On MSVC targets rustc will add symbols prefixed with `_imp_` to LLVM modules to
"emulate" dllexported statics as that workaround is still in place after #27438
hasn't been solved otherwise. These statics, however, were getting gc'd by
ThinLTO accidentally which later would cause linking failures.

This commit updates the location we add such symbols to happen just before
codegen to ensure that (a) they're not eliminated by the optimizer and (b) the
optimizer doesn't even worry about them.

Closes #45347

First the `addPreservedGUID` function forgot to take care of "alias" summaries.
I'm not 100% sure what this is but the current code now matches upstream. Next
the `computeDeadSymbols` return value wasn't actually being used, but it needed
to be used! Together these should...

Closes #45195

Some targets, like msp430 and nvptx, don't work with multiple codegen units
right now for bugs or fundamental reasons. To expose this allow targets to
express a default.

Closes #45000

Fixes #43402 (take 2)

This commit is an implementation of LLVM's ThinLTO for consumption in rustc
itself. Currently today LTO works by merging all relevant LLVM modules into one
and then running optimization passes. "Thin" LTO operates differently by having
more sharded work and allowing parallelism opportunities between optimizing
codegen units. Further down the road Thin LTO also allows *incremental* LTO
which should enable even faster release builds without compromising on the
performance we have today.

This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then
also implements two forms of ThinLTO:

* In one mode we'll *only* perform ThinLTO over the codegen units produced in a
  single compilation. That is, we won't load upstream rlibs, but we'll instead
  just perform ThinLTO amongst all codegen units produced by the compiler for
  the local crate. This is intended to emulate a desired end point where we have
  codegen units turned on by default for all crates and ThinLTO allows us to do
  this without performance loss.

* In anther mode, like full LTO today, we'll optimize all upstream dependencies
  in "thin" mode. Unlike today, however, this LTO step is fully parallelized so
  should finish much more quickly.

There's a good bit of comments about what the implementation is doing and where
it came from, but the tl;dr; is that currently most of the support here is
copied from upstream LLVM. This code duplication is done for a number of
reasons:

* Controlling parallelism means we can use the existing jobserver support to
  avoid overloading machines.
* We will likely want a slightly different form of incremental caching which
  integrates with our own incremental strategy, but this is yet to be
  determined.
* This buys us some flexibility about when/where we run ThinLTO, as well as
  having it tailored to fit our needs for the time being.
* Finally this allows us to reuse some artifacts such as our `TargetMachine`
  creation, where all our options we used today aren't necessarily supported by
  upstream LLVM yet.

My hope is that we can get some experience with this copy/paste in tree and then
eventually upstream some work to LLVM itself to avoid the duplication while
still ensuring our needs are met. Otherwise I fear that maintaining these
bindings may be quite costly over the years with LLVM updates!

Fixes #43402 now there's no multithreaded panic printouts

Also update a comment

This commit is a refactoring of the LTO backend in Rust to support compilations
with multiple codegen units. The immediate result of this PR is to remove the
artificial error emitted by rustc about `-C lto -C codegen-units-8`, but longer
term this is intended to lay the groundwork for LTO with incremental compilation
and ultimately be the underpinning of ThinLTO support.

The problem here that needed solving is that when rustc is producing multiple
codegen units in one compilation LTO needs to merge them all together.
Previously only upstream dependencies were merged and it was inherently relied
on that there was only one local codegen unit. Supporting this involved
refactoring the optimization backend architecture for rustc, namely splitting
the `optimize_and_codegen` function into `optimize` and `codegen`. After an LLVM
module has been optimized it may be blocked and queued up for LTO, and only
after LTO are modules code generated.

Non-LTO compilations should look the same as they do today backend-wise, we'll
spin up a thread for each codegen unit and optimize/codegen in that thread. LTO
compilations will, however, send the LLVM module back to the coordinator thread
once optimizations have finished. When all LLVM modules have finished optimizing
the coordinator will invoke the LTO backend, producing a further list of LLVM
modules. Currently this is always a list of one LLVM module. The coordinator
then spawns further work to run LTO and code generation passes over each module.

In the course of this refactoring a number of other pieces were refactored:

* Management of the bytecode encoding in rlibs was centralized into one module
  instead of being scattered across LTO and linking.
* Some internal refactorings on the link stage of the compiler was done to work
  directly from `CompiledModule` structures instead of lists of paths.
* The trans time-graph output was tweaked a little to include a name on each
  bar and inflate the size of the bars a little

This commit changes the default of rustc to use 32 codegen units when compiling
in debug mode, typically an opt-level=0 compilation. Since their inception
codegen units have matured quite a bit, gaining features such as:

* Parallel translation and codegen enabling codegen units to get worked on even
  more quickly.
* Deterministic and reliable partitioning through the same infrastructure as
  incremental compilation.
* Global rate limiting through the `jobserver` crate to avoid overloading the
  system.

The largest benefit of codegen units has forever been faster compilation through
parallel processing of modules on the LLVM side of things, using all the cores
available on build machines that typically have many available. Some downsides
have been fixed through the features above, but the major downside remaining is
that using codegen units reduces opportunities for inlining and optimization.
This, however, doesn't matter much during debug builds!

In this commit the default number of codegen units for debug builds has been
raised from 1 to 32. This should enable most `cargo build` compiles that are
bottlenecked on translation and/or code generation to immediately see speedups
through parallelization on available cores.

Work is being done to *always* enable multiple codegen units (and therefore
parallel codegen) but it requires #44841 at least to be landed and stabilized,
but stay tuned if you're interested in that aspect!

This commit moves the actual code generation in the compiler behind a query
keyed by a codegen unit's name. This ended up entailing quite a few internal
refactorings to enable this, along with a few cut corners:

* The `OutputFilenames` structure is now tracked in the `TyCtxt` as it affects a
  whole bunch of trans and such. This is now behind a query and threaded into
  the construction of the `TyCtxt`.

* The `TyCtxt` now has a channel "out the back" intended to send data to worker
  threads in rustc_trans. This is used as a sort of side effect of the codegen
  query but morally what's happening here is the return value of the query
  (currently unit but morally a path) is only valid once the background threads
  have all finished.

* Dispatching work items to the codegen threads was refactored to only rely on
  data in `TyCtxt`, which mostly just involved refactoring where data was
  stored, moving it from the translation thread to the controller thread's
  `CodegenContext` or the like.

* A new thread locals was introduced in trans to work around the query
  system. This is used in the implementation of `assert_module_sources` which
  looks like an artifact of the old query system and will presumably go away
  once red/green is up and running.

This commit attaches a channel to the LLVM workers to the `TyCtxt` which will
later be used during the codegen query to actually send work to LLVM workers.
Otherwise this commit is just plumbing this channel throughout the compiler to
ensure it reaches the right consumers.

This is a big map that ends up inside of a `CrateContext` during translation for
all codegen units. This means that any change to the map may end up causing an
incremental recompilation of a codegen unit! In order to reduce the amount of
dependencies here between codegen units and the actual input crate this commit
refactors dealing with exported symbols and such into various queries.

The new queries are largely based on existing queries with filled out
implementations for the local crate in addition to external crates, but the main
idea is that while translating codegen untis no unit needs the entire set of
exported symbols, instead they only need queries about particulare `DefId`
instances every now and then.

The linking stage, however, still generates a full list of all exported symbols
from all crates, but that's going to always happen unconditionally anyway, so no
news there!

This commit moves the definition of the `ExportedSymbols` structure to the
`rustc` crate and then creates a query that'll be used to construct the
`ExportedSymbols` set. This in turn uses the reachablity query exposed in the
previous commit.

This commit removes the `dep_graph` field from the `Session` type according to
issue #44390. Most of the fallout here was relatively straightforward and the
`prepare_session_directory` function was rejiggered a bit to reuse the results
in the later-called `load_dep_graph` function.

Closes #44390

The main use of `CrateStore` *before* the `TyCtxt` is created is during
resolution, but we want to be sure that any methods used before resolution are
not used after the `TyCtxt` is created. This commit starts moving the methods
used by resolve to all be named `{name}_untracked` where the rest of the
compiler uses just `{name}` as a query.

During this transition a number of new queries were added to account for
post-resolve usage of these methods.

This comit applies the following changes:

* Deletes the `is_allocator` query as it's no longer used
* Moves the `is_sanitizer_runtime` method to a query
* Moves the `is_profiler_runtime` method to a query
* Moves the `panic_strategy` method to a query
* Moves the `is_no_builtins` method to a query
* Deletes the cstore method of `is_compiler_builtins`. The query was added in
  #42588 but the `CrateStore` method was not deleted

A good bit of these methods were used late in linking during trans so a new
dedicated structure was created to ship a calculated form of this information
over to the linker rather than having to ship the whole of `TyCtxt` over to
linking.

Like #43008 (f6689991), but _much more aggressive_.

async-llvm(6): Make the LLVM work coordinator get its work package through a channel instead of upfront.