Julia: [v1.4] "LLVM ERROR: Symbols not found:" when running AccurateArithmetic tests
This is not an MWE but I want to get the ball rolling. Running the test suite of ~AstroBase.jl~ AccurateArithmetic.jl fails with this as the only output:
LLVM ERROR: Symbols not found: { _llvm.experimental.vector.reduce.v2.fadd.f64.v4f64 }
Steps to reproduce
~- Clone AstroBase.jl and switch to branch he/refactor~
- Run tests of AccurateArithmetic.jl
Additional Info
- On v1.3 the tests fail because of some
UndefVarErrors. That is the expected behavior.
- Commenting out those offending tests does not help on v1.4
- The culprit seems to be AccurateArithmetic.jl
CC: @chriselrod, @ffevotte, @JeffreySarnoff
helgee
All 6 comments
I can reproduce this, both on my machine and on travis.
What seems strange to me is that tests pass with julia nightly. See for example the last travis test results:
ffevotte
on 22 Mar 2020
File a bug report with SIMDPirates.jl (the v2 experimental reductions are only available on LLVM 9 and later).
KristofferC
on 22 Mar 2020
But how come everything works with 1.3 and nightly, but not 1.4?
ffevotte
on 22 Mar 2020
Because it is a bug in SIMDPirates.jl... Here, the reason is that different LLVM versions are used in all of those releases.
KristofferC
on 22 Mar 2020
File a bug report with SIMDPirates.jl (the
v2experimental reductions are only available on LLVM 9 and later).
Done and PR: https://github.com/chriselrod/SIMDPirates.jl/pull/6
helgee
on 22 Mar 2020
That was my mistake. I checked the LLVM 8 language reference, found experimental reduce fadd, and didn't look closely enough to see that it was different from the v2, two-argument, variants present in LLVM 9 and beyond.
Given that AccurateArithmetic has had a few SIMDPirates-related issues before, and SIMDPirates didn't have a test suite, this is its new runtests.jl. A bad solution, but still an improvement.
I normally run LoopVectorization's tests on my own computer after every change, but these take a long time on Travis.
They also allowed this (an LLVM 8-bug) to slip through the cracks.
chriselrod
on 23 Mar 2020
Related issues
sbromberger
路
3Comments
manor
路
3Comments
iamed2
路
3Comments
wilburtownsend
路
3Comments
helgee
路
3Comments
Most helpful comment
That was my mistake. I checked the LLVM 8 language reference, found experimental reduce fadd, and didn't look closely enough to see that it was different from the v2, two-argument, variants present in LLVM 9 and beyond.
Given that AccurateArithmetic has had a few SIMDPirates-related issues before, and SIMDPirates didn't have a test suite, this is its new
runtests.jl. A bad solution, but still an improvement.I normally run LoopVectorization's tests on my own computer after every change, but these take a long time on Travis.
They also allowed this (an LLVM 8-bug) to slip through the cracks.