Re: [PATCH net-next v7 1/5] rust: core abstractions for network PHY drivers

[Benno Lossin <benno.lossin@proton.me>]

On 28.10.23 01:26, Boqun Feng wrote:
> On Fri, Oct 27, 2023 at 10:50:45PM +0000, Benno Lossin wrote:
> [...]
>>>
>>> Hmm... but does it mean even `set_speed()` has the similar issue?
>>>
>>> 	let phydev: *mut phy_device = self.0.get();
>>> 	unsafe { (*phydev).speed = ...; }
>>
>> No that should be fine, take a look at the MIR output of the following
>> code [1]:
>>
>>      struct Foo {
>>          a: usize,
>>          b: usize,
>>      }
>>
>>      fn foo(ptr: *mut Foo) {
>>          unsafe { (*ptr).b = 0; }
>>      }
>>
>>      fn bar(ptr: *mut Foo) {
>>          unsafe { (&mut *ptr).b = 0; }
>>      }
>>
>> Aside from some alignment checking, foo's MIR looks like this:
>>
>>      bb1: {
>>          ((*_1).1: usize) = const 0_usize;
>>          return;
>>      }
>>
>> And bar's MIR like this:
>>
>>      bb1: {
>>          _2 = &mut (*_1);
>>          ((*_2).1: usize) = const 0_usize;
>>          return;
>>      }
>>
>> [1]: https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=f7c4d87bf29a64af0acc09ff75d3716d
>>
>> So I think that is fine, but maybe Gary has something else to say about it.
>>
> 
> Well when `-C opt-level=2`, they are the same:
> 
> 	https://godbolt.org/z/hxxo75YYh

It doesn't matter what the optimizer does, `bar` is unsound in our use-case
and `foo` is fine (I think).

-- 
Cheers,
Benno