Bounds-check with PtrMetadata instead of Len in MIR

Author: scottmcm

compiler/rustc_const_eval/src/check_consts/check.rs

@@ -573,12 +573,27 @@ impl<'tcx> Visitor<'tcx> for Checker<'_, 'tcx> {
             ) => {}
             Rvalue::ShallowInitBox(_, _) => {}
 
-            Rvalue::UnaryOp(_, operand) => {
+            Rvalue::UnaryOp(op, operand) => {
                 let ty = operand.ty(self.body, self.tcx);
-                if is_int_bool_float_or_char(ty) {
-                    // Int, bool, float, and char operations are fine.
-                } else {
-                    span_bug!(self.span, "non-primitive type in `Rvalue::UnaryOp`: {:?}", ty);
+                match op {
+                    UnOp::Not | UnOp::Neg => {
+                        if is_int_bool_float_or_char(ty) {
+                            // Int, bool, float, and char operations are fine.
+                        } else {
+                            span_bug!(
+                                self.span,
+                                "non-primitive type in `Rvalue::UnaryOp{op:?}`: {ty:?}",
+                            );
+                        }
+                    }
+                    UnOp::PtrMetadata => {
+                        if !ty.is_ref() && !ty.is_unsafe_ptr() {
+                            span_bug!(
+                                self.span,
+                                "non-pointer type in `Rvalue::UnaryOp({op:?})`: {ty:?}",
+                            );
+                        }
+                    }
                 }
             }
 

compiler/rustc_mir_build/src/build/expr/as_place.rs

@@ -630,6 +630,63 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
         block.and(base_place.index(idx))
     }
 
+    /// Given a place that's either an array or a slice, returns an operand
+    /// with the length of the array/slice.
+    ///
+    /// For arrays it'll be `Operand::Constant` with the actual length;
+    /// For slices it'll be `Operand::Move` of a local using `PtrMetadata`.
+    fn len_of_slice_or_array(
+        &mut self,
+        block: BasicBlock,
+        slice: Place<'tcx>,
+        expr_span: Span,
+        source_info: SourceInfo,
+    ) -> Operand<'tcx> {
+        let slice_ty = slice.ty(&self.local_decls, self.tcx).ty;
+        let usize_ty = self.tcx.types.usize;
+        if let ty::Array(_elem_ty, len_const) = slice_ty.kind() {
+            Operand::Constant(Box::new(ConstOperand {
+                span: expr_span,
+                user_ty: None,
+                const_: Const::from_ty_const(len_const.clone(), usize_ty, self.tcx),
+            }))
+        } else if slice_ty.is_slice() {
+            let ptr_or_ref = if let [PlaceElem::Deref] = slice.projection[..]
+                && let local_ty = self.local_decls[slice.local].ty
+                && local_ty.is_trivially_pure_clone_copy()
+            {
+                // It's extremely common that we have something that can be
+                // directly passed to `PtrMetadata`, so avoid an unnecessary
+                // temporary and statement in those cases. Note that we can
+                // only do that for `Copy` types -- not `&mut [_]` -- because
+                // the MIR we're building here needs to pass NLL later.
+                Operand::Copy(Place::from(slice.local))
+            } else {
+                let ptr_ty = Ty::new_imm_ptr(self.tcx, slice_ty);
+                let ptr = self.temp(ptr_ty, expr_span);
+                self.cfg.push_assign(
+                    block,
+                    source_info,
+                    ptr,
+                    Rvalue::RawPtr(Mutability::Not, slice),
+                );
+                Operand::Move(ptr)
+            };
+
+            let len = self.temp(usize_ty, expr_span);
+            self.cfg.push_assign(
+                block,
+                source_info,
+                len,
+                Rvalue::UnaryOp(UnOp::PtrMetadata, ptr_or_ref),
+            );
+
+            Operand::Move(len)
+        } else {
+            span_bug!(expr_span, "len called on something that's neither slice nor array")
+        }
+    }
+
     fn bounds_check(
         &mut self,
         block: BasicBlock,
@@ -638,25 +695,25 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
         expr_span: Span,
         source_info: SourceInfo,
     ) -> BasicBlock {
-        let usize_ty = self.tcx.types.usize;
-        let bool_ty = self.tcx.types.bool;
-        // bounds check:
-        let len = self.temp(usize_ty, expr_span);
-        let lt = self.temp(bool_ty, expr_span);
+        let slice = slice.to_place(self);
 
         // len = len(slice)
-        self.cfg.push_assign(block, source_info, len, Rvalue::Len(slice.to_place(self)));
+        let len = self.len_of_slice_or_array(block, slice, expr_span, source_info);
+
         // lt = idx < len
+        let bool_ty = self.tcx.types.bool;
+        let lt = self.temp(bool_ty, expr_span);
         self.cfg.push_assign(
             block,
             source_info,
             lt,
             Rvalue::BinaryOp(
                 BinOp::Lt,
-                Box::new((Operand::Copy(Place::from(index)), Operand::Copy(len))),
+                Box::new((Operand::Copy(Place::from(index)), len.to_copy())),
             ),
         );
-        let msg = BoundsCheck { len: Operand::Move(len), index: Operand::Copy(Place::from(index)) };
+        let msg = BoundsCheck { len, index: Operand::Copy(Place::from(index)) };
+
         // assert!(lt, "...")
         self.assert(block, Operand::Move(lt), true, msg, expr_span)
     }

compiler/rustc_mir_transform/src/instsimplify.rs

@@ -47,7 +47,6 @@ impl<'tcx> crate::MirPass<'tcx> for InstSimplify {
                         }
                         ctx.simplify_bool_cmp(rvalue);
                         ctx.simplify_ref_deref(rvalue);
-                        ctx.simplify_len(rvalue);
                         ctx.simplify_ptr_aggregate(rvalue);
                         ctx.simplify_cast(rvalue);
                     }
@@ -132,18 +131,6 @@ impl<'tcx> InstSimplifyContext<'_, 'tcx> {
         }
     }
 
-    /// Transform `Len([_; N])` ==> `N`.
-    fn simplify_len(&self, rvalue: &mut Rvalue<'tcx>) {
-        if let Rvalue::Len(ref place) = *rvalue {
-            let place_ty = place.ty(self.local_decls, self.tcx).ty;
-            if let ty::Array(_, len) = *place_ty.kind() {
-                let const_ = Const::from_ty_const(len, self.tcx.types.usize, self.tcx);
-                let constant = ConstOperand { span: DUMMY_SP, const_, user_ty: None };
-                *rvalue = Rvalue::Use(Operand::Constant(Box::new(constant)));
-            }
-        }
-    }
-
     /// Transform `Aggregate(RawPtr, [p, ()])` ==> `Cast(PtrToPtr, p)`.
     fn simplify_ptr_aggregate(&self, rvalue: &mut Rvalue<'tcx>) {
         if let Rvalue::Aggregate(box AggregateKind::RawPtr(pointee_ty, mutability), fields) = rvalue

compiler/rustc_mir_transform/src/validate.rs

@@ -1130,14 +1130,6 @@ impl<'a, 'tcx> Visitor<'tcx> for TypeChecker<'a, 'tcx> {
                         );
                     }
                     UnOp::PtrMetadata => {
-                        if !matches!(self.mir_phase, MirPhase::Runtime(_)) {
-                            // It would probably be fine to support this in earlier phases, but at
-                            // the time of writing it's only ever introduced from intrinsic
-                            // lowering or other runtime-phase optimization passes, so earlier
-                            // things can just `bug!` on it.
-                            self.fail(location, "PtrMetadata should be in runtime MIR only");
-                        }
-
                         check_kinds!(
                             a,
                             "Cannot PtrMetadata non-pointer non-reference type {:?}",

tests/mir-opt/array_index_is_temporary.main.SimplifyCfg-pre-optimizations.after.panic-abort.mir

@@ -7,8 +7,7 @@ fn main() -> () {
     let mut _5: u32;
     let mut _6: *mut usize;
     let _7: usize;
-    let mut _8: usize;
-    let mut _9: bool;
+    let mut _8: bool;
     scope 1 {
         debug x => _1;
         let mut _2: usize;
@@ -41,9 +40,8 @@ fn main() -> () {
         StorageDead(_6);
         StorageLive(_7);
         _7 = copy _2;
-        _8 = Len(_1);
-        _9 = Lt(copy _7, copy _8);
-        assert(move _9, "index out of bounds: the length is {} but the index is {}", move _8, copy _7) -> [success: bb2, unwind unreachable];
+        _8 = Lt(copy _7, const 3_usize);
+        assert(move _8, "index out of bounds: the length is {} but the index is {}", const 3_usize, copy _7) -> [success: bb2, unwind unreachable];
     }
 
     bb2: {

tests/mir-opt/array_index_is_temporary.main.SimplifyCfg-pre-optimizations.after.panic-unwind.mir

@@ -7,8 +7,7 @@ fn main() -> () {
     let mut _5: u32;
     let mut _6: *mut usize;
     let _7: usize;
-    let mut _8: usize;
-    let mut _9: bool;
+    let mut _8: bool;
     scope 1 {
         debug x => _1;
         let mut _2: usize;
@@ -41,9 +40,8 @@ fn main() -> () {
         StorageDead(_6);
         StorageLive(_7);
         _7 = copy _2;
-        _8 = Len(_1);
-        _9 = Lt(copy _7, copy _8);
-        assert(move _9, "index out of bounds: the length is {} but the index is {}", move _8, copy _7) -> [success: bb2, unwind continue];
+        _8 = Lt(copy _7, const 3_usize);
+        assert(move _8, "index out of bounds: the length is {} but the index is {}", const 3_usize, copy _7) -> [success: bb2, unwind continue];
     }
 
     bb2: {

tests/mir-opt/building/index_array_and_slice.index_array.built.after.mir

@@ -0,0 +1,30 @@
+// MIR for `index_array` after built
+
+fn index_array(_1: &[i32; 7], _2: usize) -> &i32 {
+    debug array => _1;
+    debug index => _2;
+    let mut _0: &i32;
+    let _3: &i32;
+    let _4: usize;
+    let mut _5: bool;
+
+    bb0: {
+        StorageLive(_3);
+        StorageLive(_4);
+        _4 = copy _2;
+        _5 = Lt(copy _4, const 7_usize);
+        assert(move _5, "index out of bounds: the length is {} but the index is {}", const 7_usize, copy _4) -> [success: bb1, unwind: bb2];
+    }
+
+    bb1: {
+        _3 = &(*_1)[_4];
+        _0 = &(*_3);
+        StorageDead(_4);
+        StorageDead(_3);
+        return;
+    }
+
+    bb2 (cleanup): {
+        resume;
+    }
+}

tests/mir-opt/building/index_array_and_slice.index_custom.built.after.mir

@@ -0,0 +1,34 @@
+// MIR for `index_custom` after built
+
+fn index_custom(_1: &WithSliceTail, _2: usize) -> &i32 {
+    debug custom => _1;
+    debug index => _2;
+    let mut _0: &i32;
+    let _3: &i32;
+    let _4: usize;
+    let mut _5: *const [i32];
+    let mut _6: usize;
+    let mut _7: bool;
+
+    bb0: {
+        StorageLive(_3);
+        StorageLive(_4);
+        _4 = copy _2;
+        _5 = &raw const ((*_1).1: [i32]);
+        _6 = PtrMetadata(move _5);
+        _7 = Lt(copy _4, copy _6);
+        assert(move _7, "index out of bounds: the length is {} but the index is {}", move _6, copy _4) -> [success: bb1, unwind: bb2];
+    }
+
+    bb1: {
+        _3 = &((*_1).1: [i32])[_4];
+        _0 = &(*_3);
+        StorageDead(_4);
+        StorageDead(_3);
+        return;
+    }
+
+    bb2 (cleanup): {
+        resume;
+    }
+}

tests/mir-opt/building/index_array_and_slice.index_mut_slice.built.after.mir

@@ -0,0 +1,34 @@
+// MIR for `index_mut_slice` after built
+
+fn index_mut_slice(_1: &mut [i32], _2: usize) -> &i32 {
+    debug slice => _1;
+    debug index => _2;
+    let mut _0: &i32;
+    let _3: &i32;
+    let _4: usize;
+    let mut _5: *const [i32];
+    let mut _6: usize;
+    let mut _7: bool;
+
+    bb0: {
+        StorageLive(_3);
+        StorageLive(_4);
+        _4 = copy _2;
+        _5 = &raw const (*_1);
+        _6 = PtrMetadata(move _5);
+        _7 = Lt(copy _4, copy _6);
+        assert(move _7, "index out of bounds: the length is {} but the index is {}", move _6, copy _4) -> [success: bb1, unwind: bb2];
+    }
+
+    bb1: {
+        _3 = &(*_1)[_4];
+        _0 = &(*_3);
+        StorageDead(_4);
+        StorageDead(_3);
+        return;
+    }
+
+    bb2 (cleanup): {
+        resume;
+    }
+}

tests/mir-opt/building/index_array_and_slice.index_slice.built.after.mir

@@ -0,0 +1,32 @@
+// MIR for `index_slice` after built
+
+fn index_slice(_1: &[i32], _2: usize) -> &i32 {
+    debug slice => _1;
+    debug index => _2;
+    let mut _0: &i32;
+    let _3: &i32;
+    let _4: usize;
+    let mut _5: usize;
+    let mut _6: bool;
+
+    bb0: {
+        StorageLive(_3);
+        StorageLive(_4);
+        _4 = copy _2;
+        _5 = PtrMetadata(copy _1);
+        _6 = Lt(copy _4, copy _5);
+        assert(move _6, "index out of bounds: the length is {} but the index is {}", move _5, copy _4) -> [success: bb1, unwind: bb2];
+    }
+
+    bb1: {
+        _3 = &(*_1)[_4];
+        _0 = &(*_3);
+        StorageDead(_4);
+        StorageDead(_3);
+        return;
+    }
+
+    bb2 (cleanup): {
+        resume;
+    }
+}