Proc Macro Crate and Workspace Miri and Clippy Malloc Error

[KSBilodeau]

When compiling a workspace that has a proc macro crate in it, rustc emits a malloc error causing tools such as miri and clippy to fail. This bug affects nightly versions 2021-12-05+.

I tried this code:

playground_workspace.zip

I expected to see this happen:
Miri and clippy successfully assess my code

Instead, this happened:
the following crashes (I can supply the actual Mac IPS files if it would be more helpful than these txt copies):
clippy.txt
miri.txt

Meta

rustc --version --verbose:

rustc 1.59.0-nightly (23f69235a 2021-12-20)
binary: rustc
commit-hash: 23f69235ad2eb9b44ac1a55eeaa3f9b484d9de4a
commit-date: 2021-12-20
host: x86_64-apple-darwin
release: 1.59.0-nightly
LLVM version: 13.0.0

[hkratz]

Looks like it is during proc macro compilation, not execution? Shouldn't that use jemalloc instead of the system malloc?

The error message is the same as in #92173.

[ehuss]

clippy and miri are not built with jemalloc.

I think this is likely the same as #92173.

I've been trying to dig into this some. I bisected to #91505, and I've been trying to bisect within that rollup, but I'm not really finding out much. I'm still trying to go through them, it is a pretty slow process.

[ehuss]

I'm a bit at a loss. Reverting #91385 makes the problem go away (both here and my tests with #92173). However, that PR is not doing anything that should affect any of this, since it is only within an error path. So my best guess is that PR shuffles the code layout/generation just enough to trigger the underlying issue.

cc @rust-lang/clippy and @rust-lang/miri just as a heads up that clippy and miri are crashing on macOS. I haven't tested RLS, but I would imagine it might also have issues.

I would be reluctant to enable jemalloc in miri and clippy. I don't know what is causing the underlying error. However, that is a route we could take. It would involve replicating all of this in their respective main binaries. I'm having dejavu about this, but I don't remember where I last discussed this.

[RalfJung]

I assume this is the same problem as rust-lang/miri#1941

[RalfJung]

I'm a bit at a loss. Reverting #91385 makes the problem go away

At least that is consistent with the regression range from rust-lang/miri#1941.

Miri itself acts as rustc for proc-macros, but it seems (I was not aware of that) it is a different rustc than the "real" one -- it does not use jemalloc? So yeah I assume this is the same as #92173. Though I also assume that just masks the error since jemalloc doesn't do the realloc check that leads to the abort. That's my guess anyway.

[rdornin]

I just want to report that issue is currently blocking me in development. If there are any work around for the macos darwin tool chain and clippy errors... would love to hear about it.