Offset calculation in suitesparse_wrapper

[andreasnoack]

I think that the way we are calculating the offsets in suitesparse_wrapper.c is broken because the offsets are calculated for the machine where suitesparse_wrapper.c is compiled. These offsets need not be the same when Julia is from a repository because the version of SuiteSparse can be different from the machine where the package was build.

This can be seen on Julia release on Ubuntu where CHOLMOD objects are printed wrongly because the print level is set in specific location in a vector and this location has changed over time because the cholmod_common_struct has changed over time.

[andreasnoack]

@staticfloat I tend to think that this is mainly a packaging issue. The suitesparse_wrapper shim is correct when build together with the linked SuiteSparse, but not when the library is substituted. Do you know how we can handle this?

[dmbates]

If the libcholmod is sufficiently recent to have a cholmod_version function then you can pass the three integers in the Julia array version_array to calls to functions defined in SuiteSparse_wrapper.c and compare the three integers to CHOLMOD_MAIN_VERSION, CHOLMOD_SUB_VERSION and CHOLMOD_SUBSUB_VERSION. If they don't match then all bets are off. Also, if the version of cholmod in the shared library is too old to have a cholmod_version function then all bets are off.

[andreasnoack]

We should check this when Julia is launched, and, as mentioned in the other issue, not (only) when the system image is build, but I don't how to do that.

[ivarne]

See

julia/base/gmp.jl

Line 54 in 32aee08

function __init__()

, for a partially working solution. The only problem is that the error is misleading, as it won't terminate the program, but just print a warning.

[andreasnoack]

@ivarne Thanks. That was exactly what I was looking for.

[andreasnoack]

I think the best we can do here is #10362, so I'll close.