Miscompilation caused by incorrectly-deduced readonly on virtual call

I tried the following code. It requires a binary crate and a library crate.

main.rs

fn main() {
    let mut struct_c = lib::StructC::new();
    let struct_a = lib::StructA { a: 1, b: 2, c: 3 };
    lib::run(&mut struct_c, struct_a);
}

main's Cargo.toml

[package]
name = "compilerbug"
version = "0.1.0"
edition = "2024"

[dependencies]
lib = { path = "./lib" }

[profile.dev]
opt-level = 1
incremental = false

lib.rs

#[derive(Clone, Copy)]
pub struct StructA {
    pub a: i32,
    pub b: i32,
    pub c: i32,
}

pub trait TraitA {}

pub struct StructB {}

impl StructB {
    pub fn new() -> Self {
        Self {}
    }
}

impl TraitA for StructB {}

pub trait TraitB {
    fn get_trait_a(&self, struct_a: StructA) -> Option<Box<dyn TraitA>> {
        Some(Box::new(StructB::new()))
    }
}

pub struct StructC {}

impl StructC {
    pub fn new() -> Self {
        Self {}
    }
}

impl TraitB for StructC {
    fn get_trait_a(&self, mut struct_a: StructA) -> Option<Box<dyn TraitA>> {
        struct_a.a = i32::MAX;
        Some(Box::new(StructB::new()))
    }
}

#[inline(never)]
fn print_struct_a(struct_a: StructA) {
    println!("{}", struct_a.a);
}

#[inline(never)]
pub fn run(trait_b: &mut dyn TraitB, struct_a: StructA) {
    print_struct_a(struct_a);
    trait_b.get_trait_a(struct_a);
    print_struct_a(struct_a);

    // Uncommenting the following will fix the code.
    // println!("{}", struct_a.a);
}

lib's Cargo.toml

[package]
name = "lib"
version = "0.1.0"
edition = "2024"

Since struct_a conforms to Copy, passing struct_a to get_trait_a() should make a copy. Instead, when modifying struct_a inside of get_trait_a(), it modifies the original. When printing the struct a second time, the struct has been modified. To repro the bug, it requires opt-level >= 1 and incremental = false.

I expected the output in my command line to be:

1
1

Instead, this happened:

1
2147483647

Meta

rustc --version --verbose:

rustc 1.85.0 (4d91de4e4 2025-02-17)
binary: rustc
commit-hash: 4d91de4e48198da2e33413efdcd9cd2cc0c46688
commit-date: 2025-02-17
host: aarch64-apple-darwin
release: 1.85.0
LLVM version: 19.1.7

searched toolchains nightly-2023-10-01 through nightly-2023-11-11


********************************************************************************
Regression in nightly-2023-10-09
********************************************************************************

fetching https://static.rust-lang.org/dist/2023-10-08/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2023-10-08: 40 B / 40 B [==================================] 100.00 % 501.91 KB/s converted 2023-10-08 to 97c81e1b537088f1881c8894ee8579812ed9b6d1
fetching https://static.rust-lang.org/dist/2023-10-09/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2023-10-09: 40 B / 40 B [==================================] 100.00 % 348.89 KB/s converted 2023-10-09 to bf9a1c8a193fc373897196321215794c8bebbeec
looking for regression commit between 2023-10-08 and 2023-10-09
fetching (via remote github) commits from max(97c81e1b537088f1881c8894ee8579812ed9b6d1, 2023-10-06) to bf9a1c8a193fc373897196321215794c8bebbeec
ending github query because we found starting sha: 97c81e1b537088f1881c8894ee8579812ed9b6d1
get_commits_between returning commits, len: 10
  commit[0] 2023-10-07: Auto merge of #116416 - Kobzol:ci-host-llvm-17-0-2, r=Mark-Simulacrum
  commit[1] 2023-10-08: Auto merge of #114623 - Kobzol:opt-dist-gha-summaries, r=Mark-Simulacrum
  commit[2] 2023-10-08: Auto merge of #116450 - Kobzol:automation-try-bors-ci, r=Mark-Simulacrum
  commit[3] 2023-10-08: Auto merge of #116486 - van-ema:master, r=nikic
  commit[4] 2023-10-08: Auto merge of #116487 - tamird:avoid-unwrap-absolute, r=bjorn3
  commit[5] 2023-10-08: Auto merge of #116509 - Enselic:rustc-test-op, r=Mark-Simulacrum
  commit[6] 2023-10-08: Auto merge of #116514 - petrochenkov:nogccld, r=lqd
  commit[7] 2023-10-08: Auto merge of #116183 - cjgillot:debug-dse-always, r=oli-obk
  commit[8] 2023-10-08: Auto merge of #116454 - tmiasko:small-dominators, r=cjgillot
  commit[9] 2023-10-08: Auto merge of #116515 - petrochenkov:nolegflavor, r=lqd
ERROR: no CI builds available between 97c81e1b537088f1881c8894ee8579812ed9b6d1 and bf9a1c8a193fc373897196321215794c8bebbeec within last 167 days

well… among those that’s probably the LLVM bump then, right?

~~Bump host compiler on x64 dist Linux to LLVM 17.0.2 #116416~~ ~ or maybe not (see answers below)

Might also be #116454

Interestingly, with the

// Uncommenting the following will fix the code.

marked section uncommented (and fixed to refer to struct_a by the correct name), it’s also still broken for old Rust versions; regressing at the same point as before (2023-10-09), and actually being “fixed” again only at the following point:

searched toolchains nightly-2024-06-07 through nightly-2024-07-20


********************************************************************************
Regression in nightly-2024-06-29
********************************************************************************

fetching https://static.rust-lang.org/dist/2024-06-28/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2024-06-28: 40 B / 40 B [====================================] 100.00 % 1.15 MB/s converted 2024-06-28 to 9c3bc805dd9cb84019c124b9a50fdff1e62a7ec9
fetching https://static.rust-lang.org/dist/2024-06-29/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2024-06-29: 40 B / 40 B [====================================] 100.00 % 1.06 MB/s converted 2024-06-29 to e9e6e2e444c30c23a9c878a88fbc3978c2acad95
looking for regression commit between 2024-06-28 and 2024-06-29
fetching (via remote github) commits from max(9c3bc805dd9cb84019c124b9a50fdff1e62a7ec9, 2024-06-26) to e9e6e2e444c30c23a9c878a88fbc3978c2acad95
ending github query because we found starting sha: 9c3bc805dd9cb84019c124b9a50fdff1e62a7ec9
get_commits_between returning commits, len: 5
  commit[0] 2024-06-27: Auto merge of #127049 - flip1995:clippy-subtree-update, r=Manishearth
  commit[1] 2024-06-28: Auto merge of #127010 - GuillaumeGomez:update-puppeteer, r=notriddle
  commit[2] 2024-06-28: Auto merge of #127076 - matthiaskrgr:rollup-l01gm36, r=matthiaskrgr
  commit[3] 2024-06-28: Auto merge of #127000 - Oneirical:no-test-for-the-wicked, r=Kobzol
  commit[4] 2024-06-28: Auto merge of #126701 - onur-ozkan:build-lld-if-enabled, r=Kobzol
ERROR: no CI builds available between 9c3bc805dd9cb84019c124b9a50fdff1e62a7ec9 and e9e6e2e444c30c23a9c878a88fbc3978c2acad95 within last 167 days

I don’t know, is that this one perhaps?

ignore llvm::Lld if lld is not enabled #126701

It also isn’t “fixed” by uncommenting the indicated line, if opt-level = 1 is used.

It’s really interesting how it seems to be relevant whether or not the default method body does mutate its argument. E.g.:

pub trait Trait {
    fn method(&self, mut arr: [i32; 3]) {
        println!("{:p}", &arr[0]);

        // uncomment to "fix":
        // println!("{:p}", &mut arr[0]);
    }
}

impl Trait for () {
    fn method(&self, mut arr: [i32; 3]) {
        arr[0] = 42;
    }
}

fn inspect_arr(arr: [i32; 3]) {
    assert_eq!(format!("{}", arr[0]), "0");
}

pub fn run(trait_b: & dyn Trait, arr: [i32; 3]) {
    inspect_arr(arr);
    trait_b.method(arr);
    inspect_arr(arr);
}

use name_of_crate::*;

fn main() {
    run(&(), [0, 0, 0]);
}

here the println!("{:p}", &mut arr[0]); (or various other things that mutably access or actually mutate arr) do make a difference.

@saethlin

With -Zmir-opt-level=0 (thanks @saethlin for the pointer) it reproduces way further back to nightly-2023-07-16:

Details

********************************************************************************
Regression in nightly-2023-07-16
********************************************************************************

fetching https://static.rust-lang.org/dist/2023-07-15/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2023-07-15: 40 B / 40 B [=======================================================] 100.00 % 62.59 KB/s converted 2023-07-15 to ad963232d9b987d66a6f8e6ec4141f672b8b9900
fetching https://static.rust-lang.org/dist/2023-07-16/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2023-07-16: 40 B / 40 B [======================================================] 100.00 % 174.93 KB/s converted 2023-07-16 to 4c8bb79d9f565115637cc6da739f8389e79f3a29
looking for regression commit between 2023-07-15 and 2023-07-16
fetching (via remote github) commits from max(ad963232d9b987d66a6f8e6ec4141f672b8b9900, 2023-07-13) to 4c8bb79d9f565115637cc6da739f8389e79f3a29
ending github query because we found starting sha: ad963232d9b987d66a6f8e6ec4141f672b8b9900
get_commits_between returning commits, len: 5
  commit[0] 2023-07-14: Auto merge of #113471 - compiler-errors:new-solver-norm-escaping, r=lcnr
  commit[1] 2023-07-15: Auto merge of #113514 - jyn514:more-gha-groups, r=oli-obk
  commit[2] 2023-07-15: Auto merge of #112157 - erikdesjardins:align, r=nikic
  commit[3] 2023-07-15: Auto merge of #113732 - matthiaskrgr:rollup-nm5qy4i, r=matthiaskrgr
  commit[4] 2023-07-15: Auto merge of #113697 - GuillaumeGomez:rm-unneeded-externallocation-handling, r=lqd
ERROR: no CI builds available between ad963232d9b987d66a6f8e6ec4141f672b8b9900 and 4c8bb79d9f565115637cc6da739f8389e79f3a29 within last 167 days

~~- #112157 sounds related maybe?~~

That’s funny, my bisection (of my reduced example) with -Zmir-opt-level=0 also reproduces further back but not quite as far: only up to

searched toolchains nightly-2023-08-20 through nightly-2023-10-01


********************************************************************************
Regression in nightly-2023-08-27
********************************************************************************

fetching https://static.rust-lang.org/dist/2023-08-26/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2023-08-26: 40 B / 40 B [=======================================================] 100.00 % 566.52 KB/s converted 2023-08-26 to 734a0d0aa0d5cab60f94f6d0c6a014dae12915f1
fetching https://static.rust-lang.org/dist/2023-08-27/channel-rust-nightly-git-commit-hash.txt
nightly manifest 2023-08-27: 40 B / 40 B [=======================================================] 100.00 % 351.72 KB/s converted 2023-08-27 to 69e97df5ce571a777acd654ec3697ae8d25962ea
looking for regression commit between 2023-08-26 and 2023-08-27
fetching (via remote github) commits from max(734a0d0aa0d5cab60f94f6d0c6a014dae12915f1, 2023-08-24) to 69e97df5ce571a777acd654ec3697ae8d25962ea
ending github query because we found starting sha: 734a0d0aa0d5cab60f94f6d0c6a014dae12915f1
get_commits_between returning commits, len: 10
  commit[0] 2023-08-25: Auto merge of #115202 - ouz-a:more_smir, r=spastorino
  commit[1] 2023-08-25: Auto merge of #115221 - compiler-errors:walk-path, r=estebank
  commit[2] 2023-08-26: Auto merge of #115211 - spastorino:add-missing-smir-generics-of, r=compiler-errors
  commit[3] 2023-08-26: Auto merge of #115228 - saethlin:is-interrupted, r=thomcc
  commit[4] 2023-08-26: Auto merge of #115236 - scottmcm:less-vector, r=compiler-errors
  commit[5] 2023-08-26: Auto merge of #115232 - wesleywiser:revert_114643, r=tmiasko
  commit[6] 2023-08-26: Auto merge of #115243 - weihanglo:update-cargo, r=weihanglo
  commit[7] 2023-08-26: Auto merge of #115246 - matthiaskrgr:rollup-zdiw9gt, r=matthiaskrgr
  commit[8] 2023-08-26: Auto merge of #115198 - Zoxc:query-panic-wait, r=cjgillot
  commit[9] 2023-08-26: Auto merge of #115224 - spastorino:remove-lub_empty, r=lcnr
ERROR: no CI builds available between 734a0d0aa0d5cab60f94f6d0c6a014dae12915f1 and 69e97df5ce571a777acd654ec3697ae8d25962ea within last 167 days

oh well… this might be because I use an array…

most likely that regression point was related to

Stop emitting non-power-of-two vectors in (non-portable-SIMD) codegen #115236

because the regression was also mismatching between nightly and stable, suggesting a back-port…

Ah yes, with a tuple instead of an array I do get nightly-2023-07-16 as well :-)

I looked at LLVM IR, and realized I can't really read it myself, but I can at least share it here to save other folks some trouble:
https://godbolt.org/z/6znK7E3Md (original example, 1.72.0 - 1.73.0 )
https://godbolt.org/z/bjcEGn8s1 (slightly modified steffahn's [i8;12] reduction, 1.69.0 - 1.70.0)

@rustbot

WG-prioritization assigning priority (Zulip discussion).

@rustbot label -I-prioritize +P-critical

Add a test case for rust-lang#137646

@saethlin

#137669 should fix this.

@saethlin Free to review if you would like.

Auto merge of rust-lang#137669 - DianQK:fn-atts-virtual, r=saethlin

Add a test case for rust-lang#137646

#138190 has been merged.

Navigation Menu

Repository navigation

Miscompilation caused by incorrectly-deduced readonly on virtual call #137646

Meta

11 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Miscompilation caused by incorrectly-deduced readonly on virtual call #137646

Description

Meta

Activity

steffahn commented on Feb 26, 2025

theemathas commented on Feb 26, 2025

steffahn commented on Feb 26, 2025

moxian commented on Feb 26, 2025

steffahn commented on Feb 26, 2025

steffahn commented on Feb 26, 2025

moxian commented on Feb 26, 2025

steffahn commented on Feb 26, 2025

moxian commented on Feb 26, 2025

11 remaining items

apiraino commented on Feb 26, 2025

dianqk commented on Feb 26, 2025

dianqk commented on Mar 12, 2025

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions