[jsmthrowaway]

> an int key performs better.

A UUID is a 128-bit integer. That people do not store, generate, or interact with them that way is the bug.

I'd kill for a 128-bit architecture so a UUID compare would be a single instruction, and it's too bad consensus is that we don't really need it.

[mortehu]

> I'd kill for a 128-bit architecture so a UUID compare would be a single instruction

Well, there's PCMPESTRI on Intel.

http://www.felixcloutier.com/x86/PCMPESTRI.html

[jsmthrowaway]

Yeah, I just said this in another comment, but SSE is my weak point; I stopped writing assembly and dealing with this stuff before SSE was even introduced.

That looks basically like what you need, though, yeah. Cool.

[gnaritas]

An int is either 4 or 8 bytes, it's still going to perform better no matter how you treat the GUID.

edit: bytes

[jessedhillon]

> An int is either 4 or 8 bits

This perfectly summarizes the bitwise-innumeracy of the argument against UUIDs. Even without 128 bit word sizes, most UUIDs are going to be non-matches in their lower bits, assuming a random distribution. There's no computational efficiency gained in the vastly wide critical path here.

[gnaritas]

> most UUIDs are going to be non-matches in their lower bits, assuming a random distribution

But you can't assume a random distribution because anyone using UUID's for database keys isn't using random UUID's, they're using sequential ones.

[jsmthrowaway]

Which would still make the low-order quadword not match.

[gnaritas]

It's a database though, you don't get to control how comparisons are done. Guid's are larger than ints, so using them as keys or joining on them is going to be slightly slower (not that it makes them bad, I'm not against GUID's). Just as using a bigint will be slower than an int.

[jsmthrowaway]

If I were implementing the uuid type in Postgres and did anything other than `if(unlikely(LO(uuid1) == LO(uuid2)) && unlikely(HI(uuid1) == HI(uuid2)))` I'd consider myself doing the wrong thing, personally. Two 64-bit compares, two flag checks, and a branch, which is about as fast as it's going to get in the absence of a 128-bit architecture.

You could actually vectorize the compares in SSE, I believe (but I stopped writing assembly before SSE existed and don't know much about it at a low level). This is why it's so critical to treat a UUID as an integer instead of a string, and I see a lot of code that does the latter.

[jandrese]

I hope you mean bytes. On a 128bit architecture that would be a single instruction compare, just like a 32 or 64 bit compare. Performance would be roughly the same (minus maybe a cache miss because you blew it out with your fat ints).