Tue May 11, 2010 12:15 pm
I just read the discussion on the mailing list with the topic "Indexes: why signed instead of unsigned?" and, since I don't have email access at the moment, would like to contribute to the discussion via this channel.
As an argument in favor of having signed indexes, the following for-loop has been mentioned:
which of course would fail when replacing int by size_t.
However, I don't think that this is a very strong argument, since a for-loop with decreasing index can be written as
The only thing to take care of is to check for inequality instead of greater-than.
On the pro-side, using size_t instead of int, has the advantages that
1) since it's unsigned, out-of-bounds can be checked with a single "if(i>=size())"; the check for i<0 is unnecessary.
2) on 64bit architectures, it's 64 bit wide. This might not sound like a big advantage right now, but with the ever increasing memory capacities, I'm pretty sure that it's only a matter of time until someone hits the 2^31-limit. I'm using Eigen for signal processing tasks, and matrices with many million elements are not uncommon in my programs already.
3) Also, using the same code with STL and Eigen containers becomes easier, removing the need of casting into either the signed or the unsigned domain.
Anyway, these are just my thoughts on this topic.
Thanks for the great library, and best wishes,
Tue May 11, 2010 12:49 pm
Development discussion should happen on the mailing list, not here. There's already a thread for that on the list. There's also already an entry in the issue tracker. Now this discussion is happening at 3 different places!
We would have to remember doing that everytime. In the past this has bitten me, and compilers don't always emit warnings "condition always true" when checking i>=0 with unsigned i.
Moreover, notice that this makes the check a bit more expensive. While conditionally jumping if zero is just 1 instruction on most CPUs (at least x86), conditionally jumping based on comparing to 'stop' would typically involve 2 or 3 instructions. I agree this isn't a big deal since such reversed loops typically have large enough bodies, just mentioning a tiny disadvantage.
These checks are made only in asserts, which can be disabled in "release mode"; moreover, even when left enable, measurements show that in most practical situations they only have a minimal performance impact; so this doesn't make a significant difference.
That said, I agree that what you describe is an advantage. It just isn't a huge one.
Ah yes, while int is only 32 bits. That we can consider changing. ptrdiff_t would be the signed variant of size_t. Write to the list, this is too serious to discuss on the forum.
I understand. This is an argument in favor of ptrdiff_t vs. int, not in favor of unsigned vs. signed. However, while millions is common, billions is not too common. int gives us 2 billions, which is quite a lot. Asking the list for opinions.
Not trying to follow too closely the STL's design is precisely what allowed us to come up with the current one... that said, when possible, easing interaction with the STL is of course an advantage.
Note: continuing discussion on the list, won't reply again here.
Join us on Eigen's IRC channel: #eigen on irc.freenode.net
Have a serious interest in Eigen? Then join the mailing list!
Registered users: Baidu [Spider], Bing [Bot], dcihon, Google [Bot], Sogou [Bot]