!nonnull not emitted when loading the field of nonnull

[beepster4096]

The following code generates the following LLVM IR. (Godbolt)

#![feature(rustc_attrs, bench_black_box)]

#[rustc_layout_scalar_valid_range_start(1)]
pub struct NonNull {
    pointer: *const (),
}

#[no_mangle]
pub fn use_nonnull(p: &NonNull) {
    std::hint::black_box(p.pointer);
}

define void @use_nonnull(i8** align 8 %p) unnamed_addr #1 {
  %0 = bitcast i8** %p to {}**
  %_3 = load {}*, {}** %0, align 8
  %_2 = call {}* @_ZN4core4hint9black_box17h01db5b3428eb9d8dE({}* %_3)
  br label %bb1

bb1:                                              ; preds = %start
  ret void
}

The load here does not have !nonnull metadata on it, even though rustc knows the field will never be null. This is not currently a problem in practice since std::ptr::NonNull's field is only accessed through a method that takes the NonNull by value. This copies the NonNull instead of the underlying pointer, which does properly emit !nonnull. However with #95576, dereferencing a box will now go though the underlying pointer field, which can trigger this issue.

[beepster4096]

@rustbot label +A-codegen +A-LLVM

[RalfJung]

Note that if we follow my suggested answer to https://github.com/rust-lang/unsafe-code-guidelines/issues/77, then &NonNull may actually point to 0, so the suggested optimization would be incorrect. Under my proposal, only data that you actually use by-value is required to comply with the validity invariant.

[RalfJung]

So, basically, under my proposal the following are not equivalent:

pub fn use_nonnull1(p: &NonNull) {
    std::hint::black_box(p.pointer);
}

pub fn use_nonnull2(p: &NonNull) {
    let p = *p;
    std::hint::black_box(p.pointer);
}

[CAD97]

Some variations on the theme for reference: (Godbolt)

; std::hint::black_box(p.pointer);
define void @use_nonnull_a(i8** align 8 %p) unnamed_addr #1 {
  %0 = bitcast i8** %p to {}**
  %_3 = load {}*, {}** %0, align 8
  %_2 = call {}* @core_hint_black_box({}* %_3)
  br label %bb1

bb1:                                              ; preds = %start
  ret void
}

; std::hint::black_box({*p}.pointer);
define void @use_nonnull_b(i8** align 8 %p) unnamed_addr #1 {
  %p1 = load i8*, i8** %p, align 8, !nonnull !3, !noundef !3
  %_4 = bitcast i8* %p1 to {}*
  %_3 = call {}* @core_hint_black_box({}* %_4)
  br label %bb1

bb1:                                              ; preds = %start
  ret void
}

It looks like {*p}.pointer (note the {} to induce a typed copy) does get the !nonnull metadata.

That said, even with reference validity ignoring memory contents, I'd still expect the deref/retag to assert memory validity and justify the !nonnull metadata here. But even if it doesn't, I'd still expect the rustc attribute to impact the copy from the field; instead of a typed copy *const () -> *const (), a "typed copy" #[rustc_layout_scalar_valid_range_start(1)] *const () -> *const (). The compiler literally knows that more is going on here than a simple typed copy of a *const () field, it's just forgetting to write that down.

IOW, this feels like a bug in how #[rustc_layout_scalar_valid_range_start(1)] works independent of how the memory model works; the goal of the attribute is that loading the field should be known in the valid range despite coming out as an unbounded type.

[RalfJung]

I'd still expect the deref/retag to assert memory validity

I don't think they should do anything like that (and in Miri, they don't).

This code does not have UB according to how I think Rust should be specified (and accordingly, it passes in Miri):

#![feature(rustc_attrs, bench_black_box)]

#[rustc_layout_scalar_valid_range_start(1)]
pub struct NonNull {
    pointer: *const (),
}

#[no_mangle]
pub fn use_nonnull(p: &NonNull) {
    std::hint::black_box(p.pointer);
}

fn main() {
    let ptr = 0usize as *const ();
    use_nonnull(unsafe { std::mem::transmute(&ptr) });
}

[RalfJung]

Here is a version of the code that does have UB:

#![feature(rustc_attrs, bench_black_box)]

#[rustc_layout_scalar_valid_range_start(1)]
#[derive(Copy, Clone)]
pub struct NonNull {
    pointer: *const (),
}

#[no_mangle]
pub fn use_nonnull(p: &NonNull) {
    std::hint::black_box(*p);
}

fn main() {
    let ptr = 0usize as *const ();
    use_nonnull(unsafe { std::mem::transmute(&ptr) });
}

You just have to actually tell the compiler that you care about the NonNull, not just its field.

I have no objections to making sure that that code optimizes well. :)