This is how InlineAsm diagnostics with source information are
reported now. Previously a separate InlineAsm diagnostic handler
was used.

This has regressed due to https://bugs.llvm.org/show_bug.cgi?id=51211.
It's pretty likely that we'll have to eat this regression for this
release.

This uses comdats since LLVM 13, causing various minor changes to the
output.

The #[used] attribute explicitly only requires symbols to be
retained in object files, but allows the linker to drop them
if dead. This corresponds to llvm.compiler.used semantics.

The motivation to change this *now* is that https://reviews.llvm.org/D97448
starts emitting #[used] symbols into unique sections with
SHF_GNU_RETAIN flag. This triggers a bug in some version of gold,
resulting in the ARGV_INIT_ARRAY symbol part of the .init_array
section to be incorrectly placed.

When LLVM_INCLUDE_TESTS is enabled (by default), LLVM requires
Python 3.6 for the lit test runner, otherwise only Python 3.0 is
required.

As we have many docker images using Ubuntu 16.04, which only has
Python 3.5, this avoids the need to install a newer Python version
for them.

The TargetMachine may be referencing data in the context. In
particular, at least the GlobalISel instruction selector stored
in the TM may reference a TrackedMDNode DebugLoc that destruction
of the TargetMachine will try to untrack.

Since https://reviews.llvm.org/D103355 this will usually also use
internal rather than private on Windows as well. We don't
particularly care about this implementation detail, just accept
either.

This follows what clang does in CoverageMappingGen. Using just
llvm.compiler.used is insufficient at least for MSVC targets.