This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Bug 13689 as further evidence of my point about maintainers


As further evidence of my point that maintainers should be trying to speed up the code, rather than just fixing regressions, i'd like to direct your attention to bug 13689.

This bug is interesting because it is a latent bug, with wrong-code generation potential, that we only caught because of Honza reimplementing alias sets for speed.
IOW, if Honza hadn't tried and succeeded in speeding up the compiler, we wouldn't have seen this bug, but it would still be there.
It is *highly* unlikely we would have caught it with the splay tree implementation (because of the lack of bounds checking/assertions that values we expect to be in the tree were found in the tree) until it actually generated wrong code, and it probably would have not been fun to track down (it took 5 of us collaborating on IRC around someone's debugger quite a while to figure this one out).


Speeding up the compiler through data structure changes and such doesn't just make the compiler faster, it usually forces us to fix latent bugs.

I'll close by also noting that this potential[1] wrong-code bug would have been in the 3.4 release if we didn't allow this speedup to go in during stage 3!


--Dan
[1] I only say potential because no wrong-code testcase exists for it, though i'm sure one could create one through careful use of pch and types



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]