LLVM coverage runtime included when linking with `-sMAIN_MODULE=1`

[hoodmane]

Since #24160, -sMAIN_MODULE=1 links the LLVM coverage runtime. It would be nice if it didn't.
This caused the following downstream issue:
pyodide/pyodide#6339

[sbc100]

Yes we can probably work around this. The problem with using -sMAIN_MODULE=1 is that its supposed to include literally everything.. is there any way you could eventually move away from having to use it.

On a related note I recently discoverted that compiler-rt (which is where this runtime lives) is supposed to be linked into each and every DLL and so maybe we could move to a model where easy side module gets its own copy (to match the native platforms). Then no compiler-rt code would need to be force-included in -sMAIN_MODULE=1 mode.

[sbc100]

Actually I spoke too soon, only the .builtins part of compiler-rt is supposed to be statically linked.

[sbc100]

Oh, I see that the issue is actually not that the runtime is being included but that is being activated and generating a file.

[hoodmane]

is there any way you could eventually move away from having to use it.

Yeah I'm definitely interested to move away from it but it's a bit of a chore. I've investigated a bit in the past, I think what we need to do is list out the symbols used in every dynamic library we know about and then put them all into a list of symbols we think we need to export. Then we maybe need to make our build tooling warn if it links a dynamic library that needs a symbol not in that list. After a while if people don't report any warnings we can switch to that list.