self.pointer.indexed, self.pointee._C_symbol])
+            assert False, f"Unexpected pointer type {type(self.pointer)}"
+
         return tuple(filter_ordered(ret))
 
     @property
     def defines(self):
-        if self.pointer.is_PointerArray or \
-           self.pointer.is_TempFunction or \
-           self.pointee._mem_stack:
+        if self.pointer.is_AbstractFunction or self.pointee._mem_stack:
             return (self.pointee.indexed, self.pointee)
         else:
             return (self.pointee,)

devito/ir/iet/visitors.py

@@ -506,24 +506,30 @@ def visit_PointerCast(self, o):
 
     def visit_Dereference(self, o):
         a0, a1 = o.functions
-        if a1.is_PointerArray or a1.is_TempFunction:
-            i = a1.indexed
-            cstr = self.ccode(i._C_typedata)
+        if a0.is_AbstractFunction:
+            cstr = self.ccode(a0.indexed._C_typedata)
+
+            try:
+                # Special AbstractFunctions such as PointerArray or TempFunction
+                cdim = f'[{a1.dim.name}]'
+            except AttributeError:
+                cdim = ''
+
             if o.flat is None:
                 shape = ''.join(f"[{self.ccode(i)}]" for i in a0.symbolic_shape[1:])
-                rvalue = f'({cstr} (*){shape}) {a1.name}[{a1.dim.name}]'
+                rvalue = f'({cstr} (*){shape}) {a1.name}{cdim}'
                 lvalue = c.Value(cstr, f'(*{self._restrict_keyword} {a0.name}){shape}')
             else:
-                rvalue = f'({cstr} *) {a1.name}[{a1.dim.name}]'
+                rvalue = f'({cstr} *) {a1.name}{cdim}'
                 lvalue = c.Value(cstr, f'*{self._restrict_keyword} {a0.name}')
-            if a0._data_alignment:
-                lvalue = c.AlignedAttribute(a0._data_alignment, lvalue)
+
         else:
             if a1.is_Symbol:
                 rvalue = f'*{a1.name}'
             else:
                 rvalue = f'{a1.name}->{a0._C_name}'
             lvalue = self._gen_value(a0, 0)
+
         return c.Initializer(lvalue, rvalue)
 
     def visit_Block(self, o):

examples/performance/00_overview.ipynb

@@ -701,7 +701,7 @@
       "  #pragma omp parallel num_threads(nthreads)\n",
       "  {\n",
       "    const int tid = omp_get_thread_num();\n",
-      "    float (*restrict r0)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr0[tid];\n",
+      "    float (*restrict r0)[z_size] = (float (*)[z_size]) pr0[tid];\n",
       "\n",
       "    #pragma omp for schedule(dynamic,1)\n",
       "    for (int x = x_m; x <= x_M; x += 1)\n",
@@ -836,7 +836,7 @@
       "  #pragma omp parallel num_threads(nthreads)\n",
       "  {\n",
       "    const int tid = omp_get_thread_num();\n",
-      "    float (*restrict r1)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr1[tid];\n",
+      "    float (*restrict r1)[z_size] = (float (*)[z_size]) pr1[tid];\n",
       "\n",
       "    #pragma omp for schedule(dynamic,1)\n",
       "    for (int x = x_m; x <= x_M; x += 1)\n",
@@ -968,7 +968,7 @@
       "  #pragma omp parallel num_threads(nthreads)\n",
       "  {\n",
       "    const int tid = omp_get_thread_num();\n",
-      "    float (*restrict r0)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr0[tid];\n",
+      "    float (*restrict r0)[z_size] = (float (*)[z_size]) pr0[tid];\n",
       "\n",
       "    #pragma omp for schedule(dynamic,1)\n",
       "    for (int x = x_m; x <= x_M; x += 1)\n",
@@ -1236,7 +1236,7 @@
       "    #pragma omp parallel num_threads(nthreads)\n",
       "    {\n",
       "      const int tid = omp_get_thread_num();\n",
-      "      float (*restrict r2)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr2[tid];\n",
+      "      float (*restrict r2)[z_size] = (float (*)[z_size]) pr2[tid];\n",
       "\n",
       "      #pragma omp for collapse(2) schedule(dynamic,1)\n",
       "      for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)\n",
@@ -1348,7 +1348,7 @@
       "  #pragma omp parallel num_threads(nthreads)\n",
       "  {\n",
       "    const int tid = omp_get_thread_num();\n",
-      "    float (*restrict r2)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr2[tid];\n",
+      "    float (*restrict r2)[z_size] = (float (*)[z_size]) pr2[tid];\n",
       "\n",
       "    #pragma omp for collapse(2) schedule(dynamic,1)\n",
       "    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)\n",
@@ -1527,7 +1527,7 @@
       "    #pragma omp parallel num_threads(nthreads)\n",
       "    {\n",
       "      const int tid = omp_get_thread_num();\n",
-      "      float (*restrict r2)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) pr2[tid];\n",
+      "      float (*restrict r2)[z_size] = (float (*)[z_size]) pr2[tid];\n",
       "\n",
       "      #pragma omp for schedule(dynamic,1)\n",
       "      for (int x = x_m; x <= x_M; x += 1)\n",