Segfault when using v0.1.3

[jakebolewski]

Here is a representative stack trace:
https://buildkite.com/clima/climacore-ci/builds/183#73896f51-3ec9-4099-a226-3217264ed813/96-100

Culprit looks to be a mis-sized buffer?

TriangularSolve.jl/src/TriangularSolve.jl

Lines 288 to 290 in 08ca19b

	L > length(buff) && resize!(buff, L)
	ptr = Base.unsafe_convert(Ptr{T}, buff)
	StridedPointer{T,2,1,0,(1,2)}(ptr, (VectorizationBase.static_sizeof(T), RSUF), (StaticInt(0),StaticInt(0)))

This also might be resolved with the latest release that adjusted the compat bounds with VectorizationBase

[chriselrod]

The buffer looks like it is sized correctly (N * register_size() * UF).

The error says

Unreachable reached at 0x2b0f7232a192
--


signal (4): Illegal instruction

which doesn't look memory related, although memory corruption can cause unpredictable results.

Let me know if you can get an example, or if you can try with 0.1.5.
0.1.5 did not change anything related to the buffer.

[jakebolewski]

Looks like updating / pinning to v1.5 fixed the issue, (which also had impacts with dependee package versions so it's difficult to exactly pin down what resolved it).

[chriselrod]

Hmm. Okay. Let me know if this comes up again on the latest version.

There were recent segfault issues with StrideArraysCore 0.2.0 that have since been fixed in 0.2.1, which the latest Polyester requires. However, TriangularSolve 0.1.3 requires Polyester 0.3, which is limited to StrideArraysCore 0.1.
There're have been other Polyester segfault issues, but all the known ones have been fixed in the latest from the 0.3 and 0.4 series, and they tended to involve lots of big structs in the loops, which shouldn't be the case in TriangularSolve or RecursiveFactorization.

[jakebolewski]

Thanks @chriselrod, I'll keep this in mind going forward if we see this coming up again.