Auto merge of #99102 - JakobDegen:reorder-generators, r=oli-obk

Rework definition of MIR phases to more closely reflect semantic concerns

Implements most of rust-lang/compiler-team#522 .

I tried my best to restrict this PR to the "core" parts of the MCP. In other words, this includes just enough changes to make the new definition of `MirPhase` make sense. That means there are a couple of FIXMEs lying around. Depending on what reviewers prefer, I can either fix them in this PR or send follow up PRs. There are also a couple other refactorings of the `rustc_mir_transform/src/lib.rs` file that I want to do in follow ups that I didn't leave explicit FIXMEs for.

2 files changed, 102 insertions, 55 deletions

diff --git a/compiler/rustc_middle/src/mir/mod.rs b/compiler/rustc_middle/src/mir/mod.rs
index e94e1e8a10d..7784449d605 100644
--- a/compiler/rustc_middle/src/mir/mod.rs
+++ b/compiler/rustc_middle/src/mir/mod.rs
@@ -128,8 +128,20 @@ pub trait MirPass<'tcx> {
 
 impl MirPhase {
     /// Gets the index of the current MirPhase within the set of all `MirPhase`s.
+    ///
+    /// FIXME(JakobDegen): Return a `(usize, usize)` instead.
     pub fn phase_index(&self) -> usize {
-        *self as usize
+        const BUILT_PHASE_COUNT: usize = 1;
+        const ANALYSIS_PHASE_COUNT: usize = 2;
+        match self {
+            MirPhase::Built => 1,
+            MirPhase::Analysis(analysis_phase) => {
+                1 + BUILT_PHASE_COUNT + (*analysis_phase as usize)
+            }
+            MirPhase::Runtime(runtime_phase) => {
+                1 + BUILT_PHASE_COUNT + ANALYSIS_PHASE_COUNT + (*runtime_phase as usize)
+            }
+        }
     }
 }
 
diff --git a/compiler/rustc_middle/src/mir/syntax.rs b/compiler/rustc_middle/src/mir/syntax.rs
index eb90169d0e3..3426f5f43f0 100644
--- a/compiler/rustc_middle/src/mir/syntax.rs
+++ b/compiler/rustc_middle/src/mir/syntax.rs
@@ -23,75 +23,110 @@ use rustc_span::symbol::Symbol;
 use rustc_span::Span;
 use rustc_target::asm::InlineAsmRegOrRegClass;
 
-/// The various "big phases" that MIR goes through.
+/// Represents the "flavors" of MIR.
 ///
-/// These phases all describe dialects of MIR. Since all MIR uses the same datastructures, the
-/// dialects forbid certain variants or values in certain phases. The sections below summarize the
-/// changes, but do not document them thoroughly. The full documentation is found in the appropriate
-/// documentation for the thing the change is affecting.
+/// All flavors of MIR use the same data structure, but there are some important differences. These
+/// differences come in two forms: Dialects and phases.
 ///
-/// Warning: ordering of variants is significant.
+/// Dialects represent a stronger distinction than phases. This is because the transitions between
+/// dialects are semantic changes, and therefore technically *lowerings* between distinct IRs. In
+/// other words, the same [`Body`](crate::mir::Body) might be well-formed for multiple dialects, but
+/// have different semantic meaning and different behavior at runtime.
+///
+/// Each dialect additionally has a number of phases. However, phase changes never involve semantic
+/// changes. If some MIR is well-formed both before and after a phase change, it is also guaranteed
+/// that it has the same semantic meaning. In this sense, phase changes can only add additional
+/// restrictions on what MIR is well-formed.
+///
+/// When adding phases, remember to update [`MirPhase::phase_index`].
 #[derive(Copy, Clone, TyEncodable, TyDecodable, Debug, PartialEq, Eq, PartialOrd, Ord)]
 #[derive(HashStable)]
 pub enum MirPhase {
-    /// The dialect of MIR used during all phases before `DropsLowered` is the same. This is also
-    /// the MIR that analysis such as borrowck uses.
-    ///
-    /// One important thing to remember about the behavior of this section of MIR is that drop terminators
-    /// (including drop and replace) are *conditional*. The elaborate drops pass will then replace each
-    /// instance of a drop terminator with a nop, an unconditional drop, or a drop conditioned on a drop
-    /// flag. Of course, this means that it is important that the drop elaboration can accurately recognize
-    /// when things are initialized and when things are de-initialized. That means any code running on this
-    /// version of MIR must be sure to produce output that drop elaboration can reason about. See the
-    /// section on the drop terminatorss for more details.
-    Built = 0,
-    // FIXME(oli-obk): it's unclear whether we still need this phase (and its corresponding query).
-    // We used to have this for pre-miri MIR based const eval.
-    Const = 1,
-    /// This phase checks the MIR for promotable elements and takes them out of the main MIR body
-    /// by creating a new MIR body per promoted element. After this phase (and thus the termination
-    /// of the `mir_promoted` query), these promoted elements are available in the `promoted_mir`
-    /// query.
-    ConstsPromoted = 2,
-    /// After this projections may only contain deref projections as the first element.
-    Derefered = 3,
-    /// Beginning with this phase, the following variants are disallowed:
-    /// * [`TerminatorKind::DropAndReplace`]
+    /// The MIR that is generated by MIR building.
+    ///
+    /// The only things that operate on this dialect are unsafeck, the various MIR lints, and const
+    /// qualifs.
+    ///
+    /// This has no distinct phases.
+    Built,
+    /// The MIR used for most analysis.
+    ///
+    /// The only semantic change between analysis and built MIR is constant promotion. In built MIR,
+    /// sequences of statements that would generally be subject to constant promotion are
+    /// semantically constants, while in analysis MIR all constants are explicit.
+    ///
+    /// The result of const promotion is available from the `mir_promoted` and `promoted_mir` queries.
+    ///
+    /// This is the version of MIR used by borrowck and friends.
+    Analysis(AnalysisPhase),
+    /// The MIR used for CTFE, optimizations, and codegen.
+    ///
+    /// The semantic changes that occur in the lowering from analysis to runtime MIR are as follows:
+    ///
+    ///  - Drops: In analysis MIR, `Drop` terminators represent *conditional* drops; roughly speaking,
+    ///    if dataflow analysis determines that the place being dropped is uninitialized, the drop will
+    ///    not be executed. The exact semantics of this aren't written down anywhere, which means they
+    ///    are essentially "what drop elaboration does." In runtime MIR, the drops are unconditional;
+    ///    when a `Drop` terminator is reached, if the type has drop glue that drop glue is always
+    ///    executed. This may be UB if the underlying place is not initialized.
+    ///  - Packed drops: Places might in general be misaligned - in most cases this is UB, the exception
+    ///    is fields of packed structs. In analysis MIR, `Drop(P)` for a `P` that might be misaligned
+    ///    for this reason implicitly moves `P` to a temporary before dropping. Runtime MIR has no such
+    ///    rules, and dropping a misaligned place is simply UB.
+    ///  - Unwinding: in analysis MIR, unwinding from a function which may not unwind aborts. In runtime
+    ///    MIR, this is UB.
+    ///  - Retags: If `-Zmir-emit-retag` is enabled, analysis MIR has "implicit" retags in the same way
+    ///    that Rust itself has them. Where exactly these are is generally subject to change, and so we
+    ///    don't document this here. Runtime MIR has all retags explicit.
+    ///  - Generator bodies: In analysis MIR, locals may actually be behind a pointer that user code has
+    ///    access to. This occurs in generator bodies. Such locals do not behave like other locals,
+    ///    because they eg may be aliased in surprising ways. Runtime MIR has no such special locals -
+    ///    all generator bodies are lowered and so all places that look like locals really are locals.
+    ///  - Const prop lints: The lint pass which reports eg `200_u8 + 200_u8` as an error is run as a
+    ///    part of analysis to runtime MIR lowering. This means that transformations which may supress
+    ///    such errors may not run on analysis MIR.
+    Runtime(RuntimePhase),
+}
+
+/// See [`MirPhase::Analysis`].
+#[derive(Copy, Clone, TyEncodable, TyDecodable, Debug, PartialEq, Eq, PartialOrd, Ord)]
+#[derive(HashStable)]
+pub enum AnalysisPhase {
+    Initial = 0,
+    /// Beginning in this phase, the following variants are disallowed:
     /// * [`TerminatorKind::FalseUnwind`]
     /// * [`TerminatorKind::FalseEdge`]
     /// * [`StatementKind::FakeRead`]
     /// * [`StatementKind::AscribeUserType`]
     /// * [`Rvalue::Ref`] with `BorrowKind::Shallow`
     ///
-    /// And the following variant is allowed:
-    /// * [`StatementKind::Retag`]
-    ///
-    /// Furthermore, `Drop` now uses explicit drop flags visible in the MIR and reaching a `Drop`
-    /// terminator means that the auto-generated drop glue will be invoked. Also, `Copy` operands
-    /// are allowed for non-`Copy` types.
-    DropsLowered = 4,
-    /// Beginning with this phase, the following variant is disallowed:
+    /// Furthermore, `Deref` projections must be the first projection within any place (if they
+    /// appear at all)
+    PostCleanup = 1,
+}
+
+/// See [`MirPhase::Runtime`].
+#[derive(Copy, Clone, TyEncodable, TyDecodable, Debug, PartialEq, Eq, PartialOrd, Ord)]
+#[derive(HashStable)]
+pub enum RuntimePhase {
+    /// In addition to the semantic changes, beginning with this phase, the following variants are
+    /// disallowed:
+    /// * [`TerminatorKind::DropAndReplace`]
+    /// * [`TerminatorKind::Yield`]
+    /// * [`TerminatorKind::GeneratorDrop`]
     /// * [`Rvalue::Aggregate`] for any `AggregateKind` except `Array`
     ///
-    /// And the following variant is allowed:
+    /// And the following variants are allowed:
+    /// * [`StatementKind::Retag`]
     /// * [`StatementKind::SetDiscriminant`]
-    Deaggregated = 5,
-    /// Before this phase, generators are in the "source code" form, featuring `yield` statements
-    /// and such. With this phase change, they are transformed into a proper state machine. Running
-    /// optimizations before this change can be potentially dangerous because the source code is to
-    /// some extent a "lie." In particular, `yield` terminators effectively make the value of all
-    /// locals visible to the caller. This means that dead store elimination before them, or code
-    /// motion across them, is not correct in general. This is also exasperated by type checking
-    /// having pre-computed a list of the types that it thinks are ok to be live across a yield
-    /// point - this is necessary to decide eg whether autotraits are implemented. Introducing new
-    /// types across a yield point will lead to ICEs becaues of this.
-    ///
-    /// Beginning with this phase, the following variants are disallowed:
-    /// * [`TerminatorKind::Yield`]
-    /// * [`TerminatorKind::GeneratorDrop`]
+    /// * [`StatementKind::Deinit`]
+    ///
+    /// Furthermore, `Copy` operands are allowed for non-`Copy` types.
+    Initial = 0,
+    /// Beginning with this phase, the following variant is disallowed:
     /// * [`ProjectionElem::Deref`] of `Box`
-    GeneratorsLowered = 6,
-    Optimized = 7,
+    PostCleanup = 1,
+    Optimized = 2,
 }
 
 ///////////////////////////////////////////////////////////////////////////