Fix Base.libblas_name/Base.liblapack_name

[giordano]

On Windows they need to include the major soversion of libblastrampoline.

Fix JuliaLang/LinearAlgebra.jl#981.

[staticfloat]

Can we somehow mark this as deprecated and instead tell this list of packages that they should use libblastrampoline_jll.libblastrampoline instead?

[giordano]

Can you deprecate a variable (that is, with a nice warning when you try to access it)? And I'm not sure that's a future-proof option, in case we come up with yet another blas idea in the future, Base.libblas_name looks more agnostic.

[giordano]

I added a test for the use of Base.libblas_name, to avoid regressing again like in JuliaLang/LinearAlgebra.jl#981.

[giordano]

Ehr, does anyone know what would cause

      From worker 12:	Exception: EXCEPTION_ACCESS_VIOLATION at 0x5fed810 -- dot_compute at C:\buildkite-agent\builds\win2k22-amdci6-1\julialang\julia-master\julia-86e64109fb\bin\libopenblas64_.dll (unknown line)
      From worker 12:	in expression starting at C:\buildkite-agent\builds\win2k22-amdci6-1\julialang\julia-master\julia-86e64109fb\share\julia\stdlib\v1.10\LinearAlgebra\test\blas.jl:719
      From worker 12:	dot_compute at C:\buildkite-agent\builds\win2k22-amdci6-1\julialang\julia-master\julia-86e64109fb\bin\libopenblas64_.dll (unknown line)
      From worker 12:	ddot_k_ZEN at C:\buildkite-agent\builds\win2k22-amdci6-1\julialang\julia-master\julia-86e64109fb\bin\libopenblas64_.dll (unknown line)

This is in the newly added test, it fails only on 64-bit Windows....

[giordano]

Using the 64-bit Windows build in this PR, I tried to run

using LinearAlgebra, Libdl
dot_sym = dlsym(dlopen(Base.libblas_name), "cblas_ddot" * (Sys.WORD_SIZE == 64 ? "64_" : ""))
@ccall $(dot_sym)(2::Cint, [2.0, 3.0]::Ref{Cdouble}, 1::Cint, [4.0, 5.0]::Ref{Cdouble}, 1::Cint)::Cdouble

under Wine and it worked fine for me, so I have no clue of what's going on here. I'm happy to change the test to something else if someone can suggest something which doesn't crash on Windows.

[staticfloat]

The problem is that on 64-bit systems, you need to pass the integers as Int64, not Cint. Why this works on some platforms and not others is more of an accident around calling conventions and values that just so happen to be zero, than anything else.

[giordano]

That worked, thanks! netlib documentation isn't very helpful, I found the signature of cblas_ddot at https://netlib.org/lapack/explore-html/db/d91/cblas__ddot_8c_source.html, where the integer arguments have type CBLAS_INT which is

#define CBLAS_INT   int32_t

🤷

[staticfloat]

Right, but that’s for an LP64 build. ILP64 means defining CBLSS_INT to be int64_t.

I myself have caused such issues dozens of times in the past. It’s easy to forget. :P