Merge remote-tracking branch 'aturon/ownership-variants'

active/0000-ownership-variants.md

@@ -0,0 +1,156 @@
+- Start Date: (fill me in with today's date, 2014-08-13)
+- RFC PR #: (leave this empty)
+- Rust Issue #: (leave this empty)
+
+# Summary
+
+This is a *conventions RFC* for settling naming conventions when there
+are by value, by reference, and by mutable reference variants of an
+operation.
+
+# Motivation
+
+Currently the libraries are not terribly consistent about how to
+signal mut variants of functions; sometimes it is by a `mut_` prefix,
+sometimes a `_mut` suffix, and occasionally with `_mut_` appearing in
+the middle. These inconsistencies make APIs difficult to remember.
+
+While there are arguments in favor of each of the positions, we stand
+to gain a lot by standardizing, and to some degree we just need to
+make a choice.
+
+# Detailed design
+
+Functions often come in multiple variants: immutably borrowed, mutably
+borrowed, and owned.
+
+The canonical example is iterator methods:
+
+- `iter` works with immutably borrowed data
+- `mut_iter` works with mutably borrowed data
+- `move_iter` works with owned data
+
+For iterators, the "default" (unmarked) variant is immutably borrowed.
+In other cases, the default is owned.
+
+The proposed rules depend on which variant is the default, but use
+*suffixes* to mark variants in all cases.
+
+## The rules
+
+### Immutably borrowed by default
+
+If `foo` uses/produces an immutable borrow by default, use:
+
+* The `_mut` suffix (e.g. `foo_mut`) for the mutably borrowed variant.
+* The `_move` suffix (e.g. `foo_move`) for the owned variant.
+
+A consequence is that the iterator methods become: `iter`, `iter_mut`,
+and `iter_move`.
+
+**NOTE**: This convention covers only the *method* names for
+  iterators, not the names of the iterator types. That will be the
+  subject of a follow up RFC.
+
+### Owned by default
+
+If `foo` uses/produces owned data by default, use:
+
+* The `_ref` suffix (e.g. `foo_ref`) for the immutably borrowed variant.
+* The `_mut` suffix (e.g. `foo_mut`) for the mutably borrowed variant.
+
+### Exceptions
+
+For mutably borrowed variants, if the `mut` qualifier is part of a
+type name (e.g. `as_mut_slice`), it should appear as it would appear
+in the type.
+
+### References to type names
+
+Some places in the current libraries, we say things like `as_ref` and
+`as_mut`, and others we say `get_ref` and `get_mut_ref`.
+
+Proposal: generally standardize on `mut` as a shortening of `mut_ref`.
+
+
+## The rationale
+
+### Why suffixes?
+
+Using a suffix makes it easier to visually group variants together,
+especially when sorted alphabetically. It puts the emphasis on the
+functionality, rather than the qualifier.
+
+### Why `move`?
+
+Historically, Rust has used `move` as a way to signal ownership
+transfer and to connect to C++ terminology. The main disadvantage is
+that it does not emphasize ownership, which is our current narrative.
+On the other hand, in Rust all data is owned, so using `_owned` as a
+qualifier is a bit strange.
+
+The `Copy` trait poses a problem for any terminology about ownership
+transfer. The proposed mental model is that with `Copy` data you are
+"moving a copy".
+
+See Alternatives for more discussion.
+
+### Why `mut` rather then `mut_ref`?
+
+It's shorter, and pairs like `as_ref` and `as_mut` have a pleasant harmony
+that doesn't place emphasis on one kind of reference over the other.
+
+# Alternatives
+
+## Prefix or mixed qualifiers
+
+Using prefixes for variants is another possibility, but there seems to
+be little upside.
+
+It's possible to rationalize our current mix of prefixes and suffixes
+via
+[grammatical distinctions](https://github.com/rust-lang/rust/issues/13660#issuecomment-43576378),
+but this seems overly subtle and complex, and requires a strong
+command of English grammar to work well.
+
+## No suffix exception
+
+The rules here make an exception when `mut` is part of a type name, as
+in `as_mut_slice`, but we could instead *always* place the qualifier
+as a suffix: `as_slice_mut`. This would make APIs more consistent in
+some ways, less in others: conversion functions would no longer
+consistently use a transcription of their type name.
+
+This is perhaps not so bad, though, because as it is we often
+abbreviate type names. In any case, we need a convention (separate
+RFC) for how to refer to type names in methods.
+
+## `owned` instead of `move`
+
+The overall narrative about Rust has been evolving to focus on
+*ownership* as the essential concept, with borrowing giving various
+lesser forms of ownership, so `_owned` would be a reasonable
+alternative to `_move`.
+
+On the other hand, the `ref` variants do not say "borrowed", so in
+some sense this choice is inconsistent. In addition, the terminology
+is less familiar to those coming from C++.
+
+## `val` instead of `owned`
+
+Another option would be `val` or `value` instead of `owned`. This
+suggestion plays into the "by reference" and "by value" distinction,
+and so is even more congruent with `ref` than `move` is. On the other
+hand, it's less clear/evocative than either `move` or `owned`.
+
+## `into_iter`
+
+For the case of iteration, at least, it would make some sense to
+signal ownership transfer by treating the owned version as a
+conversion, `into_iter`. The main downside is that it would go against
+the general convention for ownership variants (and cannot be used as
+the general convention, because not all cases can be seen as
+conversions).
+
+Moreover, it's strange to see just the owning variant as a conversion
+-- why not `as_iter`, `as_iter_mut`, and `into_iter`?