You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
segfault during stage2 build of dist x86_64-apple-darwin xcode 7.3 #48866
Both #48757 (log) and #48388 (log) failed building the first two crates of cargo stage2 in parallel, receiving a SIGSEGV on both the rustc processes. The sets of crates failing are different across the builds.
From the stack trace it is obviously caused by the CACHE TLS. The segfault happens during initialization of the CACHE i.e. when the HashMap was first created.
Activity
kennytm commentedon Mar 9, 2018
Interestingly this seems to be quite deterministic on #48326. The bug happened twice so far and only on that PR (besides the miri PR).
Auto merge of #48868 - kennytm:debug-48866, r=<try>
emilyalbini commentedon Mar 9, 2018
Also happened on #48757
emilyalbini commentedon Mar 9, 2018
Both #48757 (log) and #48388 (log) failed building the first two crates of cargo stage2 in parallel, receiving a SIGSEGV on both the rustc processes. The sets of crates failing are different across the builds.
kennytm commentedon Mar 9, 2018
Reproduced on Travis and the offending artifact should have been uploaded to S3.
Edit: That artifacts segfaults locally.
Details
kennytm commentedon Mar 9, 2018
@oli-obk
From the stack trace it is obviously caused by the
CACHETLS. The segfault happens during initialization of theCACHEi.e. when the HashMap was first created.The instruction throwing the error is
where
raxis not aligned on the 16-byte boundary. Thus the segfault is caused by unaligned access in themovdqainstruction.The
raxshould be the address of the CACHE. I don't know why it needs to be aligned, and why it isn't aligned in stage2.If you've got a Mac you may try to debug it yourself. There's no line info though.
Also, we may be switching to Xcode 9 for dist in #48891, and that bug might go away naturally 🤔
kennytm commentedon Mar 10, 2018
Now #48891 has been merged, if we don't see any more errors like this next week we may close this issue as no-longer-reproducible.
oli-obk commentedon Mar 10, 2018
I'll still check why it isn't aligned. This is something I can check on Linux
oli-obk commentedon Mar 22, 2018
I give up. There's absolutely no reason the address of
CACHEshould ever be stored without alignment.alexcrichton commentedon Apr 10, 2018
I think this was fixed with the xcode 9 updates, so presuambly this was a linker bug?
Auto merge of #51828 - kennytm:no-simd-swap-for-mac, r=<try>
Rollup merge of rust-lang#51828 - kennytm:no-simd-swap-for-mac, r=ale…