Move metadata generation

[nnethercote]

It's currently done within start_codegen, which is called from codegen_and_build_linker, and it gets included in the codegen timing section. All this is surprising, because it's not part of codegen.

This commit moves it into run_compiler, just before the call to codegen_and_build_linker. This is a more sensible place for it. The commit also adds some extra non-obvious information about the has_errors_or_delayed_bugs check, which I learned when I tried moving metadata generation earlier.

Likewise, the nearby rustc_delayed_bug_from_inside_query code is also moved, because (a) it's not part of codegen, and (b) it needs to be nearby.

r? @bjorn3

[nnethercote]

The background here: I was trying to see if metadata generation could be moved significantly earlier to increase the amount of overlap provided by compilation pipelining. This turns out to be difficult; lots of things start to break if metadata generation precedes some or all of analysis. But I figured I could do this cleanup and improve some comments with some things I learned.

[Kobzol]

Oh, I guess that metadata generation was part of the "backend" section, as recorded in rustc-perf and --json=timings. But I suppose that it does indeed make more sense to move it to the "frontend" section, as metadata generation also happens in cargo check.

[bjorn3]

This way it is easy to forget to call this function before calling codegen_and_build_linker. rustc_interface is meant to be called directly by custom drivers if rustc_driver doesn't provide enough flexibility. For example for rustdoc. mutest-rs also uses it directly. Also by moving this call before report_symbol_names, errors in the latter no longer block codegen. Maybe instead just move the tcx.sess.timings.start_section(tcx.sess.dcx(), TimingSection::Codegen); later in codegen_and_build_linker?