`[expr; 0]` ignores lifetimes

[theemathas]

I tried this code:

use std::marker::PhantomData;

pub struct Thing<'a> {
    _phantom: PhantomData<fn(&'a ()) -> &'a ()>,
}

pub fn foo<'a, 'b>(x: Thing<'a>) -> [Thing<'b>; 0] {
    [x; 0]
}

pub fn lol<'a, 'b>(x: &'a i32) -> [&'b i32; 0] {
    [x; 0]
}

I expected the code not to compile, but instead it compiles without errors or warnings.

See also #79580, which is about [expr; 0] being weird.

Meta

Reproducible on the playground with 1.90.0-nightly (2025-07-15 3014e79f9c8d5510ea7b)

@rustbot labels +A-zst +A-lifetimes

[compiler-errors]

This is probably not unsound, but the types being connected here seems like a "nice" property to have.

[compiler-errors]

Though I'm also not totally certain how we'd enforce this.

We could emit a fake edge that accesses and assigns _0[0] = expr or something, but that seems dumb; and it doesn't seem clear to me how else we'd enforce that connection between the type of the expr temp that ends up getting dropped and the rvalue zero length array that we evaluate.

[zachs18]

This used to fail to compile in 1.62.0 (and nightly-2022-06-07), it appears to have started compiling in 1.63.0 (and nightly-2022-06-08) godbolt link.

Possibly notably, 1.63.0 was the version that fully removed the pre-NLL borrowchecker IIUC. Note that the bisection is the same in all editions (that existed at the time: 2015, 2018, and 2021), so it doesn't seem to be related to the "migration mode" for Rust 2015 mentioned in the release notes.

cargo bisect-rustc

$ cargo bisect-rustc --regress success --preserve --start 1.62.0 --end 1.63.0

// ... snip

searched toolchains nightly-2022-05-14 through nightly-2022-06-24


********************************************************************************
Regression in nightly-2022-06-08
********************************************************************************

fetching https://static.rust-lang.org/dist/2022-06-07/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2022-06-07: 40 B / 40 B [=====================================================================================================================] 100.00 % 1.24 MB/s converted 2022-06-07 to 50b00252aeb77b10db04d65dc9e12ce758def4b5
fetching https://static.rust-lang.org/dist/2022-06-08/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2022-06-08: 40 B / 40 B [=====================================================================================================================] 100.00 % 1.07 MB/s converted 2022-06-08 to 5435ed6916a59e8d5acba2149316a841c3905cbd
looking for regression commit between 2022-06-07 and 2022-06-08
fetching (via remote github) commits from max(50b00252aeb77b10db04d65dc9e12ce758def4b5, 2022-06-05) to 5435ed6916a59e8d5acba2149316a841c3905cbd
ending github query because we found starting sha: 50b00252aeb77b10db04d65dc9e12ce758def4b5
get_commits_between returning commits, len: 7
  commit[0] 2022-06-06: Auto merge of #97730 - flip1995:clippyup, r=Manishearth
  commit[1] 2022-06-06: Auto merge of #97809 - matthiaskrgr:rollup-ajyvjd3, r=matthiaskrgr
  commit[2] 2022-06-07: Auto merge of #97801 - RalfJung:miri, r=RalfJung
  commit[3] 2022-06-07: Auto merge of #95565 - jackh726:remove-borrowck-mode, r=nikomatsakis
  commit[4] 2022-06-07: Auto merge of #97512 - scottmcm:add-coldcc, r=nagisa,lcnr
  commit[5] 2022-06-07: Auto merge of #97825 - Dylan-DPC:rollup-ya51k1k, r=Dylan-DPC
  commit[6] 2022-06-07: Auto merge of #97835 - Dylan-DPC:rollup-0ae3pwp, r=Dylan-DPC
ERROR: no CI builds available between 50b00252aeb77b10db04d65dc9e12ce758def4b5 and 5435ed6916a59e8d5acba2149316a841c3905cbd within last 167 days

[clarfonthey]

I mean, it would make most sense to me that the code provided in the OP:

use std::marker::PhantomData;

pub struct Thing<'a> {
    _phantom: PhantomData<fn(&'a ()) -> &'a ()>,
}

pub fn foo<'a, 'b>(x: Thing<'a>) -> [Thing<'b>; 0] {
    [x; 0]
}

pub fn lol<'a, 'b>(x: &'a i32) -> [&'b i32; 0] {
    [x; 0]
}

is equivalent to:

use std::marker::PhantomData;

pub struct Thing<'a> {
    _phantom: PhantomData<fn(&'a ()) -> &'a ()>,
}

pub fn foo<'a, 'b>(x: Thing<'a>) -> [Thing<'b>; 0] {
    let _: Thing<'b> = x;
    []
}

pub fn lol<'a, 'b>(x: &'a i32) -> [&'b i32; 0] {
    let _: &'b i32 = x;
    []
}

Ultimately, the value is discarded, but it should still perform a type-level assertion, which should honour lifetimes.