On 08/08/18 19:33, Bernd Edlinger wrote:
> Hi Martin,
>
> sorry, I hope you forgive me, when I add a few comments to your patch.
>
>> + unsigned HOST_WIDE_INT nelts = CONSTRUCTOR_NELTS (ctor);
>> + tree eltype = TREE_TYPE (type);
> ...
>> + /* Bail if the CTOR has a block of more than 256 embedded nuls
>> + due to implicitly initialized elements. */
>> + unsigned nelts = (idx - str.length ()) + 1;
>> + if (nelts > 256)
>> + return NULL_TREE;
>
> nelts shadows previous nelts.
>
>> + if (!nelts || str.length () < i)
>
> I don't understand when is str.length () < i ?
>
>> + /* Append a nul for the empty initializer { } and for the last
>> + explicit initializer in the loop above that is a nul. */
>> + str.safe_push (0);
>> +
>> + /* Build a string literal but return the embedded STRING_CST. */
>> + tree res = build_string_literal (str.length (), str.begin ());
>> + res = TREE_OPERAND (TREE_OPERAND (res, 0), 0);
>> + return res;
>> +}
>
>
> res has a different type than the object you initialize.
> I think you should use:
>
> tree res = build_string (str.length (), str.begin ());
> TREE_TYPE (res) = type;
> return res;
>
Sorry, again, but could it be possible that this test case changed
with your patch?
$ cat w.c
const char x[] = { };
int main()
{
__builtin_printf("%ld\n", sizeof(x));
return 0;
}
$ gcc w.c
$ ./a.out
1
without your patch
$ ./a.out
0
Bernd.
