[meindnoch]

>Another one is that span has a performance overhead due to it being implemented as a normal type. If it was part of the language rather than a library type, the compiler could make assumptions about it, but instead it’s treated equivalent to me writing it myself.

False. Nothing prevents compilers from giving their own stdlib types special treatment under the hood.

[maccard]

That would be an ABI break which is just not happening, and you know it. As it is we’ve decided it’s more important to be able to use std span from libc++ on clang than it is to have an optimised version for people on their tool chain.

[meindnoch]

So you're saying that turning std::span from a standard library class into a language feature wouldn't break the ABI? How so? How would such a language construct fit into the existing ABI?

(for context: parent is referring to the fact that x64 calling conventions mandate structs larger than 64 bits to be passed in memory, which means that passing a 128bit std::span is going to be less efficient than passing a separate 64bit index and 64bit length, as those can go into registers)

[maccard]

No, the cat’s out of the bag with span (and unique pointer) now, we can’t go back. We knew this was a problem from unique_ptr and had an opportunity to not make the same mistake again, but instead we chose back compat for a new feature over something novel and performant

[gpderetta]

What is the issue exactly? Span is trivially copyable and destructible and, at least with the Itanium ABI, it can be passed (and returned) via registers: https://gcc.godbolt.org/z/4rbcshve4 .

Other ABIs might have different constraints but there is no reason why they couldn't special case std::span. In fact if span was a built-in type there is nothing preventing a compiler form picking a suboptimal ABI and being stuck with it. In any case it is not a standardization issue, but purely a QoI.

[meindnoch]

Yes, GCC can pass it in two registers. On the other hand Microsoft's x64 ABI doesn't:

>Structs and unions of size 8, 16, 32, or 64 bits, and __m64 types, are passed as if they were integers of the same size. Structs or unions of other sizes are passed as a pointer to memory allocated by the caller.

https://learn.microsoft.com/en-us/cpp/build/x64-calling-conv...

[maccard]

The parent commenter handled the ABI question for me, but I want to respond to :

> In any case it is not a standardization issue, but purely a QoI.

This is an enormous problem in the C++ ecosystem - playing hot-potato with whose fault it is, instead of trying to actually fix it. Span is a decent example, the standards committee get to say it's a vendor issue, and the vendors get to say that their hands are tied by the ABI. The result is that people spend time arguing about who should fix it rather than actually fixing it.

[gpderetta]

That's all well and good, but what would you do exaclty? The standard comitee cannot impose an ABI as it would be ignored by implementors. Implementors either own the ABI (MS for example) and are responsible for their owns screwups or there are other bodies that are responsible (the informal forum for the inter-vendor Itanium ABI for another example).

In any case this has nothing to do with std::span being an technically a library type or a built in. There is really no fundamental difference between the two.

For example std::complex and std::initializer_list have special handling on many compilers, just to mention two types.

[maccard]

> That's all well and good, but what would you do exaclty?

Start with having the standards committee accept that they are in fact where the buck stops, and that the language includes, whether they like it or not, the toolchain. They don't have to decide upon the toolchain, but their current MO of "toolchain/ABI issue == not our problem (except for when we decide we're not willing to make any backwards incompatible ABI changes, but only sometimes)." The vendors are already jumping through hoops to support what is being standardised (modules being the perfect example here).

I can't speak for std.complex as I've never had to use it, but I think initializer list would be a great example of "how much better would this be if it was special cased into the compiler". The benefit we would get from initialisation being consistent with the compiler far outweighs the benefit of being able to use libc++'s initialiser list with clang.

> There is really no fundamental difference between the two.

Except there is. If I write an implementation of the standard library, and provide an implementation of std span as (abbreviated) - https://gcc.godbolt.org/z/c1sz4neKG it's got to respect the various conventions instead of being treated as an opaque type (like a slice in go). If it's a `_Span`, the compiler is free to go "ok you're using this thing that I know all the internals of, and can reason about. I can elide bounds checks that don't pass muster, I can specify that I will generate code for this specific type that puts extent and data as registers in the following cases". But instead, on x64 (where I work 99% of the time so it's where my effort/knowledge is, sorry), we're bound by >64 == memory.

Now, you might call that a QOI issue, but I'd call it a design flaw that could have avoided an implementation issue, that we see on many features.