Recent strstream regression: reversion needed?

[Paolo Carlini]

Paolo Carlini wrote:

> In the next hours I hope to make progress in understanding /why/! 

I bit of additional information: with the _M_buf = buf change the 
segmentation fault happens during the output, after the first 
reallocation, in basic_streambuf<_CharT, _Traits>::sputc(char_type __c). 
It looks like there is something wrong in fiddling with _M_buf after the 
initial allocation.

Benjamin, could you possibly check with 'valgrind' that the memory leak 
reported in libstdc++/6594 does /really/ happen with the old code? 
Really, that is not at all obvious to me given the _M_free(eback()) call 
in the original ~strstreambuf().

Perhaps I should add some information supporting this, therefore let us 
consider the testcase printing the integers 0-999 and the old code 
(before libstdc++/6594 commit).

- The value of buf at the end of each overflow call:
overflow::buf = 0x8049980
overflow::buf = 0x80499a8
overflow::buf = 0x80499f0
overflow::buf = 0x8049a78
overflow::buf = 0x8049b80
overflow::buf = 0x8049d88
overflow::buf = 0x804a190
overflow::buf = 0x804a998

- The value of this->eback() in ~strstreambuf():
this->eback() = 0x804a998

The fact that the last value of overflow::buf is identical to 
this->eback() seems to me sort of a proof that there are no memory leaks.

What do you think?

Ciao, Paolo.