------- Comment #7 from zackw at panix dot com 2006-07-12 23:19 ------- Created an attachment (id=11875) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=11875&action=view) C test case (with interesting implications)
I've found a plain C test case. In the process, I've found that the way libstdc++ <string> is coded interacts interestingly with the optimizer. In the attached file, has_bad_chars_bad is a literal translation to C of the code seen by the optimizers after inlining for the original C++ test case. Yes, libstdc++ <string> does the moral equivalent of ((struct rep*)path)[-1].len. This function compiles to the same bad code as my original test case. has_bad_chars_good, on the other hand, is how I naively thought <string> worked on the first read-through. That one compiles to code which looks optimal to me. I suspect some optimizer or other is not smart enough to see through this particular construct ... it would be good to make it do so, since we want libstdc++ <string> to generate good code. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28364
