On 06/25/2012 05:25 AM, Jason Merrill wrote:
> On 06/11/2012 12:11 PM, Florian Weimer wrote:
>> + tree inner_nelts_cst = maybe_constant_value (inner_nelts);
>> + if (!TREE_CONSTANT (inner_nelts_cst))
>> + {
>> + if (complain & tf_error)
>> + error_at (EXPR_LOC_OR_HERE (inner_nelts),
>> + "array size in operator new must be constant");
>
> Please use cxx_constant_value to give a more specific error about what
> is non-constant.

Thanks for your review.

I tried this, but for this example program,

/* 1 */ void f(int n, int m)
/* 2 */ {
/* 3 */   typedef char T[n];
/* 4 */   new T[m];
/* 5 */ }

GCC reports these errors:

/tmp/t.C: In function ‘void f(int, int)’:
/tmp/t.C:4:18: error: array size in operator new must be constant
    /* 4 */   new T[m];
                     ^
/tmp/t.C:4:18: error: ‘n’ is not a constant expression

The message should point to the typedef, but instead, it references the 
line with operator new (which doesn't even contain the variable n).

For the non-VLA typedef case, it is an improvement.  But I would like to 
leave in both errors, as in the attached patch.

If you have suggestions how to improve cxx_constant_value error 
reporting, I can look into that in a separate patch.

>> + /* Warn if we performed the (T[N]) to T[N] transformation and N is
>> + variable. */
>> + if (outer_nelts_from_type
>> + && !TREE_CONSTANT (maybe_constant_value (outer_nelts))
>> + && (complain & tf_warning_or_error))
>> + pedwarn(EXPR_LOC_OR_HERE (outer_nelts), OPT_Wvla,
>> + "ISO C++ does not support variable-length array types");
>
> Here, if we aren't complaining we should return error_mark_node; we
> always need to act pedantic in SFINAE context.

The new version should do that.

I've added a new test case about error reporting in templates (where we 
previously accepted ill-formed code such as new long[n][T::n] with n, 
T::n both variables).

Bootstrapped and tested on x86_64-unknown-linux-gnu, with no new 
regressions.  OK for trunk?

-- 
Florian Weimer / Red Hat Product Security Team