Reduce the amount of untracked state in TyCtxt
Access to untracked global state may generate instances of #84970.
The GlobalCtxt contains the lowered HIR, the resolver outputs and interners.
By wrapping the resolver inside a query, we make sure those accesses are properly tracked.
As a no_hash query, all dependent queries essentially become `eval_always`,
what they should have been from the beginning.
Previously, we sorted the vec prior to hashing, making the hash
independent of the original (command-line argument) order. However, the
original vec was still always kept in the original order, so we were
relying on the rest of the compiler always working with it in an
'order-independent' way.
This assumption was not being upheld by the `native_libraries` query -
the order of the entires in its result depends on the order of entries
in `Options.libs`. This lead to an 'unstable fingerprint' ICE when the
`-l` arguments were re-ordered.
This PR removes the sorting logic entirely. Re-ordering command-line
arguments (without adding/removing/changing any arguments) seems like a
really niche use case, and correctly optimizing for it would require
additional work. By always hashing arguments in their original order, we
can entirely avoid a cause of 'unstable fingerprint' errors.
Preserve `SyntaxContext` for invalid/dummy spans in crate metadata
Fixes#85197
We already preserved the `SyntaxContext` for invalid/dummy spans in the
incremental cache, but we weren't doing the same for crate metadata.
If an invalid (lo/hi from different files) span is written to the
incremental cache, we will decode it with a 'dummy' location, but keep
the original `SyntaxContext`. Since the crate metadata encoder was only
checking for `DUMMY_SP` (dummy location + root `SyntaxContext`),
the metadata encoder would treat it as a normal span, encoding the
`SyntaxContext`. As a result, the final span encoded to the metadata
would change across sessions, even if the crate itself was unchanged.
This could lead to an 'unstable fingerprint' ICE under the following conditions:
1. We compile a crate with an invalid span using incremental compilation. The metadata encoder discards the `SyntaxContext` since the span is invalid, while the incremental cache encoder preserves the `SyntaxContext`
2. From another crate, we execute a foreign query, decoding the invalid span from the metadata as `DUMMY_SP` (e.g. with `SyntaxContext::root()`). This span gets hashed into the query fingerprint. So far, this has always happened through the `optimized_mir` query.
3. We recompile the first crate using our populated incremental cache, without changing anything. We load the (previously) invalid span from our incremental cache - it gets converted to a span with a dummy (but valid) location, along with the original `SyntaxContext`. This span gets written out to the crate metadata - since it now has a valid location, we preserve its `SyntaxContext`.
4. We recompile the second crate, again using a populated incremental cache. We now re-run the foreign query `optimized_mir` - the foreign crate hash is unchanged, but we end up decoding a different span (it now ha a non-root `SyntaxContext`). This results in the fingerprint changing, resulting in an ICE.
This PR updates our encoding of spans in the crate metadata to mirror
the encoding of spans into the incremental cache. We now always encode a
`SyntaxContext`, and encode location information for spans with a
non-dummy location.
Fixes#85197
We already preserved the `SyntaxContext` for invalid/dummy spans in the
incremental cache, but we weren't doing the same for crate metadata.
If an invalid (lo/hi from different files) span is written to the
incremental cache, we will decode it with a 'dummy' location, but keep
the original `SyntaxContext`. Since the crate metadata encoder was only
checking for `DUMMY_SP` (dummy location + root `SyntaxContext`),
the metadata encoder would treat it as a normal span, encoding the
`SyntaxContext`. As a result, the final span encoded to the metadata
would change across sessions, even if the crate itself was unchanged.
This PR updates our encoding of spans in the crate metadata to mirror
the encoding of spans into the incremental cache. We now always encode a
`SyntaxContext`, and encode location information for spans with a
non-dummy location.
Fixes#80691
When we evaluate a trait predicate, we convert an
`EvaluatedToOk` result to `EvaluatedToOkModuloRegions` if we erased any
regions. We cache the result under a region-erased 'freshened'
predicate, so `EvaluatedToOk` may not be correct for other predicates
that have the same cache key.
Fixes issue #82920
Even if an item does not change between compilation sessions, it may end
up with a different `DefId`, since inserting/deleting an item affects
the `DefId`s of all subsequent items. Therefore, we use a `DefPathHash`
in the incremental compilation system, which is stable in the face of
changes to unrelated items.
In particular, the query system will consider the inputs to a query to
be unchanged if any `DefId`s in the inputs have their `DefPathHash`es
unchanged. Queries are pure functions, so the query result should be
unchanged if the query inputs are unchanged.
Unfortunately, it's possible to inadvertantly make a query result
incorrectly change across compilations, by relying on the specific value
of a `DefId`. Specifically, if the query result is a slice that gets
sorted by `DefId`, the precise order will depend on how the `DefId`s got
assigned in a particular compilation session. If some definitions end up
with different `DefId`s (but the same `DefPathHash`es) in a subsequent
compilation session, we will end up re-computing a *different* value for
the query, even though the query system expects the result to unchanged
due to the unchanged inputs.
It turns out that we have been sorting the predicates computed during
`astconv` by their `DefId`. These predicates make their way into the
`super_predicates_that_define_assoc_type`, which ends up getting used to
compute the vtables of trait objects. This, re-ordering these predicates
between compilation sessions can lead to undefined behavior at runtime -
the query system will re-use code built with a *differently ordered*
vtable, resulting in the wrong method being invoked at runtime.
This PR avoids sorting by `DefId` in `astconv`, fixing the
miscompilation. However, it's possible that other instances of this
issue exist - they could also be easily introduced in the future.
To fully fix this issue, we should
1. Turn on `-Z incremental-verify-ich` by default. This will cause the
compiler to ICE whenver an 'unchanged' query result changes between
compilation sessions, instead of causing a miscompilation.
2. Remove the `Ord` impls for `CrateNum` and `DefId`. This will make it
difficult to introduce ICEs in the first place.
Fixes#80336
Due to macro expansion, we may end up with spans with an invalid
location and non-root `SyntaxContext`. This commits preserves the
`SyntaxContext` of such spans in the incremental cache, and ensures
that we always hash the `SyntaxContext` when computing the `Fingerprint`
of a `Span`
Previously, we would discard the `SyntaxContext` during serialization to
the incremental cache, causing the span's `Fingerprint` to change across
compilation sessions.
Fixes#79890
Previously, we just copied a `RawDefId` from the 'old' map to the 'new'
map. However, the `RawDefId` for a given `DefPathHash` may be different
in the current compilation session. Using `def_path_hash_to_def_id`
ensures that the `RawDefId` we use is valid in the current session.
On the nopt builders, we disable optimization by default for all tests
which causes the new behavior to take effect and causes the tests to
fail when they should not. By passing the `-O` flag explicitly, we will
always run these tests with optimizations enabled.
Fixes#79661
In incremental compilation mode, we update a `DefPathHash -> DefId`
mapping every time we create a `DepNode` for a foreign `DefId`.
This mapping is written out to the on-disk incremental cache, and is
read by the next compilation session to allow us to lazily decode
`DefId`s.
When we decode a `DepNode` from the current incremental cache, we need
to ensure that any previously-recorded `DefPathHash -> DefId` mapping
gets recorded in the new mapping that we write out. However, PR #74967
didn't do this in all cases, leading to us being unable to decode a
`DefPathHash` in certain circumstances.
This PR refactors some of the code around `DepNode` deserialization to
prevent this kind of mistake from happening again.
Update affected ui & incremental tests to use a user declared variable
bindings instead of temporaries. The former are preserved because of
debuginfo, the latter are not.
Use llvm::computeLTOCacheKey to determine post-ThinLTO CGU reuse
During incremental ThinLTO compilation, we attempt to re-use the
optimized (post-ThinLTO) bitcode file for a module if it is 'safe' to do
so.
Up until now, 'safe' has meant that the set of modules that our current
modules imports from/exports to is unchanged from the previous
compilation session. See PR #67020 and PR #71131 for more details.
However, this turns out be insufficient to guarantee that it's safe
to reuse the post-LTO module (i.e. that optimizing the pre-LTO module
would produce the same result). When LLVM optimizes a module during
ThinLTO, it may look at other information from the 'module index', such
as whether a (non-imported!) global variable is used. If this
information changes between compilation runs, we may end up re-using an
optimized module that (for example) had dead-code elimination run on a
function that is now used by another module.
Fortunately, LLVM implements its own ThinLTO module cache, which is used
when ThinLTO is performed by a linker plugin (e.g. when clang is used to
compile a C proect). Using this cache directly would require extensive
refactoring of our code - but fortunately for us, LLVM provides a
function that does exactly what we need.
The function `llvm::computeLTOCacheKey` is used to compute a SHA-1 hash
from all data that might influence the result of ThinLTO on a module.
In addition to the module imports/exports that we manually track, it
also hashes information about global variables (e.g. their liveness)
which might be used during optimization. By using this function, we
shouldn't have to worry about new LLVM passes breaking our module re-use
behavior.
In LLVM, the output of this function forms part of the filename used to
store the post-ThinLTO module. To keep our current filename structure
intact, this PR just writes out the mapping 'CGU name -> Hash' to a
file. To determine if a post-LTO module should be reused, we compare
hashes from the previous session.
This should unblock PR #75199 - by sheer chance, it seems to have hit
this issue due to the particular CGU partitioning and optimization
decisions that end up getting made.
During incremental ThinLTO compilation, we attempt to re-use the
optimized (post-ThinLTO) bitcode file for a module if it is 'safe' to do
so.
Up until now, 'safe' has meant that the set of modules that our current
modules imports from/exports to is unchanged from the previous
compilation session. See PR #67020 and PR #71131 for more details.
However, this turns out be insufficient to guarantee that it's safe
to reuse the post-LTO module (i.e. that optimizing the pre-LTO module
would produce the same result). When LLVM optimizes a module during
ThinLTO, it may look at other information from the 'module index', such
as whether a (non-imported!) global variable is used. If this
information changes between compilation runs, we may end up re-using an
optimized module that (for example) had dead-code elimination run on a
function that is now used by another module.
Fortunately, LLVM implements its own ThinLTO module cache, which is used
when ThinLTO is performed by a linker plugin (e.g. when clang is used to
compile a C proect). Using this cache directly would require extensive
refactoring of our code - but fortunately for us, LLVM provides a
function that does exactly what we need.
The function `llvm::computeLTOCacheKey` is used to compute a SHA-1 hash
from all data that might influence the result of ThinLTO on a module.
In addition to the module imports/exports that we manually track, it
also hashes information about global variables (e.g. their liveness)
which might be used during optimization. By using this function, we
shouldn't have to worry about new LLVM passes breaking our module re-use
behavior.
In LLVM, the output of this function forms part of the filename used to
store the post-ThinLTO module. To keep our current filename structure
intact, this PR just writes out the mapping 'CGU name -> Hash' to a
file. To determine if a post-LTO module should be reused, we compare
hashes from the previous session.
This should unblock PR #75199 - by sheer chance, it seems to have hit
this issue due to the particular CGU partitioning and optimization
decisions that end up getting made.
If a symbol name can only be imported from one place for a type, and
as long as it was not glob-imported anywhere in the current crate, we
can trim its printed path and print only the name.
This has wide implications on error messages with types, for example,
shortening `std::vec::Vec` to just `Vec`, as long as there is no other
`Vec` importable anywhere.
This adds a new '-Z trim-diagnostic-paths=false' option to control this
feature.
On the good path, with no diagnosis printed, we should try to avoid
issuing this query, so we need to prevent trimmed_def_paths query on
several cases.
This change also relies on a previous commit that differentiates
between `Debug` and `Display` on various rustc types, where the latter
is trimmed and presented to the user and the former is not.
