Travis Vitek wrote:
Martin Sebor wrote:
The test fails to compile with HP aCC 3.63. The error messages
point to the patch for the issue:
http://svn.apache.org/viewvc?view=rev&revision=629550
Actually, I think it was me that caused the problem, in the following
change:
http://svn.apache.org/viewvc?view=rev&revision=669747
The fix was for STDCXX-170 and friends, but break on every compiler when
the Iterator::operator*() returns a temporary (which is legal).
That's what I thought was your objection to Farid's patch. Am I
to understand that the code somehow detects whether operator*()
returns a rvalue or lvalue and the branch with the cast is only
supposed to be entered for lvalues? (I'm still uncomfortable
with the cast and would like to understand why it's safe).
It looks like I'd need to do a special case when the iterator type is
pointer. I don't see any way to legally check for no overlap without
that, so the only option I can see then is to always make the copy and
fix it with an overload selection trick (which would only be appropriate
for 4.3.x).
I think it should be fine to optimize just the common case
(const_pointer) and leave the rest unoptimized (i.e., make
a copy). Or can you think of another common use that should
be optimized as well?
Travis
Looking at the patch I don't see how the reinterpret_cast to
const_reference can possibly be correct, and I'm not sure we
satisfactorily resolved the same question the first time it
was raised back in March:
http://markmail.org/message/eijfmt3wxhg25bvs
Farid?
Thanks
Martin
