Stop emitting `.debug_pubnames` and `.debug_pubtypes` into Linux object files and binaries

[rocallahan]

As far as I can tell, nothing uses these. They bloat executables and make linking more expensive.

At least stop emitting them for debuginfo=1.

[sfackler]

It seems like gcc and clang don't emit these anymore, so it'd probably make sense for us to skip them as well.

[tromey]

I have a patch for this but I'm still looking at how to write a test for it.

[tromey]

I learned on #llvm that these sections may be read by lld to create a .gdb_index section. So perhaps they are desirable after all. I'll look into lld to see.

[tromey]

.debug_pubnames and .debug_pubtypes are indeed read by lld when --gdb-index is given. This doesn't appear to be the default. This emits an index that speeds up gdb symbol reading. So, it seems that Linux is the platform where these sections might be useful. However, since it requires a linker flag, perhaps the thing to do is just drop this everywhere.

[DenialAdams]

.debug_pubnames and .debug_pubtypes are indeed read by lld when --gdb-index is given. This doesn't appear to be the default. This emits an index that speeds up gdb symbol reading. So, it seems that Linux is the platform where these sections might be useful. However, since it requires a linker flag, perhaps the thing to do is just drop this everywhere.

It seems like we could put emitting them behind a rustc flag by default, but add a cargo option that causes them to be emitted and turns the linker flag on?

[tromey]

It seems like we could skip emitting them by default, but add a cargo option that causes them to be emitted and turns the linker flag on?

Yeah, that could be done. I have other priorities, though, so I'm going to drop this for now. I can share my rustc patch if anybody wants to pick it up.

[khuey]

What blocks landing this?