C++ constructor bug? (and what are the semantics of a RANGE_EXPR?)

[Steven Bosscher <stevenb at suse dot de>]

Hi,

With a patch for the wrong-code bug PR18191 I'm seeing a new failure
for g++.dg/init/new4.C.  That test case looks like this:

int *x = new int [2] ();

which looks like this in the .original tree dump:

  (void) (x = (int *) (TARGET_EXPR <D.1910, (int[2] *) operator new [] (8)>, try
    {
      *D.1910 = {0};
    }
  catch
    {
      operator delete [] ((void *) D.1910);
    }, D.1910))


When I count the number of elements in the constructor for this test,
I get 3 elements.  The tree looks like this:

 <constructor 0x2a95999c00
    type <array_type 0x2a959e1b60
        type <integer_type 0x2a958979a0 int public SI
            size <integer_cst 0x2a95890a20 constant invariant 32>
            unit size <integer_cst 0x2a95890540 constant invariant 4>
            align 32 symtab 0 alias set -1 precision 32 min <integer_cst 0x2a95890990 -2147483648> max <integer_cst 0x2a958909c0 2147483647>
            pointer_to_this <pointer_type 0x2a958b3c40>>
        BLK
        size <integer_cst 0x2a95890bd0 constant invariant 64>
        unit size <integer_cst 0x2a95890c00 constant invariant 8>
        align 32 symtab 0 alias set -1
        domain <integer_type 0x2a9596a540 type <integer_type 0x2a95897380 long unsigned int>
            DI size <integer_cst 0x2a95890bd0 64> unit size <integer_cst 0x2a95890c00 8>
            align 64 symtab 0 alias set -1 precision 64 min <integer_cst 0x2a95890570 0> max <integer_cst 0x2a95890660 1>>
        pointer_to_this <pointer_type 0x2a959e1ee0>>
    constant invariant
    arg 0 <tree_list 0x2a9599bf00
        purpose <range_expr 0x2a958968c0 type <integer_type 0x2a95897380 long unsigned int>
            constant invariant
            arg 0 <integer_cst 0x2a95890570 constant invariant 0>
            arg 1 <integer_cst 0x2a95890870 constant invariant 2>>
        value <integer_cst 0x2a958ac690 constant invariant 0>>>

Note especially the last two lines.  There is a RANGE_EXPR with two
operands, with values 0 and 2.  This comes out of the C++ front end
and I suspect it's wrong.  From the comment for CONSTRUCTOR:

    If the TREE_PURPOSE is a RANGE_EXPR, it is a short-hand for many nodes,
    one for each index in the range.

But the index "2" is not in the range for the target object of this
constructor.  Smells like off-by-one in the C++ front end.

But I can't be sure because RANGE_EXPR is not documented.  Not in
doc/c-tree.texi and not in tree.def (ARRAY_RANGE_REF is also not in
the texi file btw).

Help! anyone?  :-)

Gr.
Steven