RFC/RFH: My new library call rewriting optimization pass is ICEing

[David Daney <ddaney at avtrex dot com>]

The essence of the pass is that it transforms

new java.lang.StringBuffer();

into

new gnu.gcj.runtime.StringBuffer();

My first attempt is to insert my pass immediatly after pass_build_ssa. 
I am trying to do the first part of transform with this code:

.
.
.
tree stmt = bsi_stmt (bsi);
tc = TREE_CODE(stmt);
if (MODIFY_EXPR == tc) {
  tree rhs = TREE_OPERAND(stmt, 1);
  tc = TREE_CODE(rhs);
  if (CALL_EXPR == tc) {
    tree fn = TREE_OPERAND(rhs, 0);
    tc = TREE_CODE(fn);
    if (ADDR_EXPR == tc) {
      tree ddd = TREE_OPERAND(fn, 0);
      tc = TREE_CODE(ddd);
      if (FUNCTION_DECL == tc) {
        tree id = DECL_NAME(ddd);
        tc = TREE_CODE(id);
        const char *name = IDENTIFIER_POINTER(id);
        if (name && 0 == strcmp(name, "_Jv_AllocObjectNoFinalizer")) {
          tree args = TREE_OPERAND(rhs, 1);
          tc = TREE_CODE(args);
          tree arg1 = TREE_VALUE(args);
          tc = TREE_CODE(arg1);
          tree tch = TREE_CHAIN(args);
          if (!tch && ADDR_EXPR == tc) {
            tree jj = TREE_OPERAND(arg1, 0);
            tc = TREE_CODE(jj);
            if (VAR_DECL == tc) {
              tree jjid = DECL_NAME(jj);
              const char *jjname = IDENTIFIER_POINTER(jjid);
              if (jjname && 0 == strcmp(jjname,
                                        "java.lang.StringBuffer.class")) {
                tree jjtype = TREE_TYPE(jj);
                tree id = get_identifier ("gnu.gcj.runtime.StringBuffer");
                tree cls = resolve_and_layout(id, NULL);
                tree overridden = build_class_ref(TREE_TYPE(cls));
                java_mark_addressable (TREE_OPERAND (overridden, 0));
                TREE_VALUE(args) = overridden;
              }
            }
          }
        }
      }
    }
  }
}
.
.
.

The tree dumps show that the call to _Jv_AllocObjectNoFinalizer is being 
 transformed in the desired manner, but I am getting an ICE in 
compute_may_aliases().  Because the gnu.gcj.runtime.StringBuffer.class 
var has not been properly initialized (missing var_ann).  This 
annotation is added by one of the passes previous to pass_build_ssa.

I am starting to think that even if I figure out how to add the proper 
annotations to my new var, that this will be very brittle and there will 
be an unending series of similar problems.

Q: Is this type of rewriting possible or advisable at this point in the 
compilation process?

I could move the transformation to a point much earlier (perhaps in the 
java front end) but that might make it much more difficult to analyze 
the control flow to see if the optimization is possible.

Opinions?

Thanks,

David Daney