[PATCH] Fix PR c++/23180 (4.1 regression)

[Josh Conner <jconner at apple dot com>]

In this PR, g++ is segfaulting on structure initializers containing a  
value like this:

    ((long) (&((struct type *) 42)->field[0])) - 42

(i.e., an attempt to mimic the offsetof functionality).

This is occurring because the tree that is passed to output_constant:

<addr_expr 0x24b47c0
    type <pointer_type 0x2422120
        type <integer_type 0x24141e0 char public unsigned QI
            size <integer_cst 0x2409340 constant invariant 8>
            unit size <integer_cst 0x2409360 constant invariant 1>
            align 8 symtab 0 alias set -1 precision 8 min  
<integer_cst 0x2409420 0> max <integer_cst 0x2409440 255>
            pointer_to_this <pointer_type 0x2422120>>
        unsigned SI
        size <integer_cst 0x24095e0 constant invariant 32>
        unit size <integer_cst 0x24092a0 constant invariant 4>
        align 32 symtab 0 alias set -1>
    constant invariant
    arg 0 <array_ref 0x240e390 type <integer_type 0x24141e0 char>
         arg 0 <component_ref 0x240e300 type <array_type 0x24b1de0>
             arg 0 <indirect_ref 0x24b46a0 type <record_type  
0x24b1d20 Track>
                 arg 0 <integer_cst 0x2491900 constant invariant  
42>> arg 1 <field_decl 0x24b1e40 soundName>>
        arg 1 <integer_cst 0x24b4680 constant invariant -42>>>

Is being evaluated through the function expand_simple_binop to  
generate the code for calculating "42-42".  However,  
expand_simple_binop isn't appropriate for calculating values outside  
of a function, as in this case.  The attached patch addresses this  
issue by detecting this condition in expand_expr_addr_expr_1, and  
computing the (42-42) expression in place.

Note that this behavior doesn't appear with the C front-end, only C+ 
+.  So, if there's some reason that output_constant shouldn't ever  
see a tree of this form, please let me know and I'll take a look at  
the front-end instead.

Tested on i686-pc-linux-gnu and arm-none-elf with no regressions.  OK  
for mainline?

- Josh

2005-09-08  Josh Conner  <jconner@apple.com>

        PR c++/23180
        * expr.c (expand_expr_addr_expr_1): Don't invoke
        expand_simple_binop if addr and offset are both constants.

Attachment: pr23180.patch
Description: Binary data