coreos-installer test segfaults on s390x-unknown-linux-gnu

[cuviper]

xref: https://bugzilla.redhat.com/show_bug.cgi?id=1883457

In the Fedora build of coreos-installer 0.7.0, one of the tests crashed with SIGSEGV on s390x, while it passed on every other architecture. The reporter in Red Hat bugzilla found that it happens equally with system LLVM 11 or 10. The crate's Cargo.toml also has lto = true, but they didn't see any improvement after turning that off.

The distro builds run tests in release mode (to avoid recompiling from the main build) with flags in .cargo/config:

[build]
rustc = "/usr/bin/rustc"
rustdoc = "/usr/bin/rustdoc"
rustflags = ["-Copt-level=3", "-Cdebuginfo=2", "-Clink-arg=-Wl,-z,relro,-z,now", "-Ccodegen-units=1", "--cap-lints=warn", ]

The tests pass if we don't set -Ccodegen-units=1, which suggests to me that some relevant (mis)optimization is only triggered with the full unit.

The tests are also fine with Fedora's rust-1.45.2, only crashing on rust-1.46.0. I also reproduced the crash on rustup's stable 1.46.0, and I used cargo bisect-rustc to narrow down to nightly-2020-07-05 passing, while nightly-2020-07-06 crashes. That's 0cd7ff7...2753fab, which is just #73879 (cc @ecstatic-morse). I don't know if codegen-units would affect that, or if it's just triggering something new in LLVM.

Here's the reporter's backtrace:

#0  0x000002aa0c3616c4 in core::ptr::drop_in_place () at /builddir/build/BUILD/rustc-1.46.0-src/src/libcore/ptr/mod.rs:184
#1  core::ptr::drop_in_place () at /builddir/build/BUILD/rustc-1.46.0-src/src/libcore/ptr/mod.rs:184
#2  core::ptr::drop_in_place () at /builddir/build/BUILD/rustc-1.46.0-src/src/libcore/ptr/mod.rs:184
#3  core::ptr::drop_in_place () at /builddir/build/BUILD/rustc-1.46.0-src/src/libcore/ptr/mod.rs:184
#4  core::ptr::drop_in_place () at /builddir/build/BUILD/rustc-1.46.0-src/src/libcore/ptr/mod.rs:184
#5  <libcoreinst::source::FileLocation as libcoreinst::source::ImageLocation>::sources (self=0x3ffcc3fc4f8) at src/source.rs:135
#6  libcoreinst::download::tests::test_write_image_limit () at src/download.rs:489

5. https://github.com/coreos/coreos-installer/blob/79b59763bd17afba9ac6b13dfdfef401e2d9bde9/src/source.rs#L135
6. https://github.com/coreos/coreos-installer/blob/79b59763bd17afba9ac6b13dfdfef401e2d9bde9/src/download.rs#L489

[ecstatic-morse]

cc #74551, another miscompilation on a tier 2 architecture that appeared after a similar optimization. If the root cause is the same, we should have a better chance of diagnosing it now, since it occurs in a crate's test case instead of while compiling Firefox. cc @pnkfelix as well.

[apiraino]

Assigning P-critical as discussed as part of the Prioritization Working Group procedure and removing I-prioritize.