[Issue 15624] opApply with @safe and @system variants can't be used with foreach syntax
d-bugmail at puremagic.com
d-bugmail at puremagic.com
Sat Nov 6 01:02:11 UTC 2021
https://issues.dlang.org/show_bug.cgi?id=15624
--- Comment #5 from Dlang Bot <dlang-bot at dlang.rocks> ---
dlang/dmd pull request #13226 "Rework fix for attribute based opApply overloads
(issue 15624)" was merged into master:
- a28680a845a06d0cac476442c25e7c7fe8e0d16d by MoonlightSentinel:
Rework fix for attribute based opApply overloads (issue 15624)
The old implementation uses `Type.covariant` to suppress overzealous
ambiguity errors for `opApply` specialisations. Deferring the ambiguity
checks after the attribute inference on the loop body allows for
`opApply` overloads with different attributes, e.g. `@safe / @system` for
`@safe / @system` delegates.
Example:
```d
struct Collection
{
/* A. */ int opApply(int delegate(int) @safe dg) @safe;
/* B. */ int opApply(int delegate(int) @system dg) @system;
}
void main() @safe
{
Collection col;
foreach (entry; col) {} // Resolves to A.
}
```
The previous code achieved this by checking whether the previously
selected `opApply` and the current overload are covariant in any
direction (i.e. `A.covariant(B)` or `B.covariant(A)`).
This implementation builds on the flawed assumption that the `opApply`
functions in this example are covariant - which they are not
(falsely reported due to a bug in `Type.covariant`).
Consider the types of the `opApply` overloads:
```d
A: int function(int delegate(int) @safe dg) @safe;
B: int function(int delegate(int) @system dg) @system;
```
Neither of those is an appropriate substitute for the other:
- A <= B is invalid due to the `opApply`'s attributes. A `@system`
function may not appear in place of a `@safe` function
- A => B is invalid due to the parameter types (`dg` attributes).
A function expecting a `@safe` delegate may not be treated as a
function accepting `@system` callbacks. Otherwise it could silently
execute a `@system` callback which is disallowed in `@safe` functions.
Note that `Type.covariant` already rejects covariance when using
`function` instead of `delegate` for the callbacks in the example.
---
The new implementation does not rely on the buggy behaviour of
`Type.covariant`. It checks two different aspects to defer the
`opApply` resolution:
1. different attributes on the `opApply` function
2. different attributes on the `opApply` function
3. covariance of the callbacks receiving the foreach body
(2) should handle different qualifiers on the foreach body
because missmatched parameter types are usually rejected by
`matchParamsToOpApply`.
---
This patch enables us to fix the aforementioned bug without breaking
existing code. It also enables the code added the test case wich
is valid despite being utterly useless.
https://github.com/dlang/dmd/pull/13226
--
More information about the Digitalmars-d-bugs
mailing list