LLVM assertion error on release-0.6 while building docs #24855
Labels
help wanted
Indicates that a maintainer wants help on an issue or pull request
priority
This should be addressed urgently
Milestone
Comments
|
Possibly caused by Lines 2092 to 2093 in 3835d80 This code is intended to trigger runtime UB, but according to Keno, it also triggers LLVM bugs, so on master, he replaced it with a large performance regression (https://github.com/JuliaLang/julia/blame/c1cf539c6cf6915f2870fcddbbe7aeb47a0e11ba/src/cgutils.cpp#L2214 - the comment on that line is wrong now btw). |
|
I can't reproduce with a stock build nor with a build with these options: But it happens on every Travis Linux build on release-0.6. :/ |
|
I assume this is blocking the release of 0.6.2, so any help would be much appreciated. |
ararslan
added
the
help wanted
|
Did the release-0.6 branch get any patches applied to lllvm in the cache at some point? Maybe reset that and try again? |
|
I'm not sure how to do that. Do you have any pointers, @vtjnash? |
|
Just click the trash can icon https://travis-ci.org/JuliaLang/julia/caches |
|
It looks like I don't have the necessary permissions to do that. |
|
@ararslan should be tied to the commit bit, are you signed in on Travis? |
|
Ah yeah, I was getting confused by redirects. I've deleted the cache and restarted the most recent build on release-0.6, so we'll see what happens. |
|
Seems not to have helped: https://travis-ci.org/JuliaLang/julia/builds/309339079. Same error. |
|
Had a quick look with the archive @ararslan put on anubis. A better backtrace: However, lacking EDIT: OK, managed to get a hold of the function that makes mem2reg crash (doing a manual |
|
MWE after bugpoint: ; ModuleID = 'bugpoint-reduced-simplified.bc'
source_filename = "bugpoint-output-59d46bf.bc"
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
; Function Attrs: argmemonly nounwind
declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture writeonly, i8* nocapture readonly, i32, i32, i1) #0
define void @julia_push() {
top:
%item = alloca i8
call void @llvm.memcpy.p0i8.p0i8.i32(i8* %item, i8* %item, i32 1, i32 1, i1 false)
unreachable
}
attributes #0 = { argmemonly nounwind }Strangely only crashes on a build from |
|
As expected, bisected to 820ec5f #23750. Note this is the first bad commit, and it shouldn't be part of the LLVM version as used by the So this is probably the same issue as #24173. With respect to the crash, @Keno will be able to judge whether this is a problem with the patch from #23750, or whether the MWE is invalid and cannot get generated post-#23750. |
|
Yeah, so that IR as illegal and we don't generate it after #23750, though as @vtjnash will point out to me, it should not crash the compiler (it may still crash your program at runtime though). In either case the release-0.6 version of LLVM should not have that patch, so that's the bug we should look at. |
|
closed by #24953 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Travis is triggering an LLVM assertions on release-0.6 while building the docs: https://travis-ci.org/JuliaLang/julia/jobs/309198066
I haven't managed to reproduce that locally. cc: @ararslan
The text was updated successfully, but these errors were encountered: