We shouldn't be able to tail-call #[track_caller] functions

I tried this code:

struct Struct;

fn foo<T: ?Sized>(t: &T) -> usize {
    become bar(t);
}

#[track_caller]
fn bar<T: ?Sized>(_: &T) -> usize {
    456
}

fn main() {
    assert_eq!(foo(&Struct), 456);
}

I expected to see: Not sure actually.

Instead, LLVM's opt --passes=verify fails b/c we're performing a tail call with an incompatible signature.

Meta

rustc --version --verbose:

whatever is nightly on august 12, 2025

Not really an ICE per se, but a llvm bug. It doesn't show up on nightly b/c llvm is not built with debug assertions, I think, but it does show up when you pass the generated llvm to opt. I did this in alive2.

I think we need to rework how tail calls handle both pre-mono and post-mono #[track_caller] :3

cc #145270 (comment)

I think the problem with #[track_caller] is deeper than this -- specifically, I think we actually need to disallow calling #[track_caller] functions altogether (which I think we can only check for during monomorphization), or at least we need to be adjusting such calls to a "reify shim" when monomorphizing those tail calls, though I think that ends up leading to us pushing a stack frame which seems not good.

waffle has #144865 where these are linted as not doing anything until we can check for them post/during mono.

gamer

Duplicate of #144755 then

Navigation Menu

Repository navigation

We shouldn't be able to tail-call `#[track_caller]` functions #145321

Meta

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

We shouldn't be able to tail-call #[track_caller] functions #145321

Description

Meta

Activity

compiler-errors commented on Aug 12, 2025

lqd commented on Aug 12, 2025

compiler-errors commented on Aug 12, 2025

compiler-errors commented on Aug 12, 2025

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

We shouldn't be able to tail-call `#[track_caller]` functions #145321