SIGSEGV in rustc upon compiling with --release #119179
Comments
rustbot
added
the
needs-triage
nikic
added
I-crash
This comment was marked as outdated.
This comment was marked as outdated.
|
Reduced: ; RUN: opt -passes="loop-mssa(loop-instsimplify,simple-loop-unswitch<nontrivial>)" < %s
@vtable = constant ptr @foo
declare void @foo() memory(none)
define void @test(i1 %c) {
br label %loop
loop:
%fn = load ptr, ptr @vtable, align 8
call void %fn()
br i1 %c, label %exit, label %loop
exit:
ret void
} |
|
Upstream fix: llvm/llvm-project#76142 |
|
@cfstras, could you please try adding edit: Ah nikic might have already found a more specific issue, probably nevermind :) |
|
I verified successfully that this also fixes the build for my codebase where I originally found this 👍🏻 (by cherry-picking only your commit on top of the currently used nightly llvm branch in rust-lang/rust, 606bc11367b4) edit @jamesmunns: Sorry, just saw your comment as I was leaving. I can try that later, but since it also reproduces on the (linux/ guessing amd64?) rust playground, I doubt this is ARM-related. |
|
@jamesmunns: The argument doesn't seem to make a difference, neither on stable/beta, nor on nightly. |
nikic
added
the
llvm-fixed-upstream
|
I'm confused, this code still crashes with yesterday's nightly? Could it be that the alternative fix you selected doesn't fully fix the issue I'm seeing? |
|
I had also tested my code using your cherry-picked fix (llvm/llvm-project#76819) based on llvm 606bc11367b4 again, and saw the crash still happen. Apologies, I seemingly had written the comment but forgot to send. |
|
Hm yeah, I tested my original fix against the original test case, but not the final version. Let me take another look. |
|
New reduction: ; RUN: opt -passes="loop-mssa(loop-instsimplify,simple-loop-unswitch<nontrivial>)" < %s
@vtable = constant ptr @foo
declare void @foo() memory(none)
define void @test(i1 %c) {
br label %loop
loop:
%fn = load ptr, ptr @vtable, align 8
call void %fn()
br i1 %c, label %exit, label %loop
exit:
ret void
}Took me a while to spot how this differs from the previous one: The branch destinations are swapped... |
|
New upstream fix: llvm/llvm-project#78272 |
|
When cherry-picking this onto latest rust master (fa0dc20, llvm 919a630125a0), I get: Is this expected? |
|
@cfstras Are you sure the commit was picked up? The patch removes that assertion. |
|
Reopening, the auto closing on a merge in @saethlin's fork seems unintentional. |
|
🤦 yes, unintentional. If someone has a way to prevent that I'd appreciate it. |
|
@nikic duh, forgot to commit the llvm submodule, thanks. Confirmed ✅ no crash, also builds my original project without issue. |
|
Fixed in beta & nightly. |
I tried compiling this code with
--release:(Thanks to @Xiretza for helping me find this shorter reproducer)
I expected to see this happen: (nothing, since there's no main method) a clean compile
Instead, this happened:
(Or, when compiling through
cargo build --release:)Meta
rustc --version --verbose:Note: this crashes on the playground on nightly (2023-12-20 5ac4c8a), beta (1.75.0-beta.7), and stable (1.74.1). Also crashes on aarch64-apple-darwin nightly (2023-12-19 f704f3b), beta (1.75.0-beta.7) and stable (1.74.1)
Backtrace
(From Playground since I don't have a debug build on macOS yet)
The text was updated successfully, but these errors were encountered: