Transforms match into an assignment statement

Author: dianqk

compiler/rustc_middle/src/mir/terminator.rs

@@ -74,6 +74,12 @@ impl SwitchTargets {
     pub fn target_for_value(&self, value: u128) -> BasicBlock {
         self.iter().find_map(|(v, t)| (v == value).then_some(t)).unwrap_or_else(|| self.otherwise())
     }
+
+    /// Returns true if all targets (including the fallback target) are distinct.
+    #[inline]
+    pub fn is_distinct(&self) -> bool {
+        self.targets.iter().collect::<FxHashSet<_>>().len() == self.targets.len()
+    }
 }
 
 pub struct SwitchTargetsIter<'a> {

compiler/rustc_mir_transform/src/match_branches.rs

@@ -1,6 +1,6 @@
 use rustc_index::IndexVec;
 use rustc_middle::mir::*;
-use rustc_middle::ty::{ParamEnv, Ty, TyCtxt};
+use rustc_middle::ty::{ParamEnv, ScalarInt, Ty, TyCtxt};
 use std::iter;
 
 use super::simplify::simplify_cfg;
@@ -38,6 +38,11 @@ impl<'tcx> MirPass<'tcx> for MatchBranchSimplification {
                 should_cleanup = true;
                 continue;
             }
+            if SimplifyToExp::default().simplify(tcx, &mut body.local_decls, bbs, bb_idx, param_env)
+            {
+                should_cleanup = true;
+                continue;
+            }
         }
 
         if should_cleanup {
@@ -48,7 +53,7 @@ impl<'tcx> MirPass<'tcx> for MatchBranchSimplification {
 
 trait SimplifyMatch<'tcx> {
     fn simplify(
-        &self,
+        &mut self,
         tcx: TyCtxt<'tcx>,
         local_decls: &mut IndexVec<Local, LocalDecl<'tcx>>,
         bbs: &mut IndexVec<BasicBlock, BasicBlockData<'tcx>>,
@@ -72,7 +77,7 @@ trait SimplifyMatch<'tcx> {
         let source_info = bbs[switch_bb_idx].terminator().source_info;
         let discr_local = local_decls.push(LocalDecl::new(discr_ty, source_info.span));
 
-        // We already checked that first and second are different blocks,
+        // We already checked that targets are different blocks,
         // and bb_idx has a different terminator from both of them.
         let new_stmts = self.new_stmts(tcx, targets, param_env, bbs, discr_local.clone(), discr_ty);
         let (_, first) = targets.iter().next().unwrap();
@@ -91,7 +96,7 @@ trait SimplifyMatch<'tcx> {
     }
 
     fn can_simplify(
-        &self,
+        &mut self,
         tcx: TyCtxt<'tcx>,
         targets: &SwitchTargets,
         param_env: ParamEnv<'tcx>,
@@ -144,7 +149,7 @@ struct SimplifyToIf;
 /// ```
 impl<'tcx> SimplifyMatch<'tcx> for SimplifyToIf {
     fn can_simplify(
-        &self,
+        &mut self,
         tcx: TyCtxt<'tcx>,
         targets: &SwitchTargets,
         param_env: ParamEnv<'tcx>,
@@ -250,3 +255,207 @@ impl<'tcx> SimplifyMatch<'tcx> for SimplifyToIf {
         new_stmts.collect()
     }
 }
+
+#[derive(Default)]
+struct SimplifyToExp {
+    transfrom_types: Vec<TransfromType>,
+}
+
+#[derive(Clone, Copy)]
+enum CompareType<'tcx, 'a> {
+    Same(&'a StatementKind<'tcx>),
+    Eq(&'a Place<'tcx>, Ty<'tcx>, ScalarInt),
+    Discr(&'a Place<'tcx>, Ty<'tcx>),
+}
+
+enum TransfromType {
+    Same,
+    Eq,
+    Discr,
+}
+
+impl From<CompareType<'_, '_>> for TransfromType {
+    fn from(compare_type: CompareType<'_, '_>) -> Self {
+        match compare_type {
+            CompareType::Same(_) => TransfromType::Same,
+            CompareType::Eq(_, _, _) => TransfromType::Eq,
+            CompareType::Discr(_, _) => TransfromType::Discr,
+        }
+    }
+}
+
+/// If we find that the value of match is the same as the assignment,
+/// merge a target block statements into the source block,
+/// using cast to transform different integer types.
+///
+/// For example:
+///
+/// ```ignore (MIR)
+/// bb0: {
+///     switchInt(_1) -> [1: bb2, 2: bb3, 3: bb4, otherwise: bb1];
+/// }
+///
+/// bb1: {
+///     unreachable;
+/// }
+///
+/// bb2: {
+///     _0 = const 1_i16;
+///     goto -> bb5;
+/// }
+///
+/// bb3: {
+///     _0 = const 2_i16;
+///     goto -> bb5;
+/// }
+///
+/// bb4: {
+///     _0 = const 3_i16;
+///     goto -> bb5;
+/// }
+/// ```
+///
+/// into:
+///
+/// ```ignore (MIR)
+/// bb0: {
+///    _0 = _3 as i16 (IntToInt);
+///    goto -> bb5;
+/// }
+/// ```
+impl<'tcx> SimplifyMatch<'tcx> for SimplifyToExp {
+    fn can_simplify(
+        &mut self,
+        tcx: TyCtxt<'tcx>,
+        targets: &SwitchTargets,
+        param_env: ParamEnv<'tcx>,
+        bbs: &IndexVec<BasicBlock, BasicBlockData<'tcx>>,
+    ) -> bool {
+        if targets.iter().len() < 2 || targets.iter().len() > 64 {
+            return false;
+        }
+        // We require that the possible target blocks all be distinct.
+        if !targets.is_distinct() {
+            return false;
+        }
+        if !bbs[targets.otherwise()].is_empty_unreachable() {
+            return false;
+        }
+        let mut iter = targets.iter();
+        let (first_val, first_target) = iter.next().unwrap();
+        let first_terminator_kind = &bbs[first_target].terminator().kind;
+        // Check that destinations are identical, and if not, then don't optimize this block
+        if !targets
+            .iter()
+            .all(|(_, other_target)| first_terminator_kind == &bbs[other_target].terminator().kind)
+        {
+            return false;
+        }
+
+        let first_stmts = &bbs[first_target].statements;
+        let (second_val, second_target) = iter.next().unwrap();
+        let second_stmts = &bbs[second_target].statements;
+
+        let mut compare_types = Vec::new();
+        for (f, s) in iter::zip(first_stmts, second_stmts) {
+            let compare_type = match (&f.kind, &s.kind) {
+                // If two statements are exactly the same, we can optimize.
+                (f_s, s_s) if f_s == s_s => CompareType::Same(f_s),
+
+                // If two statements are assignments with the match values to the same place, we can optimize.
+                (
+                    StatementKind::Assign(box (lhs_f, Rvalue::Use(Operand::Constant(f_c)))),
+                    StatementKind::Assign(box (lhs_s, Rvalue::Use(Operand::Constant(s_c)))),
+                ) if lhs_f == lhs_s
+                    && f_c.const_.ty() == s_c.const_.ty()
+                    && f_c.const_.ty().is_integral() =>
+                {
+                    match (
+                        f_c.const_.try_eval_scalar_int(tcx, param_env),
+                        s_c.const_.try_eval_scalar_int(tcx, param_env),
+                    ) {
+                        (Some(f), Some(s)) if f == s => CompareType::Eq(lhs_f, f_c.const_.ty(), f),
+                        (Some(f), Some(s))
+                            if Some(f) == ScalarInt::try_from_uint(first_val, f.size())
+                                && Some(s) == ScalarInt::try_from_uint(second_val, s.size()) =>
+                        {
+                            CompareType::Discr(lhs_f, f_c.const_.ty())
+                        }
+                        _ => return false,
+                    }
+                }
+
+                // Otherwise we cannot optimize. Try another block.
+                _ => return false,
+            };
+            compare_types.push(compare_type);
+        }
+
+        for (other_val, other_target) in iter {
+            let other_stmts = &bbs[other_target].statements;
+            if compare_types.len() != other_stmts.len() {
+                return false;
+            }
+            for (f, s) in iter::zip(&compare_types, other_stmts) {
+                match (*f, &s.kind) {
+                    (CompareType::Same(f_s), s_s) if f_s == s_s => {}
+                    (
+                        CompareType::Eq(lhs_f, f_ty, val),
+                        StatementKind::Assign(box (lhs_s, Rvalue::Use(Operand::Constant(s_c)))),
+                    ) if lhs_f == lhs_s
+                        && s_c.const_.ty() == f_ty
+                        && s_c.const_.try_eval_scalar_int(tcx, param_env) == Some(val) => {}
+                    (
+                        CompareType::Discr(lhs_f, f_ty),
+                        StatementKind::Assign(box (lhs_s, Rvalue::Use(Operand::Constant(s_c)))),
+                    ) if lhs_f == lhs_s && s_c.const_.ty() == f_ty => {
+                        let Some(f) = s_c.const_.try_eval_scalar_int(tcx, param_env) else {
+                            return false;
+                        };
+                        if Some(f) != ScalarInt::try_from_uint(other_val, f.size()) {
+                            return false;
+                        }
+                    }
+                    _ => return false,
+                }
+            }
+        }
+        self.transfrom_types = compare_types.into_iter().map(|c| c.into()).collect();
+        true
+    }
+
+    fn new_stmts(
+        &self,
+        _tcx: TyCtxt<'tcx>,
+        targets: &SwitchTargets,
+        _param_env: ParamEnv<'tcx>,
+        bbs: &IndexVec<BasicBlock, BasicBlockData<'tcx>>,
+        discr_local: Local,
+        discr_ty: Ty<'tcx>,
+    ) -> Vec<Statement<'tcx>> {
+        let (_, first) = targets.iter().next().unwrap();
+        let first = &bbs[first];
+
+        let new_stmts =
+            iter::zip(&self.transfrom_types, &first.statements).map(|(t, s)| match (t, &s.kind) {
+                (TransfromType::Same, _) | (TransfromType::Eq, _) => (*s).clone(),
+                (
+                    TransfromType::Discr,
+                    StatementKind::Assign(box (lhs, Rvalue::Use(Operand::Constant(f_c)))),
+                ) => {
+                    let operand = Operand::Copy(Place::from(discr_local));
+                    let r_val = if f_c.const_.ty() == discr_ty {
+                        Rvalue::Use(operand)
+                    } else {
+                        Rvalue::Cast(CastKind::IntToInt, operand, f_c.const_.ty())
+                    };
+                    Statement {
+                        source_info: s.source_info,
+                        kind: StatementKind::Assign(Box::new((*lhs, r_val))),
+                    }
+                }
+                _ => unreachable!(),
+            });
+        new_stmts.collect()
+    }
+}

tests/codegen/match-optimized.rs

@@ -26,12 +26,12 @@ pub fn exhaustive_match(e: E) -> u8 {
 // CHECK-NEXT: store i8 1, ptr %_0, align 1
 // CHECK-NEXT: br label %[[EXIT]]
 // CHECK: [[C]]:
-// CHECK-NEXT: store i8 2, ptr %_0, align 1
+// CHECK-NEXT: store i8 3, ptr %_0, align 1
 // CHECK-NEXT: br label %[[EXIT]]
     match e {
         E::A => 0,
         E::B => 1,
-        E::C => 2,
+        E::C => 3,
     }
 }
 

tests/mir-opt/matches_reduce_branches.match_i128_u128.MatchBranchSimplification.diff

@@ -5,37 +5,42 @@
       debug i => _1;
       let mut _0: u128;
       let mut _2: i128;
++     let mut _3: i128;
 
       bb0: {
           _2 = discriminant(_1);
-          switchInt(move _2) -> [1: bb3, 2: bb4, 3: bb5, 340282366920938463463374607431768211455: bb1, otherwise: bb2];
-      }
-  
-      bb1: {
-          _0 = const _;
-          goto -> bb6;
-      }
-  
-      bb2: {
-          unreachable;
-      }
-  
-      bb3: {
-          _0 = const 1_u128;
-          goto -> bb6;
-      }
-  
-      bb4: {
-          _0 = const 2_u128;
-          goto -> bb6;
-      }
-  
-      bb5: {
-          _0 = const 3_u128;
-          goto -> bb6;
-      }
-  
-      bb6: {
+-         switchInt(move _2) -> [1: bb3, 2: bb4, 3: bb5, 340282366920938463463374607431768211455: bb1, otherwise: bb2];
+-     }
+- 
+-     bb1: {
+-         _0 = const _;
+-         goto -> bb6;
+-     }
+- 
+-     bb2: {
+-         unreachable;
+-     }
+- 
+-     bb3: {
+-         _0 = const 1_u128;
+-         goto -> bb6;
+-     }
+- 
+-     bb4: {
+-         _0 = const 2_u128;
+-         goto -> bb6;
+-     }
+- 
+-     bb5: {
+-         _0 = const 3_u128;
+-         goto -> bb6;
+-     }
+- 
+-     bb6: {
++         StorageLive(_3);
++         _3 = move _2;
++         _0 = _3 as u128 (IntToInt);
++         StorageDead(_3);
           return;
       }
   }