Use versioned libblastrampoline and GMP

[gbaraldi]

Basically the title.

[staticfloat]

Very interesting; both win32 and win64 died with approximately the same error. Any ideas as to why?

[staticfloat]

It's minimally reproducible with the following:

julia> using LinearAlgebra, SparseArrays
       using SparseArrays: fixed
       b = sprandn(10, 10, 0.99) + I
       a = fixed(b)
       lu(a)

Please submit a bug report with steps to reproduce this fault, and any error messages that follow (in their entirety). Thanks.
Exception: EXCEPTION_ACCESS_VIOLATION at 0x0 -- unknown function (ip: 0000000000000000)
in expression starting at REPL[6]:5
unknown function (ip: 0000000000000000)
Allocations: 1954934 (Pool: 1953138; Big: 1796); GC: 3

[gbaraldi]

I imagine it's the blas trampoline change, because the GMP change works in the mimalloc PR.

[giordano]

It's minimally reproducible with the following:

JuliaLinearAlgebra/libblastrampoline#89

[staticfloat]

That looks quite probable:

julia> println.(filter(x -> occursin("blast", x), Libdl.dllist()))
C:\Users\Administrator\AppData\Local\Temp\julia-0f526b01f8\bin\libblastrampoline-5.dll
C:\Users\Administrator\AppData\Local\Temp\julia-0f526b01f8\bin\libblastrampoline.dll

[giordano]

So there is still something pulling in the wrong library.

[staticfloat]

It's SuiteSparse:

$ for f in *.dll; do if [[ -n $(objdump.exe -p $f | grep "libblastrampoline.dll") ]]; then echo $f; fi; done
libblastrampoline.dll
libcholmod.dll
libspqr.dll
libumfpack.dll

[vchuravy]

Can we pull in the version from the buildsystem? See base/Makefile the build_h.jl generation

[giordano]

@gbaraldi bump

[gbaraldi]

I'm not sure, maybe we can pull the version from the jll from inside the makefile? It's kind of annoying not being able to depend on the jll proper here.

[gbaraldi]

In fact, is there a reason for BigNums being part of base instead of being stdlibs?

[giordano]

They're used in Base, no?

[staticfloat]

They’re needed quite early because we can parse literal bignums

[staticfloat]

We can get version numbers from the makefile system, then embed those into build_h.jl. I suggest we try some awk trickery to figure out the so major version number.

[vchuravy]

One minor nit we should move the so version into a place where packagers can override them, and it would be great to just have a single place to change them

[giordano]

SuiteSparse is still failing on Windows despite the update to the new build 😕

[giordano]

sandbox:${WORKSPACE} # for f in bin/*.dll; do if [[ -n $(${target}-objdump -p $f | grep "libblastrampoline.dll") ]]; then echo $f; fi; done
bin/libblastrampoline.dll
sandbox:${WORKSPACE} # for f in bin/*.dll; do if [[ -n $(${target}-objdump -p $f | grep "libblastrampoline-5.dll") ]]; then echo $f; fi; done
bin/libblastrampoline-5.dll
sandbox:${WORKSPACE} # for f in bin/*.dll; do if [[ -n $(${target}-objdump -p $f | grep "libblastrampoline-5-2-0.dll") ]]; then echo $f; fi; done
bin/libblastrampoline-5-2-0.dll
bin/libblastrampoline-5.dll
bin/libblastrampoline.dll
bin/libcholmod.dll
bin/libspqr.dll
bin/libumfpack.dll

So now we're linking to libblastrampoline-5-2-0.dll, which is even worse than libblastrampoline.dll

[giordano]

All Windows tests are green 🥳