Proto/Env: remove Bytes.unsafe_of_string

This commit is contained in:
Grégoire Henry 2017-02-24 18:24:28 +01:00
parent c277b5b56d
commit 863869eb40

View File

@ -301,125 +301,6 @@ val equal: t -> t -> bool
always-correct {!Bytes.to_string} and {!Bytes.of_string} instead.
*)
val unsafe_to_string : bytes -> string
(** Unsafely convert a byte sequence into a string.
To reason about the use of [unsafe_to_string], it is convenient to
consider an "ownership" discipline. A piece of code that
manipulates some data "owns" it; there are several disjoint ownership
modes, including:
- Unique ownership: the data may be accessed and mutated
- Shared ownership: the data has several owners, that may only
access it, not mutate it.
Unique ownership is linear: passing the data to another piece of
code means giving up ownership (we cannot write the
data again). A unique owner may decide to make the data shared
(giving up mutation rights on it), but shared data may not become
uniquely-owned again.
[unsafe_to_string s] can only be used when the caller owns the byte
sequence [s] -- either uniquely or as shared immutable data. The
caller gives up ownership of [s], and gains ownership of the
returned string.
There are two valid use-cases that respect this ownership
discipline:
1. Creating a string by initializing and mutating a byte sequence
that is never changed after initialization is performed.
{[
let string_init len f : string =
let s = Bytes.create len in
for i = 0 to len - 1 do Bytes.set s i (f i) done;
Bytes.unsafe_to_string s
]}
This function is safe because the byte sequence [s] will never be
accessed or mutated after [unsafe_to_string] is called. The
[string_init] code gives up ownership of [s], and returns the
ownership of the resulting string to its caller.
Note that it would be unsafe if [s] was passed as an additional
parameter to the function [f] as it could escape this way and be
mutated in the future -- [string_init] would give up ownership of
[s] to pass it to [f], and could not call [unsafe_to_string]
safely.
We have provided the {!String.init}, {!String.map} and
{!String.mapi} functions to cover most cases of building
new strings. You should prefer those over [to_string] or
[unsafe_to_string] whenever applicable.
2. Temporarily giving ownership of a byte sequence to a function
that expects a uniquely owned string and returns ownership back, so
that we can mutate the sequence again after the call ended.
{[
let bytes_length (s : bytes) =
String.length (Bytes.unsafe_to_string s)
]}
In this use-case, we do not promise that [s] will never be mutated
after the call to [bytes_length s]. The {!String.length} function
temporarily borrows unique ownership of the byte sequence
(and sees it as a [string]), but returns this ownership back to
the caller, which may assume that [s] is still a valid byte
sequence after the call. Note that this is only correct because we
know that {!String.length} does not capture its argument -- it could
escape by a side-channel such as a memoization combinator.
The caller may not mutate [s] while the string is borrowed (it has
temporarily given up ownership). This affects concurrent programs,
but also higher-order functions: if [String.length] returned
a closure to be called later, [s] should not be mutated until this
closure is fully applied and returns ownership.
*)
val unsafe_of_string : string -> bytes
(** Unsafely convert a shared string to a byte sequence that should
not be mutated.
The same ownership discipline that makes [unsafe_to_string]
correct applies to [unsafe_of_string]: you may use it if you were
the owner of the [string] value, and you will own the return
[bytes] in the same mode.
In practice, unique ownership of string values is extremely
difficult to reason about correctly. You should always assume
strings are shared, never uniquely owned.
For example, string literals are implicitly shared by the
compiler, so you never uniquely own them.
{[
let incorrect = Bytes.unsafe_of_string "hello"
let s = Bytes.of_string "hello"
]}
The first declaration is incorrect, because the string literal
["hello"] could be shared by the compiler with other parts of the
program, and mutating [incorrect] is a bug. You must always use
the second version, which performs a copy and is thus correct.
Assuming unique ownership of strings that are not string
literals, but are (partly) built from string literals, is also
incorrect. For example, mutating [unsafe_of_string ("foo" ^ s)]
could mutate the shared string ["foo"] -- assuming a rope-like
representation of strings. More generally, functions operating on
strings will assume shared ownership, they do not preserve unique
ownership. It is thus incorrect to assume unique ownership of the
result of [unsafe_of_string].
The only case we have reasonable confidence is safe is if the
produced [bytes] is shared -- used as an immutable byte
sequence. This is possibly useful for incremental migration of
low-level programs that manipulate immutable sequences of bytes
(for example {!Marshal.from_bytes}) and previously used the
[string] type for this purpose.
*)
(** Functions reading and writing bytes *)
val get_char: t -> int -> char