invalid opcode regression in x86_64-unknown-linux-musl release builds while compiling code using generic-array

[SohumB]

Repro.

searched nightlies: from nightly-2024-11-21 to nightly-2025-01-24
regressed nightly: nightly-2025-01-10
searched commit range: a580b5c...8247594
regressed commit: b6b8361

bisected with cargo-bisect-rustc v0.6.9

Host triple: x86_64-unknown-linux-gnu
Reproduce with:

cargo bisect-rustc -vv --start=2024-11-21 --target=x86_64-unknown-linux-musl --script=run.sh

I definitely don't understand MIR, but I was curious enough to look at it, so in case it helps anyone else, here's what appears to be the relevant section (without the StorageLive/StorageDead calls)

succeeding

        _2 = copy (_1.0: Box<GenericArray<usize, U1>>);

        _3 = copy (((_2.0: Unique<GenericArray<usize, U1>>).0: NonNull<GenericArray<usize, U1>>).0: 
                    *const GenericArray<usize, U1>);

        _4 = *const [usize] from (copy _3, const 1_usize);
        _6 = NonNull::<[usize]> { pointer: copy _4 };
        _11 = copy _6 as *mut [usize] (Transmute);
        _10 = copy _11 as *const usize (PtrToPtr);

        _5 = NonNull::<usize> { pointer: move _10 };
        _9 = copy _5 as *mut usize (Transmute);
        _8 = Offset(copy _9, const 1_usize);
        _7 = move _8 as *const usize (PtrToPtr);
        drop(_1) -> [return: bb2, unwind continue];

failing

        _2 = copy (_1.0: Box<GenericArray<usize, U1>>);

        _3 = copy ((_2.0: Unique<GenericArray<usize, U1>>).0: NonNull<GenericArray<usize, U1>>) 
                    as *const GenericArray<usize, U1> (Transmute);

        _4 = *const [usize] from (copy _3, const 1_usize);
        _9 = copy _4 as *const usize (Transmute);

        _5 = NonNull::<usize> { pointer: move _9 };
        _8 = copy _4 as *mut usize (Transmute);
        _7 = Offset(copy _8, const 1_usize);
        _6 = move _7 as *const usize (PtrToPtr);
        drop(_1) -> [return: bb2, unwind continue];

[lqd]

(cc @scottmcm on that bisection)

[SohumB]

        _5 = NonNull::<usize> { pointer: move _9 };
        _8 = copy _4 as *mut usize (Transmute);

Is this the issue? _5 is created but then never used in the failing MIR. In the succeeding MIR, it's what is copy _5'd.

[scottmcm]

Hmm, the metadata checks should be keeping things from merging that *const [usize] to *const usize cast

rust/compiler/rustc_mir_transform/src/gvn.rs

Lines 1440 to 1452 in 8231e85

	// PtrToPtr-then-Transmute is fine so long as the pointer cast is identity:
	// `*const T -> *mut T -> NonNull<T>` is fine, but we need to check for narrowing
	// to skip things like `*const [i32] -> *const i32 -> NonNull<T>`.
	(PtrToPtr, Transmute)
	if self.pointers_have_same_metadata(inner_from, inner_to) =>
	{
	Some(Transmute)
	}
	// Similarly, for Transmute-then-PtrToPtr. Note that we need to check different
	// variables for their metadata, and thus this can't merge with the previous arm.
	(Transmute, PtrToPtr) if self.pointers_have_same_metadata(from, to) => {
	Some(Transmute)
	}

But clearly something in there is going wrong. I'll take a look.