Re: [PATCH] Tighten length error check in string::replace

[Paolo Carlini <pcarlini at unitus dot it>]

Benjamin Kosnik wrote:

>>+void
>>+test05()
>>+{
>>    
>>
>
>bool test = true;
>
Ok.

>>+  std::string str01 = "beyond skin";
>>+  std::string str02 = "glam";
>>+
>>+  try {
>>+    str01.replace(8, 7, str02.c_str(), str02.max_size() - 4);
>>+  }
>>+  catch(std::length_error& fail) {
>>+    VERIFY( std::string(fail.what()) == "basic_string::replace" );
>>+  }
>>+  catch(...) {
>>+    VERIFY( false );
>>+  }
>>+}
>>    
>>
>Is this really correct behavior, to throw length_error here? I'm
>uncertain.
>
Well, in the past I followed mostly Josuttis in working on such replace 
members.

Preliminarily, we should consider that, in case of replace(..., ..., 
const char* chars, size_type chars_len) it is assumed that chars has at 
least chars_len characters, '\0' has no special meaning and we cannot 
control the effective available length of the source data (str02 above 
is there only to make available the pointer str01.c_str()).

Second, according to Josuttis at least, replace throws lenght_error if 
the /resulting/ size (notice that in order to compute it exactly we need 
__foldn1 not __n1) exceeds the maximum number of characters, thus, I 
understand, the test I wrote some time ago must be amended to __size - 
__foldn1 > this->max_size() - __n2.

For example, in the testcase, __size == 11, __foldn1 == 3, 
this->max_size - __n2 == 4 and the test for __throw_length_error is 
definitely true.

Notice, moreover, that (my bad!) the current test (__size - __n1 > 
this->max_size() - __n2) for sure is not self-consistent, since both 
__size and __n1 are unsigned quantities and therefore if the user 
specifies __n1 > __size (he can) false positives may result from the 
comparison.

Ciao,
Paolo.