By splitting the `FnSig` within `TyKind::FnPtr` into `FnSigTys` and
`FnHeader`, which can be packed more efficiently. This reduces the size
of the hot `TyKind` type from 32 bytes to 24 bytes on 64-bit platforms.
This reduces peak memory usage by a few percent on some benchmarks. It
also reduces cache misses and page faults similarly, though this doesn't
translate to clear cycles or wall-time improvements on CI.
We already do this for a number of crates, e.g. `rustc_middle`,
`rustc_span`, `rustc_metadata`, `rustc_span`, `rustc_errors`.
For the ones we don't, in many cases the attributes are a mess.
- There is no consistency about order of attribute kinds (e.g.
`allow`/`deny`/`feature`).
- Within attribute kind groups (e.g. the `feature` attributes),
sometimes the order is alphabetical, and sometimes there is no
particular order.
- Sometimes the attributes of a particular kind aren't even grouped
all together, e.g. there might be a `feature`, then an `allow`, then
another `feature`.
This commit extends the existing sorting to all compiler crates,
increasing consistency. If any new attribute line is added there is now
only one place it can go -- no need for arbitrary decisions.
Exceptions:
- `rustc_log`, `rustc_next_trait_solver` and `rustc_type_ir_macros`,
because they have no crate attributes.
- `rustc_codegen_gcc`, because it's quasi-external to rustc (e.g. it's
ignored in `rustfmt.toml`).
Stabilize the size of incr comp object file names
The current implementation does not produce stable-length paths, and we create the paths in a way that makes our allocation behavior is nondeterministic. I think `@eddyb` fixed a number of other cases like this in the past, and this PR fixes another one. Whether that actually matters I have no idea, but we still have bimodal behavior in rustc-perf and the non-uniformity in `find` and `ls` was bothering me.
I've also removed the truncation of the mangled CGU names. Before this PR incr comp paths look like this:
```
target/debug/incremental/scratch-38izrrq90cex7/s-gux6gz0ow8-1ph76gg-ewe1xj434l26w9up5bedsojpd/261xgo1oqnd90ry5.o
```
And after, they look like this:
```
target/debug/incremental/scratch-035omutqbfkbw/s-gux6borni0-16r3v1j-6n64tmwqzchtgqzwwim5amuga/55v2re42sztc8je9bva6g8ft3.o
```
On the one hand, I'm sure this will break some people's builds because they're on Windows and only a few bytes from the path length limit. But if we're that seriously worried about the length of our file names, I have some other ideas on how to make them smaller. And last time I deleted some hash truncations from the compiler, there was a huge drop in the number if incremental compilation ICEs that were reported: https://github.com/rust-lang/rust/pull/110367https://github.com/rust-lang/rust/pull/110367
---
Upon further reading, this PR actually fixes a bug. This comment says the CGU names are supposed to be a fixed-length hash, and before this PR they aren't: ca7d34efa9/compiler/rustc_monomorphize/src/partitioning.rs (L445-L448)
KCFI: Use legal charset in shim encoding
To separate `ReifyReason::FnPtr` from `ReifyReason::VTable`, we hyphenated the shims. Hyphens are not actually legal, but underscores are, so use those instead.
r? `@compiler-errors`
To separate `ReifyReason::FnPtr` from `ReifyReason::VTable`, we
hyphenated the shims. Hyphens are not actually legal, but underscores
are, so use those instead.
Create the rustc_sanitizers crate and move the source code for the CFI
and KCFI sanitizers to it.
Co-authored-by: David Wood <agile.lion3441@fuligin.ink>
CFI: Don't rewrite ty::Dynamic directly
Now that we're using a type folder, the arguments in predicates are processed automatically - we don't need to descend manually.
We also want to keep projection clauses around, and this does so.
r? `@compiler-errors`
Now that we're using a type folder, the arguments in predicates are
processed automatically - we don't need to descend manually.
We also want to keep projection clauses around, and this does so.
Restore typeid_for_instance default behavior of performing self type
erasure, since it's the most common case and what it does most of the
time. Using concrete self (or not performing self type erasure) is for
assigning a secondary type id, and secondary type ids are only assigned
when they're unique and to methods, and also are only tested for when
methods are used as function pointers.
CFI: Support function pointers for trait methods
Adds support for both CFI and KCFI for function pointers to trait methods by attaching both concrete and abstract types to functions.
KCFI does this through generation of a `ReifyShim` on any function pointer for a method that could go into a vtable, and keeping this separate from `ReifyShim`s that are *intended* for vtable us by setting a `ReifyReason` on them.
CFI does this by setting both the concrete and abstract type on every instance.
This should land after #123024 or a similar PR, as it diverges the implementation of CFI vs KCFI.
r? `@compiler-errors`
Adds support for both CFI and KCFI for attaching concrete and abstract
types to functions. KCFI does this through generation of `ReifyShim` on
any function pointer that could go in a vtable, and checking the
`ReifyReason` when emitting the instance. CFI does this by attaching
both the concrete and abstract type to every instance.
TypeID codegen tests are switched to be anchored on the left rather than
the right in order to allow emission of additional type attachments.
Fixes#115953
KCFI needs to be able to tell which kind of `ReifyShim` it is examining
in order to decide whether to use a concrete type (`FnPtr` case) or an
abstract case (`Vtable` case). You can *almost* tell this from context,
but there is one case where you can't - if a trait has a method which is
*not* `#[track_caller]`, with an impl that *is* `#[track_caller]`, both
the vtable and a function pointer created from that method will be
`ReifyShim(def_id)`.
Currently, the reason is optional to ensure no additional unique
`ReifyShim`s are added without KCFI on. However, the case in which an
extra `ReifyShim` is created is sufficiently rare that this may be worth
revisiting to reduce complexity.
Previously, we assumed all `ty::Coroutine` were general coroutines and
attempted to generalize them through the `Coroutine` trait. Select
appropriate traits for each kind of coroutine.
Similar to methods on a trait object, the most common way to indirectly
call a closure or coroutine is through the vtable on the appropriate
trait. This uses the same approach as we use for trait methods, after
backing out the trait arguments from the type.
CFI: Support calling methods on supertraits
Automatically adjust `Virtual` calls to supertrait functions to use the supertrait's trait object type as the receiver rather than the child trait.
cc `@compiler-errors` - this is the next usage of `trait_object_ty` I intend to have, so I thought it might be relevant while reviewing the existing one.
CFI: Fix methods as function pointer cast
Fix casting between methods and function pointers by assigning a secondary type id to methods with their concrete self so they can be used as function pointers.
This was split off from #116404.
cc `@compiler-errors` `@workingjubilee`
Fix casting between methods and function pointers by assigning a
secondary type id to methods with their concrete self so they can be
used as function pointers.
CFI: Fix drop and drop_in_place
Fix drop and drop_in_place by transforming self of drop and drop_in_place methods into a Drop trait objects.
This was split off from https://github.com/rust-lang/rust/pull/116404.
cc `@compiler-errors` `@workingjubilee`
CFI: (actually) check that methods are object-safe before projecting their receivers to `dyn Trait` in CFI
`trait_object_ty` assumed that associated types would be fully determined by the trait. This is *almost* true - const parameters and type parameters are no longer allowed, but lifetime parameters are. Since we erase all lifetime parameters anyways, instantiate it with as many erased regions as it needs.
Fixes: #123053
r? `@compiler-errors`
`trait_object_ty` assumed that associated types would be fully
determined by the trait. This is *almost* true - const parameters and
type parameters are no longer allowed, but lifetime parameters are.
Since we erase all lifetime parameters anyways, instantiate it with as
many erased regions as it needs.
Fixes: #123053
CFI: Support complex receivers
Right now, we only support rewriting `&self` and `&mut self` into `&dyn MyTrait` and `&mut dyn MyTrait`. This expands it to handle the full gamut of receivers by calculating the receiver based on *substitution* rather than based on a rewrite. This means that, for example, `Arc<Self>` will become `Arc<dyn MyTrait>` appropriately with this change.
This approach also allows us to support associated type constraints as well, so we will correctly rewrite `&self` into `&dyn MyTrait<T=i32>`, for example.
r? ```@workingjubilee```
Previously, we only rewrote `&self` and `&mut self` receivers. By
instantiating the method from the trait definition, we can make this
work work with arbitrary legal receivers instead.
In user-facing Rust, `dyn` always has at least one predicate following
it. Unfortunately, because we filter out marker traits from receivers at
callsites and `dyn Sync` is, for example, legal, this results in us
having `dyn` types with no predicates on occasion in our alias set
encoding. This patch handles cases where there are no predicates in a
`dyn` type which are relevant to its alias set.
Fixes#122998
CFI: Strip auto traits off Virtual calls
We already use `Instance` at declaration sites when available to glean additional information about possible abstractions of the type in use. This does the same when possible at callsites as well.
The primary purpose of this change is to allow CFI to alter how it generates type information for indirect calls through `Virtual` instances.
This is needed for the "separate machinery" version of my approach to the vtable issues (#122573), because we need to respond differently to a `Virtual` call to the same type as a non-virtual call, specifically [stripping auto traits off the receiver's `Self`](54b15b0c36) because there isn't a separate vtable for `Foo` vs `Foo + Send`.
This would also make a more general underlying mechanism that could be used by rcvalle's [proposed drop detection / encoding](edcd1e20a1) if we end up using his approach, as we could condition out on the `def_id` in the CFI code rather than requiring the generating code to explicitly note whether it was calling drop.
Additional trait bounds beyond the principal trait and its implications
are not possible in the vtable. This means that if a receiver is
`&dyn Foo + Send`, the function will only be expecting `&dyn Foo`.
This strips those auto traits off before CFI encoding.
Current `transform_ty` attempts to avoid cycles when normalizing
`#[repr(transparent)]` types to their interior, but runs afoul of this
pattern used in `self_cell`:
```
struct X<T> {
x: u8,
p: PhantomData<T>,
}
#[repr(transparent)]
struct Y(X<Y>);
```
When attempting to normalize Y, it will still cycle indefinitely. By
using a types-visited list, this will instead get expanded exactly
one layer deep to X<Y>, and then stop, not attempting to normalize `Y`
any further.
Nicholas Nethercote