Add documentation for the semantics of MIR rvalues

Author: JakobDegen

compiler/rustc_middle/src/lib.rs

@@ -59,6 +59,7 @@
 #![feature(unwrap_infallible)]
 #![feature(decl_macro)]
 #![feature(drain_filter)]
+#![feature(intra_doc_pointers)]
 #![recursion_limit = "512"]
 #![allow(rustc::potential_query_instability)]
 

compiler/rustc_middle/src/mir/mod.rs

@@ -2378,57 +2378,134 @@ impl<'tcx> Operand<'tcx> {
 #[derive(Clone, TyEncodable, TyDecodable, Hash, HashStable, PartialEq)]
 /// The various kinds of rvalues that can appear in MIR.
 ///
-/// Not all of these are allowed at every [`MirPhase`]. Check the documentation there to see which
-/// ones you do not have to worry about. The MIR validator will generally enforce such restrictions,
-/// causing an ICE if they are violated.
+/// Not all of these are allowed at every [`MirPhase`] - when this is the case, it's stated below.
+///
+/// Computing any rvalue begins by evaluating the places and operands in the rvalue in the order in
+/// which they appear. These are then used to produce a "value" - the same kind of value that an
+/// [`Operand`] is.
 pub enum Rvalue<'tcx> {
-    /// x (either a move or copy, depending on type of x)
+    /// Yields the operand unchanged
     Use(Operand<'tcx>),
 
-    /// [x; 32]
+    /// Creates an array where each element is the value of the operand. This currently does not
+    /// drop the value even if the number of repetitions is zero, see [#74836].
+    ///
+    /// Corresponds to source code like `[x; 32]`.
+    ///
+    /// [#74836]: https://github.com/rust-lang/rust/issues/74836
     Repeat(Operand<'tcx>, ty::Const<'tcx>),
 
-    /// &x or &mut x
+    /// Creates a reference of the indicated kind to the place.
+    ///
+    /// There is not much to document here, because besides the obvious parts the semantics of this
+    /// are essentially entirely a part of the aliasing model. There are many UCG issues discussing
+    /// exactly what the behavior of this operation should be.
+    ///
+    /// `Shallow` borrows are disallowed after drop lowering.
     Ref(Region<'tcx>, BorrowKind, Place<'tcx>),
 
-    /// Accessing a thread local static. This is inherently a runtime operation, even if llvm
-    /// treats it as an access to a static. This `Rvalue` yields a reference to the thread local
-    /// static.
+    /// Returns a pointer/reference to the given thread local.
+    ///
+    /// The yielded type is a `*mut T` if the static is mutable, otherwise if the static is extern a
+    /// `*const T`, and if neither of those apply a `&T`.
+    ///
+    /// **Note:** This is a runtime operation that actually executes code and is in this sense more
+    /// like a function call. Also, DSEing these causes `fn main() {}` to SIGILL for some reason
+    /// that I never got a chance to look into.
+    ///
+    /// **Needs clarification**: Are there weird additional semantics here related to the runtime
+    /// nature of this operation?
     ThreadLocalRef(DefId),
 
-    /// Create a raw pointer to the given place
-    /// Can be generated by raw address of expressions (`&raw const x`),
-    /// or when casting a reference to a raw pointer.
+    /// Creates a pointer with the indicated mutability to the place.
+    ///
+    /// This is generated by pointer casts like `&v as *const _` or raw address of expressions like
+    /// `&raw v` or `addr_of!(v)`.
+    ///
+    /// Like with references, the semantics of this operation are heavily dependent on the aliasing
+    /// model.
     AddressOf(Mutability, Place<'tcx>),
 
-    /// length of a `[X]` or `[X;n]` value
+    /// Yields the length of the place, as a `usize`.
+    ///
+    /// If the type of the place is an array, this is the array length. This also works for slices
+    /// (`[T]`, not `&[T]`) through some mechanism that depends on how exactly places work (see
+    /// there for more details).
     Len(Place<'tcx>),
 
+    /// Performs essentially all of the casts that can be performed via `as`.
+    ///
+    /// This allows for casts from/to a variety of types.
+    ///
+    /// **FIXME**: Document exactly which `CastKind`s allow which types of casts. Figure out why
+    /// `ArrayToPointer` and `MutToConstPointer` are special.
     Cast(CastKind, Operand<'tcx>, Ty<'tcx>),
 
+    /// * `Offset` has the same semantics as [`offset`](pointer::offset), except that the second
+    ///   paramter may be a `usize` as well.
+    /// * The comparison operations accept `bool`s, `char`s, signed or unsigned integers, floats,
+    ///   raw pointers, or function pointers and return a `bool`.
+    /// * Left and right shift operations accept signed or unsigned integers not necessarily of the
+    ///   same type and return a value of the same type as their LHS. For all other operations, the
+    ///   types of the operands must match.
+    /// * The `Bit*` operations accept signed integers, unsigned integers, or bools and return a
+    ///   value of that type.
+    /// * The remaining operations accept signed integers, unsigned integers, or floats of any
+    ///   matching type and return a value of that type.
     BinaryOp(BinOp, Box<(Operand<'tcx>, Operand<'tcx>)>),
+
+    /// Same as `BinaryOp`, but yields `(T, bool)` instead of `T`. In addition to performing the
+    /// same computation as the matching `BinaryOp`, checks if the infinite precison result would be
+    /// unequal to the actual result and sets the `bool` if this is the case. `BinOp::Offset` is not
+    /// allowed here.
+    ///
+    /// **FIXME**: What about division/modulo? Are they allowed here at all? Are zero divisors still
+    /// UB? Also, which other combinations of types are disallowed?
     CheckedBinaryOp(BinOp, Box<(Operand<'tcx>, Operand<'tcx>)>),
 
+    /// Yields the size or alignment of the type as a `usize`.
     NullaryOp(NullOp, Ty<'tcx>),
+
+    /// Exactly like `BinaryOp`, but less operands.
+    ///
+    /// Also does two's-complement arithmetic. Negation requires a signed integer or a float; binary
+    /// not requires a signed integer, unsigned integer, or bool. Both operation kinds return a
+    /// value with the same type as their operand.
     UnaryOp(UnOp, Operand<'tcx>),
 
-    /// Read the discriminant of an ADT.
+    /// Computes the discriminant of the place, returning it as an integer of type
+    /// [`discriminant_ty`].
+    ///
+    /// The validity requirements for the underlying value are undecided for this rvalue, see
+    /// [#91095]. Note too that the value of the discriminant is not the same thing as the
+    /// variant index; use [`discriminant_for_variant`] to convert.
+    ///
+    /// For types defined in the source code as enums, this is well behaved. This is also well
+    /// formed for other types, but yields no particular value - there is no reason it couldn't be
+    /// defined to yield eg zero though.
     ///
-    /// Undefined (i.e., no effort is made to make it defined, but there’s no reason why it cannot
-    /// be defined to return, say, a 0) if ADT is not an enum.
+    /// [`discriminant_ty`]: crate::ty::Ty::discriminant_ty
+    /// [#91095]: https://github.com/rust-lang/rust/issues/91095
+    /// [`discriminant_for_variant`]: crate::ty::Ty::discriminant_for_variant
     Discriminant(Place<'tcx>),
 
-    /// Creates an aggregate value, like a tuple or struct. This is
-    /// only needed because we want to distinguish `dest = Foo { x:
-    /// ..., y: ... }` from `dest.x = ...; dest.y = ...;` in the case
-    /// that `Foo` has a destructor. These rvalues can be optimized
-    /// away after type-checking and before lowering.
+    /// Creates an aggregate value, like a tuple or struct.
+    ///
+    /// This is needed because dataflow analysis needs to distinguish
+    /// `dest = Foo { x: ..., y: ... }` from `dest.x = ...; dest.y = ...;` in the case that `Foo`
+    /// has a destructor.
+    ///
+    /// Disallowed after deaggregation for all aggregate kinds except `Array` and `Generator`. After
+    /// generator lowering, `Generator` aggregate kinds are disallowed too.
     Aggregate(Box<AggregateKind<'tcx>>, Vec<Operand<'tcx>>),
 
     /// Transmutes a `*mut u8` into shallow-initialized `Box<T>`.
     ///
-    /// This is different a normal transmute because dataflow analysis will treat the box
-    /// as initialized but its content as uninitialized.
+    /// This is different a normal transmute because dataflow analysis will treat the box as
+    /// initialized but its content as uninitialized. Like other pointer casts, this in general
+    /// affects alias analysis.
+    ///
+    /// Disallowed after drop elaboration.
     ShallowInitBox(Operand<'tcx>, Ty<'tcx>),
 }