Merge 0cecafd647cc into 10.0.1 #45676
Description
| Bugzilla Link | 46331 |
| Resolution | FIXED |
| Resolved on | Jun 23, 2020 16:06 |
| Version | 10.0 |
| OS | All |
| Blocks | #44654 |
| CC | @alinas,@emaste,@zmodem,@hfinkel,@tstellar |
| Fixed by commit(s) | 0cecafd caa755f |
Extended Description
Please merge https://reviews.llvm.org/rG0cecafd647cc into 10.0.1. As described in https://bugs.freebsd.org/246630, clang's new -fintegrated-cc1 default (introduced in https://reviews.llvm.org/rGc4a134a5107c and some follow-up commits) appears to have uncovered non-determinism in the assembly output. This is further described in https://reviews.llvm.org/D69825, the original review for
https://reviews.llvm.org/rGc4a134a5107c.
In FreeBSD I have worked around the problem temporarily, by setting CLANG_SPAWN_CC1 back to 1, so a fresh clang -cc1 process is executed.
But after some more testing and bisecting, I found out that the non-determinism is actually fixed by https://reviews.llvm.org/rG0cecafd647cc ("[BasicAA] Make BasicAA a cfg pass"). In https://reviews.llvm.org/D69825#2091632 I've asked @asbirlea if this was an intended side-effect of the commit, or if there is something different going on.
In any case, this bug should be solved one way or another in 10.0.1, as non-determinism is obviously a blocker.