Adjust documentation of SetDiscriminant to describe the semantics more precisely.

The changed semantics also make one analysis work better as written, so we remove the documentation
in it that is now out of date.

Author: JakobDegen

compiler/rustc_middle/src/mir/mod.rs

@@ -1553,7 +1553,21 @@ pub enum StatementKind<'tcx> {
     /// never accessed still get some sanity checks for, e.g., `let x: ! = ..;`
     FakeRead(Box<(FakeReadCause, Place<'tcx>)>),
 
-    /// Write the discriminant for a variant to the enum Place.
+    /// Initialize the place with the given variant and all fields uninit.
+    ///
+    /// If `place` is an ADT, this corresponds to the Rust code `place = Variant(uninit, uninit,
+    /// uninit)`. If each of the fields is initialized after the `SetDiscriminant`, the place is
+    /// completely initialized. Of course, you cannot initialize the fields first, as
+    /// `SetDiscriminant` invalidates them.
+    ///
+    /// Computing the `Rvalue::Discriminant` of this place after `SetDiscriminant` is not
+    /// necessarily well defined if the fields have not also been initialized. See [#91095][91095]
+    /// for discussion around this topic.
+    ///
+    /// If `place` is a generator, then this invalidates only those fields which are not also
+    /// present in another variant. Those fields that are present in another variant are unchanged.
+    ///
+    /// [91095]: https://github.com/rust-lang/rust/issues/91095
     SetDiscriminant { place: Box<Place<'tcx>>, variant_index: VariantIdx },
 
     /// Start a live range for the storage of the local.

compiler/rustc_mir_dataflow/src/impls/liveness.rs

@@ -18,30 +18,6 @@ use crate::{AnalysisDomain, Backward, CallReturnPlaces, GenKill, GenKillAnalysis
 /// such an assignment is currently marked as a "use" of `x` in an attempt to be maximally
 /// conservative.
 ///
-/// ## Enums and `SetDiscriminant`
-///
-/// Assigning a literal value to an `enum` (e.g. `Option<i32>`), does not result in a simple
-/// assignment of the form `_1 = /*...*/` in the MIR. For example, the following assignment to `x`:
-///
-/// ```
-/// x = Some(4);
-/// ```
-///
-/// compiles to this MIR
-///
-/// ```
-/// ((_1 as Some).0: i32) = const 4_i32;
-/// discriminant(_1) = 1;
-/// ```
-///
-/// However, `MaybeLiveLocals` **does** mark `x` (`_1`) as "killed" after a statement like this.
-/// That's because it treats the `SetDiscriminant` operation as a definition of `x`, even though
-/// the writes that actually initialized the locals happened earlier.
-///
-/// This makes `MaybeLiveLocals` unsuitable for certain classes of optimization normally associated
-/// with a live variables analysis, notably dead-store elimination. It's a dirty hack, but it works
-/// okay for the generator state transform (currently the main consumuer of this analysis).
-///
 /// [`MaybeBorrowedLocals`]: super::MaybeBorrowedLocals
 /// [flow-test]: https://github.com/rust-lang/rust/blob/a08c47310c7d49cbdc5d7afb38408ba519967ecd/src/test/ui/mir-dataflow/liveness-ptr.rs
 /// [liveness]: https://en.wikipedia.org/wiki/Live_variable_analysis