track_caller on panic_no_unwind leads to LLVM crash

[Mark-Simulacrum]

#94290 is expecting to land having introduced #[cfg_attr(bootstrap, track_caller)] on panic_no_unwind in library/core/src/panicking.rs.

We saw these errors in stage 1 std compilation in CI (with track_caller and without it):

Incorrect number of arguments passed to called function!
  tail call void @_ZN4core9panicking15panic_no_unwind17heb5bd7bc5de639ecE(%"core::panic::location::Location"* bitcast (<{ i8*, [16 x i8] }>* @10 to %"core::panic::location::Location"*)) #15
in function _ZN12panic_unwind8real_imp5panic17exception_cleanup17h7b2c07bcac8db7c5E
LLVM ERROR: Broken function found, compilation aborted!
�[0m�[0m�[1m�[31merror�[0m�[1m:�[0m could not compile `panic_unwind`
�[0m�[0m�[1m�[33mwarning�[0m�[1m:�[0m build failed, waiting for other jobs to finish...
Incorrect number of arguments passed to called function!
  call void @_ZN4core9panicking15panic_no_unwind17heb5bd7bc5de639ecE(%"panic::location::Location"* bitcast (<{ i8*, [16 x i8] }>* @364 to %"panic::location::Location"*)) #26
in function _ZN4core9panicking15panic_no_unwind17heb5bd7bc5de639ecE
LLVM ERROR: Broken function found, compilation aborted!

Incorrect number of arguments passed to called function!
  call void @_ZN4core9panicking15panic_no_unwind17h8631ddd8246c176fE() #16
in function _ZN4core9panicking15panic_no_unwind17h8631ddd8246c176fE
LLVM ERROR: Broken function found, compilation aborted!

Locally, I'm not able to reproduce these (perhaps need to enable LLVM asserts or something), but I do see the following in stage 1 std compilation when building with track_caller.

error: failed to parse bitcode for LTO module: Invalid record (Producer: 'LLVM14.0.0-rust-1.61.0-nightly' Reader: 'LLVM 14.0.0-rust-1.61.0-nightly')

error: could not compile `core` due to previous error

cc @Amanieu @nbdd0121 -- seems plausible that something in #92911 is causing this

[Amanieu]

The relevant change is here. This function used to take a caller location but this was removed to reduce code size bloat.

[Mark-Simulacrum]

I guess it seems ok for an internal function and a lang item, but I'm surprised that we don't have some proper error being emitted if the function is somehow being manually called -- that seems generally hazardous. (Should it have some special ABI declared at the Rust syntax level?)

I also note that the function still calls Location::caller: which seems useless or at least confusing without track_caller, right?

    let pi = PanicInfo::internal_constructor(Some(&fmt), Location::caller(), false);`

[Amanieu]

The panic_no_unwind function is only expected to be called by the compiler via the lang item. It is never used directly.

I also note that the function still calls Location::caller: which seems useless or at least confusing without track_caller, right?

That's the only way to construct a Location which is required for a PanicInfo.

[nbdd0121]

Oops. I think I started working on #92911 before panic_no_unwind lands on beta and I forget to add cfg bootstrap after it's promoted into beta.

[Mark-Simulacrum]

Even if it's not used directly, we should probably be adding a comment to the function about that, right? It seems like such a function should be documented as being compiler-only.

Anyway, seems OK -- I can adjust my FIXME comment to include this information.