Add a raw "address of" operator #64588

matthewjasper · 2019-09-18T20:48:59Z

Parse and feature gate &raw [const | mut] expr (feature gate name is raw_address_of)
Add mir::Rvalue::AddressOf
Use the new Rvalue for:
- the new syntax
- reference to pointer casts
- drop shims for slices and arrays
Stop using mir::Rvalue::Cast with a reference as the operand
Correctly evaluate mir::Rvalue::{Ref, AddressOf} in constant propagation

cc @Centril @RalfJung @oli-obk @eddyb
cc #64490

src/librustc/hir/mod.rs

src/librustc/mir/mod.rs

src/librustc/mir/visit.rs

Centril · 2019-09-18T21:57:46Z

src/librustc_mir/transform/qualify_consts.rs

cc @ecstatic-morse @oli-obk re. the qualification code.

Yea I think none of it should be changed and we should just treat raw borrows as unpromotable (because any potentially promotable things had to error out before)

src/librustc_mir/hair/mod.rs

src/librustc_mir/borrow_check/mod.rs

src/librustc_codegen_ssa/mir/rvalue.rs

src/test/ui/packed/packed-struct-address-of-element.rs

Centril · 2019-09-18T22:37:29Z

Thanks for the PR. I've made an initial review above ^--- (but Ralf should also review this).

Some more remarks on the PR description:

Use the new Rvalue for:

reference to pointer casts

This is not part of the RFC and should remain UB.

drop shims for slices and arrays

Can you elaborate on this please? Is this a spec change or just internal refactoring?

Stop using mir::Rvalue::Cast with a reference as the operand

Seems like the same as &x as *const _ using Rvalue::AddrOf (i.e. not part of the RFC).

Can you please also extend the unused_parens lint to handle &raw [const|mut] $place_expr?

ecstatic-morse · 2019-09-19T17:05:12Z

@matthewjasper You need to handle Rvalue::AddressOf(_, _) here and here as well. This is the part of the code that forbids taking &mut references in const bodies. We need to forbid anything that allows mutation through indirection, since that can cause an Option<NeedsDropType> to go from None to Some.

matthewjasper · 2019-09-19T20:15:03Z

This is not part of the RFC and should remain UB.

What should be UB? &x as *const _ still is and x as *const _ could never cause UB.

Can you elaborate on this please? Is this a spec change or just internal refactoring?

It's a refactoring. If the precise stacked borrows effects of slice/array drop shims is defined, then maybe it's a change, but not a very big one since the parameter is a mutable reference, and the calls to drop the elements create mutable references.

Can you please also extend the unused_parens lint to handle &raw [const|mut] $place_expr?

This can be a follow up. The (lack of) behavior here is consistent with & AFAICT.

Centril · 2019-09-19T20:28:36Z

What should be UB? &x as *const _ still is and x as *const _ could never cause UB.

Oh I see; seems I misread. Can you add a mir-opt test to that effect?

This can be a follow up. The (lack of) behavior here is consistent with & AFAICT.

Sure, can you add a ticket to the tracking issue when this PR is merged?

ecstatic-morse · 2019-09-29T23:19:14Z

@matthewjasper @RalfJung Some tests need to be added to ensure that &raw mut is forbidden in const contexts. Also, there's now two validation passes (one uses dataflow to support #49146) that are checked against one another for compatibility. Once &raw mut references are handled, I'm happy to port the changes to the new validator in librustc_mir/transform/check_consts/validation.rs.

matthewjasper · 2019-12-19T13:01:32Z

@bors r=oli-obk

bors · 2019-12-19T13:01:34Z

📌 Commit a749116 has been approved by oli-obk

RalfJung · 2019-12-21T12:01:34Z

This is awesome, thanks so much. :)

RalfJung · 2019-12-21T12:03:43Z

src/librustc_mir/transform/add_retag.rs

+                    // Retag-as-raw after escaping to a raw pointer.
+                    StatementKind::Assign(box (ref place, Rvalue::AddressOf(..))) => {
+                        (RetagKind::Raw, place.clone())
                    }


So this is doing that for all &raw now, right? Certainly a reasonable start, though it doesn't seem to be entirely equivalent to the previous operation which also made sure that the "input" is a reference.

There wasn't any way to write &raw const *x where x is a raw pointer before.

rust-highfive assigned cramertj Sep 18, 2019