Re: [PATCH, i386, Pointer Bounds Checker 32/x] Pointer Bounds Checker hooks for i386 target

[Jeff Law]

On 09/19/14 10:21, Uros Bizjak wrote:

+static tree
+ix86_make_bounds_constant (HOST_WIDE_INT lb, HOST_WIDE_INT ub)
+{
+  tree low = lb ? build_minus_one_cst (pointer_sized_int_node)
+    : build_zero_cst (pointer_sized_int_node);
+  tree high = ub ? build_zero_cst (pointer_sized_int_node)
+    : build_minus_one_cst (pointer_sized_int_node);
+
+  /* This function is supposed to be used to create zero and
+     none bounds only.  */
+  gcc_assert (lb == 0 || lb == -1);
+  gcc_assert (ub == 0 || ub == -1);
+
+  return build_complex (NULL, low, high);
Needs a comment, even more so than normal given the restrictive nature 
of the input values.  One could bikeshed on using true/false vs 0/-1 
since it appears that lb/ub are really boolean values.


+}
+
+static int
+ix86_initialize_bounds (tree var, tree lb, tree ub, tree *stmts)
+{
+  tree size_ptr = build_pointer_type (size_type_node);
+  tree lhs, modify, var_p;
+
+  ub = build1 (BIT_NOT_EXPR, size_type_node, ub);
+  var_p = build1 (CONVERT_EXPR, size_ptr,
+                 build_fold_addr_expr (var));
+
+  lhs = build1 (INDIRECT_REF, size_type_node, var_p);
+  modify = build2 (MODIFY_EXPR, TREE_TYPE (lhs), lhs, lb);
+  append_to_statement_list (modify, stmts);
+
+  lhs = build1 (INDIRECT_REF, size_type_node,
+               build2 (POINTER_PLUS_EXPR, size_ptr, var_p,
+                       TYPE_SIZE_UNIT (size_type_node)));
+  modify = build2 (MODIFY_EXPR, TREE_TYPE (lhs), lhs, ub);
+  append_to_statement_list (modify, stmts);
+
+  return 2;
+}
Are we supposed to be emitting gimple or generic here?  I don't think 
this is valid gimple as we have a CONVERT_EXPR embedded in the 
INDIRECT_REF's argument.

I think it's valid generic and does the obvious thing.  I just Ilya to 
make sure what we create does get gimplified.

jeff