Use the object crate for metadata reading
This allows sharing the metadata reader between cg_llvm, cg_clif and other codegen backends.
This is not currently useful for rlib reading with cg_spirv ([rust-gpu](https://github.com/EmbarkStudios/rust-gpu/)) as it uses tar rather than ar as .rlib format, but it is useful for dylib reading required for loading proc macros. (cc `@eddyb)`
The object crate is already trusted as dependency of libstd through backtrace. As far as I know it supports reading all object file formats used by targets for which we support rust dylibs with crate metadata, but I am not certain. If this happens to not be the case, I could keep using LLVM for reading dylib metadata.
Marked as WIP for a perf run and as it is based on #83637.
This change tunes ahead-of-time codegening according to the amount of
concurrency available, rather than according to the number of CPUs on
the system. This can lower memory usage by reducing the number of
compiled LLVM modules in memory at once, particularly across several
rustc instances.
Previously, each rustc instance would assume that it should codegen
ahead of time to meet the demand of number-of-CPUs workers. But often, a
rustc instance doesn't have nearly that much concurrency available to
it, because the concurrency availability is split, via the jobserver,
across all active rustc instances spawned by the driving cargo process,
and is further limited by the `-j` flag argument. Therefore, each rustc
might have had several times the number of LLVM modules in memory than
it really needed to meet demand. If the modules were large, the effect
on memory usage would be noticeable.
With this change, the required amount of ahead-of-time codegen scales up
with the actual number of workers running within a rustc instance. Note
that the number of workers running can be less than the actual
concurrency available to a rustc instance. However, if more concurrency
is actually available, workers are spun up quickly as job tokens are
acquired, and the ahead-of-time codegen scales up quickly as well.
For better throughput during parallel processing by LLVM, we used to sort
CGUs largest to smallest. This would lead to better thread utilization
by, for example, preventing a large CGU from being processed last and
having only one LLVM thread working while the rest remained idle.
However, this strategy would lead to high memory usage, as it meant the
LLVM-IR for all of the largest CGUs would be resident in memory at once.
Instead, we can compromise by ordering CGUs such that the largest and
smallest are first, second largest and smallest are next, etc. If there
are large size variations, this can reduce memory usage significantly.
