This massively speeds up serialization. It also
seems to produce deterministic metadata hashes
(before I was seeing inconsistent results).
Fixes#35232.
The biggest problem, actually, is krate numbers being removed entirely,
which can lead to array-index-out-of-bounds errors.
cc #35123 -- not a complete fix, since really we ought to "map" the old
crate numbers to the new ones, not just detect changes.
The way we do HIR inlining introduces reads of the "Hir" into the graph,
but this Hir in fact belongs to other crates, so when we try to load
later, we ICE because the Hir nodes in question don't belond to the
crate (and we haven't done inlining yet). This pass rewrites those HIR
nodes to the metadata from which the inlined HIR was loaded.
In the older version, a `.o` and ` .bc` file were separate
work-products. This newer version keeps, for each codegen-unit, a set
of files of different kinds. We assume that if any kinds are available
then all the kinds we need are available, since the precise set of
switches will depend on attributes and command-line switches.
Should probably test this: the effect of changing attributes in
particular might not be successfully tracked?
