Elaborate 'ccall' requires the compiler error message

[sunho]

Recently, I was building julia with USE_BINARYBUILDER_LLVM=0 in order to try out my own llvm fork. Since I had a higher glibc in my local machine, libllvm failed to load as libstdc++ inside julia/usr/lib was outdated.

However, since the error message from load_library was intentioanlly silenced (err paramter is 0), there was no helpful error message printed out. The actual error message I got was in the screenshot below. As it points to completely unreleated direction, I had to spend lot of hours to track down what was going on.

This pr makes it better by priting warning message to stdout when codegen library failed to load so that we can identify this kind of llvm load error right away.

[vchuravy]

Hm, I understand your motivation, but Julia is supposed to be able to start without Codegen available. So maybe we can instead make the ccall error message better?

[sunho]

Hmm, clarifying ccall error is definitely an option. It would be nice if we can differentiate library not found error and other unexpected error, though. In glibc error case, we can confidently tell it to user right?

[vchuravy]

Yeah that would be fine. If the library is present and we failed to load it we should definitely warn the user about it.

[haampie]

See #45641 (comment) and below, libstdc++ detection in julia is incorrect: it locates libstdc++ in the implicit link directories of the fortran compiler assuming this can only be gfortran. But when using clang as a c++ compiler, it links to the libstdc++ of the most recent gcc it can find on the system (or in some cases llvm is compiled with a fixed, specific gcc toolchain), not the "default" gcc in the PATH. Same can happen when gfortran is from an older GCC than g++

[maleadt]

Hm, I understand your motivation, but Julia is supposed to be able to start without Codegen available.

load_library could differentiate between "library does not exist" and "library is not loadable", the former is unlikely to be intentional.

FWIW, I also lost an hour on this yesterday, so would be in favor of better reporting. Especially during build, where we don't ever want to load a stubs library. So alternatively, we could introduce an -fno-stubs and use that during build?

[sunho]

I have editted the pr to just elaborate ccall error message. Good thing is even outside the boostrap, the ccall error message will pop up when codegen library failed to load.

[vtjnash]

we will fix this as the load_library layer