Add known-bug tests for all(?) I-unsound issues

[jackh726]

I'll come back at some point and create an actual list. However, we should aim to add known-bug tests (or work to finding MCVEs) for all the I-unsound issues.

The list of issues remaining: https://github.com/rust-lang/rust/issues?q=is%3Aopen+is%3Aissue+label%3AI-unsound+-label%3AS-bug-has-test+

The list of completed issues: https://github.com/rust-lang/rust/issues?q=is%3Aopen+is%3Aissue+label%3AI-unsound+label%3AS-bug-has-test+

[sladyn98]

@jackh726 I would like to give this a shot, any hints on where I could begin

[jackh726]

@sladyn98 sure!

Okay, so we have a concept of known-bug tests. Here's an example: https://github.com/rust-lang/rust/blob/18a6d911caba59605eb03db1452848a85d2e5879/tests/ui/coherence/coherence-overlap-negative-impls.rs

Note the // known-bug: #101350 at the top. There is also // check-pass, which means that this test currently passes under during check (though that test shouldn't pass, hence the known-bug)

The goal for this issue is to go through the open I-unsound issues (https://github.com/rust-lang/rust/issues?page=1&q=is%3Aopen+is%3Aissue+label%3AI-unsound), identify a minimal reproduction that shouldn't pass, but does, and add a known-bug test for each one.

Nearly all of the issues should have a minimal repro in the comments somewhere, so I would start with those. Tests should go in src/test/ui in an appropriate directory. Test names should likely be meaningful compared to the open issue. Any comments at the top of the file briefly explaining the problem would also be helpful.

Let me know if that helps!

[sladyn98]

#105084
I found this issue, and planning to a PR with a known test bug

[KenCox94]

is there anything a new contributor can do for this? any help still needed @sladyn98 @jackh726

[jackh726]

My previous comment still applies; there haven't been any PRs opened for this.

If anything is confusing, please let me know.

[gburgessiv]

If it helps, the listing I was able to scrape is:

• Collisions in type_id #10389 (maybe a no-go? we'd need to test for TypeId collisions)
• Investigate how LLVM deals with huge stack frames #18072
• Implied bounds on nested references + variance = soundness hole #25860
• set_var/remove_var are unsound in combination with C code that reads the environment #27970
• #[no_mangle] is unsafe in the presence of name collisions #28179
• Tracking issue for specialization (RFC 1210) #31844
• Specialization and lifetime dispatch #40582
• Extend stack probe support to non-tier-1 platforms, and clarify policy for mitigating LLVM-dependent unsafety #43241
• ill-typed unused FFI declarations can cause UB #46188
• Panics in destructors can cause the return value to be leaked #47949
• Constants can contain references that are not Sync #49206
• Keeping references to #[thread_local] statics is allowed across yields. #49682
• Trait objects can call nonexistent concrete methods #50781
• Abort instead of unwinding past FFI functions #52652
• repr(simd) is unsound in C FFI #53346
• Coherence can be bypassed by an indirect impl for a trait object #57893
• Resolve unsound interaction between noalias and self-referential data (incl. generators, async fn) #63818
• -C target-feature/-C target-cpu are unsound #64609
• std::process::Command doesn't follow Unix signal safety in forked child #64718
• Switching to opt-level=z on i686-windows-msvc triggers STATUS_ACCESS_VIOLATION #67497
• Pin is unsound due to transitive effects of CoerceUnsized #68015
• Statics don't support alignments larger than the page size #70022
• Locals aligned to greater than page size can cause unsound behavior #70143
• unsized_locals fails to uphold alignment requirements #71416
• Comparing to infinity is buggy on x87 #72327
• error: could not compile gkrust since Rust 1.43 on SPARC Solaris #74551
• negative_impls and auto_traits experimental features allow trait impls to overlap #74629
• AVR: Miscompilation with Trait Object in Option #74743
• #[link_section] is unsound on Harvard architectures #76507
• With min_specialization enabled, an incomplete impl for a non-static type will delegate method calls to a less-specific impl with a 'static bound #79457
• Miscompilation of AVX2 code under --release #79865
• rustc has wrong signature for C function with 16-byte aligned stack argument in x86_64 Linux #80127
• Unsoundness in type checking of trait impls. Differences in implied lifetime bounds are not considered. #80176
• Accessing DST field of packed struct calculates the wrong field address #80925
• repr(C) on MSVC targets does not always match MSVC type layout when ZST are involved #81996
• Regressions with large (2-4GB) stack arrays on large stacks #83060
• 'static closures/FnDefs/futures with non-'static return type are unsound #84366
• Functions, closures, and HRTB-trait-objects can implement traits such that validity of associated types is never checked. #84533
• HRTB on subtrait unsoundly provides HTRB on supertrait with weaker implied bounds #84591
• auto trait candidate selection is unsound #84857
• A Pin unsoundness involving an impl DerefMut for Pin<&dyn LocalTrait> #85099
• SpecExtend for TrustedLen is unsound #89948
• TypeId exposes equality-by-subtyping vs normal-form-syntactic-equality unsoundness. #97156
• Unsoundness due to where clauses not checked for well-formedness #98117
• orphan check incorrectly handles projections #99554
• Normalization can skip WF #100041
• implied bounds from projections in impl header can be unsound #100051
• LLVM miscompiles large stack allocations #100914
• Incorrect handling of lateout pairs in inline asm #101346
• relating projection substs is unsound during coherence #102048
• extern "C" functions don't generate the same IR definitions as clang on x86, causing problems with cross-language LTO #102174
• prevent negative impl cycles #102678
• function type params are not checked for well-formedness #104005
• Box syntax and generator_clone can lead to double free #105084
• Arguments of an impl method may have stronger implied bounds than trait method  #105295
• specialization: default items completely drop candidates instead of ambiguity #105782
• occurs check with aliases results in error, not ambiguity #105787
• Miscompilation: Equal pointers comparing as unequal #107975
• TAIT shouldn't constrain impl generics #108425

rging for the list of I-Unsound issues makes it seem that there're no tests for any of these ATM, so I left them unpopulated. I added a few tests while I was in the area, which I'll build a PR for now. :)

[KisaragiEffective]

NOTE: #105084 is already added

[whtahy]

Big thanks to @gburgessiv for the big checklist! Very useful.

@jackh726 Thanks for adding the test labels and updating the issue. Looking through the remaining issues, I'm a bit lost for most of these.

I'm comfortable adding tests where the mcve compiles but shouldn't, but I'm wondering what to do for mcves that don't quite follow this pattern or have other variations: multiple crates, llvm/asm/etc, panic/warning/error, etc.

Below, I've organized the remaining issues as I see them, hopefully useful for other contributors as well. (Currently only 44 issues in my list vs 55 actual, so missing 11 issues as of 2023.04.28)

Let me know if there's anything else sort of bookkeeping that would be useful here.

1. Not able to find an mcve for these 7 issues:

issue	name
27970	set_var/remove_var are unsound in combination with C code that reads the environment
43241	Extend stack probe support to non-tier-1 platforms, and clarify policy for mitigating LLVM-dependent unsafety
52652	Abort instead of unwinding past FFI functions
64718	std::process::Command doesn't follow Unix signal safety in forked child
67497	Switching to opt-level=z on i686-windows-msvc triggers STATUS_ACCESS_VIOLATION
68015	Pin is unsound due to transitive effects of CoerceUnsized
74551	error: could not compile gkrust since Rust 1.43 on SPARC Solaris

2. Tracking issue needs test?

Wondering if the specialization tracking issue (issue 31844) needs a test, since it's sort of multiple issues in 1 issue.

3. These 4 issues already include tests on master:

issue	name	test	test status
50781	Trait objects can call nonexistent concrete methods	tests/ui/issues/issue-50781.rs	test errors and includes .stderr
97156	TypeId exposes equality-by-subtyping vs normal-form-syntactic-equality unsoundness.	tests/ui/const-generics/generic_const_exprs/typeid-equality-by-subtyping.rs	test includes known-bug and check-pass comments
102048	relating projection substs is unsound during coherence	tests/ui/traits/new-solver/coherence/issue-102048.rs	test compiles but includes .stderr
105084	Box syntax and generator_clone can lead to double free	tests/ui/generator/issue-105084.rs	test includes known-bug and check-pass comments

4. These 2 issues have an mcve, but don't work as is on playground:

issue	name	mcve	error	needs?
74743	AVR: Miscompilation with Trait Object in Option	#74743 (comment)	error: unknown feature llvm_asm	llvm_asm! macro
76507	#[link_section] is unsound on Harvard architectures	#76507 (comment)	error: expected expression, found ;	serial output

5. These 3 issues may need regression tests instead?

issue	name	mcve	status
79457	With min_specialization enabled, an incomplete impl for a non-static type will delegate method calls to a less-specific impl with a 'static bound	#79457 (comment)	fixed w/ error? #79457 (comment)
79865	Miscompilation of AVX2 code under --release	#79865 (comment)	fixed by llvm 14 upgrade: #79865 (comment)
80127	rustc has wrong signature for C function with 16-byte aligned stack argument in x86_64 Linux	#80127 (comment)	wip PR 103830 includes regression test: #103830 (comment)

6. Not sure how to test: These 11 issues involve llvm, asm, C, etc

issue	name	mcve	needs?
18072	Investigate how LLVM deals with huge stack frames	#18072 (comment)	llvm
46188	ill-typed unused FFI declarations can cause UB	#46188 (comment)	llvm
53346	repr(simd) is unsound in C FFI	#53346 (comment)	UB caused by C vs Rust ABI mismatch
63818	Resolve unsound interaction between noalias and self-referential data (incl. generators, async fn)	#63818 (comment)	miri
64609	-C target-feature/-C target-cpu are unsound	#64609 (comment)	multiple crates, rustflags, ABI
70022	Statics don't support alignments larger than the page size	#70022 (comment)	platform specific, llvm
70143	Locals aligned to greater than page size can cause unsound behavior	#70143 (comment)	platform specific, llvm
72327	Comparing to infinity is buggy on x87	#72327 (comment)	platform specific, asm
81996	repr(C) is unsound on MSVC targets	#81996 (comment)	#81996 (comment)
101346	Incorrect handling of lateout pairs in inline asm	#101346 (comment)	asm
102174	extern "C" functions don't generate the same IR definitions as clang on x86, causing problems with cross-language LTO	#102174 (comment)	C, asm

7. Not sure how to test: These 10 issues error, segfault, or panic:

issue	name	mcve	test outcome
10389	Collisions in type_id	#10389 (comment)	panic: Option::unwrap on None
47949	Panics in destructors can cause the return value to be leaked	#47949 (comment)	explicit panic
71416	unsized_locals fails to uphold alignment requirements	#71416 (comment)	fails left == right test
80176	Unsoundness in type checking of trait impls. Differences in implied lifetime bounds are not considered.	#80176 (comment)	error: #[deny(implied_bounds_entailment)] on by default
83060	Regressions with large (2-4GB) stack arrays on large stacks	#83060 (comment)	panics in playground: Result::unwrap() on Err or stack overflow
84857	auto trait candidate selection is unsound	#84857 (comment)	segfault and warning: cross-crate traits with a default impl, like UnwindSafe, should not be specialized
89948	SpecExtend for TrustedLen is unsound	#89948 (comment)	segfault, but maybe not bug? #89948 (comment)
100914	Double free on Linux with stable toolchain	#100914 (comment)	panics in playground: Result::unwrap() on Err or stack overflow. (related to 83060)
102678	prevent negative impl cycles	#102678 (comment)	should compile but errors
105295	Seg fault in Rust 1.65.0 if I don't create temporary variable	#105295 (comment)	error: #[deny(implied_bounds_entailment)] on by default

8. Not sure how to test: These 2 mcves split code across multiple crates

issue	name	mcve
28179	#[no_mangle] is unsafe	#28179 (comment)
99554	orphan check incorrectly handles projections	#99554 (comment)

9. This mcve compiles, but not sure if it's supposed to error?

issue	name	mcve
80925	Accessing DST field of packed struct calculates the wrong field address	#80925 (comment)

10. I think I'm good to PR these 4 issues: Merged PR 110878

issue	name	mcve
49682	Keeping references to #[thread_local] statics is allowed across yields.	#49682 (comment)
40582	Specialization and lifetime dispatch	#40582 (comment)
74629	negative_impls and optin_builtin_traits nightly-features allow trait impls to overlap	#74629 (comment)
105782	specialization: default items completely drop candidates instead of ambiguity	#105782 (comment)

11. These 3 issues are part of @gburgessiv's PR 108445:

issue	name
105787	occurs check with projections results in error, not ambiguity
107975	Miscompilation: Equal pointers comparing as unequal
108425	TAIT shouldn't constrain impl generics

[J-Kappes]

Okay, so we have a concept of known-bug tests. Here's an example:

That link is dead.

I'm also not sure I understand what to do with the MCVEs. Are they to be used for writing the tests?

[jackh726]

That link is dead.

Fixed

I'm also not sure I understand what to do with the MCVEs. Are they to be used for writing the tests?

The test is essentially the MCVE, with some comments describing the issue.

[YpeKingma]

For a somewhat different approach, there is a pull request for a clippy lint for three of these issues #25860 #84591 and #100051 :

rust-lang/rust-clippy#12471

The lint looks for nested references and suggests to add a lifetimes bound that is implied by the nested reference.
After adding this bound, the compiler currently gives an error message.
Since the added bound is implied, the error message should have been given before adding it.

[jackh726]

I'm going to go ahead and close this. While there are still many issues (75 as of writing this!) that don't have a known-bug test, I don't expect keeping this issue open to meaningful make progress on reducing that number.

The remaining relevant issues are mixed in the reasoning for why they don't have a known-bug test, but it's most often because either 1) there is no MCVE because the unsoundness is either "theoretical" or just hard to test, or 2) the unsoundness is only for some select targets.

Thanks all who've helped make progress on this though!