add init_box_via_move intrinsic and use it for vec!

This allows us to get rid of box_new entirely

Author: RalfJung

compiler/rustc_borrowck/src/type_check/mod.rs

@@ -1537,15 +1537,17 @@ impl<'a, 'tcx> Visitor<'tcx> for TypeChecker<'a, 'tcx> {
                         }
                     }
                     CastKind::Transmute => {
-                        let ty_from = op.ty(self.body, tcx);
-                        match ty_from.kind() {
-                            ty::Pat(base, _) if base == ty => {}
-                            _ => span_mirbug!(
-                                self,
-                                rvalue,
-                                "Unexpected CastKind::Transmute {ty_from:?} -> {ty:?}, which is not permitted in Analysis MIR",
-                            ),
-                        }
+                        // FIXME: `init_box_via_move` lowering really wants to use this.
+                        // What do we have to do here?
+                        // let ty_from = op.ty(self.body, tcx);
+                        // match ty_from.kind() {
+                        //     ty::Pat(base, _) if base == ty => {}
+                        //     _ => span_mirbug!(
+                        //         self,
+                        //         rvalue,
+                        //         "Unexpected CastKind::Transmute {ty_from:?} -> {ty:?}, which is not permitted in Analysis MIR",
+                        //     ),
+                        // }
                     }
                     CastKind::Subtype => {
                         bug!("CastKind::Subtype shouldn't exist in borrowck")

compiler/rustc_hir_analysis/src/check/intrinsic.rs

@@ -76,7 +76,6 @@ fn intrinsic_operation_unsafety(tcx: TyCtxt<'_>, intrinsic_id: LocalDefId) -> hi
         | sym::autodiff
         | sym::bitreverse
         | sym::black_box
-        | sym::box_new
         | sym::breakpoint
         | sym::bswap
         | sym::caller_location
@@ -132,6 +131,7 @@ fn intrinsic_operation_unsafety(tcx: TyCtxt<'_>, intrinsic_id: LocalDefId) -> hi
         | sym::forget
         | sym::frem_algebraic
         | sym::fsub_algebraic
+        | sym::init_box_via_move
         | sym::is_val_statically_known
         | sym::log2f16
         | sym::log2f32
@@ -553,6 +553,12 @@ pub(crate) fn check_intrinsic_type(
         sym::write_via_move => {
             (1, 0, vec![Ty::new_mut_ptr(tcx, param(0)), param(0)], tcx.types.unit)
         }
+        sym::init_box_via_move => {
+            let t = param(0);
+            let maybe_uninit_t = Ty::new_maybe_uninit(tcx, t);
+
+            (1, 0, vec![Ty::new_box(tcx, maybe_uninit_t), param(0)], Ty::new_box(tcx, t))
+        }
 
         sym::typed_swap_nonoverlapping => {
             (1, 0, vec![Ty::new_mut_ptr(tcx, param(0)); 2], tcx.types.unit)
@@ -645,8 +651,6 @@ pub(crate) fn check_intrinsic_type(
 
         sym::ub_checks => (0, 0, Vec::new(), tcx.types.bool),
 
-        sym::box_new => (1, 0, vec![param(0)], Ty::new_box(tcx, param(0))),
-
         // contract_check_requires::<C>(C) -> bool, where C: impl Fn() -> bool
         sym::contract_check_requires => (1, 0, vec![param(0)], tcx.types.unit),
         sym::contract_check_ensures => {

compiler/rustc_hir_typeck/src/fn_ctxt/suggestions.rs

@@ -3022,6 +3022,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> {
                     {
                         let deref_kind = if checked_ty.is_box() {
                             // detect Box::new(..)
+                            // FIXME: use `box_new` diagnostic item instead?
                             if let ExprKind::Call(box_new, [_]) = expr.kind
                                 && let ExprKind::Path(qpath) = &box_new.kind
                                 && let Res::Def(DefKind::AssocFn, fn_id) =

compiler/rustc_mir_build/src/builder/expr/into.rs

@@ -365,30 +365,112 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
                     None
                 })
             }
-            // The `write_via_move` intrinsic needs to be special-cased very early to avoid
-            // introducing unnecessary copies that can be hard to remove again later:
-            // `write_via_move(ptr, val)` becomes `*ptr = val` but without any dropping.
+            // Some intrinsics are handled here because they desperately want to avoid introducing
+            // unnecessary copies.
             ExprKind::Call { ty, fun, ref args, .. }
-                if let ty::FnDef(def_id, _generic_args) = ty.kind()
+                if let ty::FnDef(def_id, generic_args) = ty.kind()
                     && let Some(intrinsic) = this.tcx.intrinsic(def_id)
-                    && intrinsic.name == sym::write_via_move =>
+                    && matches!(intrinsic.name, sym::write_via_move | sym::init_box_via_move) =>
             {
                 // We still have to evaluate the callee expression as normal (but we don't care
                 // about its result).
                 let _fun = unpack!(block = this.as_local_operand(block, fun));
-                // The destination must have unit type (so we don't actually have to store anything
-                // into it).
-                assert!(destination.ty(&this.local_decls, this.tcx).ty.is_unit());
 
-                // Compile this to an assignment of the argument into the destination.
-                let [ptr, val] = **args else {
-                    span_bug!(expr_span, "invalid write_via_move call")
-                };
-                let Some(ptr) = unpack!(block = this.as_local_operand(block, ptr)).place() else {
-                    span_bug!(expr_span, "invalid write_via_move call")
-                };
-                let ptr_deref = ptr.project_deeper(&[ProjectionElem::Deref], this.tcx);
-                this.expr_into_dest(ptr_deref, block, val)
+                match intrinsic.name {
+                    sym::write_via_move => {
+                        // `write_via_move(ptr, val)` becomes `*ptr = val` but without any dropping.
+
+                        // The destination must have unit type (so we don't actually have to store anything
+                        // into it).
+                        assert!(destination.ty(&this.local_decls, this.tcx).ty.is_unit());
+
+                        // Compile this to an assignment of the argument into the destination.
+                        let [ptr, val] = **args else {
+                            span_bug!(expr_span, "invalid write_via_move call")
+                        };
+                        let Some(ptr) = unpack!(block = this.as_local_operand(block, ptr)).place()
+                        else {
+                            span_bug!(expr_span, "invalid write_via_move call")
+                        };
+                        let ptr_deref = ptr.project_deeper(&[ProjectionElem::Deref], this.tcx);
+                        this.expr_into_dest(ptr_deref, block, val)
+                    }
+                    sym::init_box_via_move => {
+                        // `write_via_move(b, val)` becomes
+                        // ```
+                        // *transmute::<_, *mut T>(b) = val;
+                        // transmute::<_, Box<T>>(b)
+                        // ```
+                        let t = generic_args.type_at(0);
+                        let [b, val] = **args else {
+                            span_bug!(expr_span, "invalid init_box_via_move call")
+                        };
+                        let Some(b) = unpack!(block = this.as_local_operand(block, b)).place()
+                        else {
+                            span_bug!(expr_span, "invalid init_box_via_move call")
+                        };
+                        // Project to the pointer inside `b`. We have to keep `b` in scope to ensure
+                        // it gets dropped. After the first projection we can transmute which is
+                        // easier.
+                        let ty::Adt(box_adt_def, box_adt_args) =
+                            b.ty(&this.local_decls, this.tcx).ty.kind()
+                        else {
+                            span_bug!(expr_span, "invalid init_box_via_move call")
+                        };
+                        let unique_field =
+                            this.tcx.adt_def(box_adt_def.did()).non_enum_variant().fields
+                                [rustc_abi::FieldIdx::ZERO]
+                                .did;
+                        let Some(unique_def) =
+                            this.tcx.type_of(unique_field).instantiate_identity().ty_adt_def()
+                        else {
+                            span_bug!(
+                                this.tcx.def_span(unique_field),
+                                "expected Box to contain Unique"
+                            )
+                        };
+                        let unique_ty =
+                            Ty::new_adt(this.tcx, unique_def, this.tcx.mk_args(&[box_adt_args[0]]));
+                        let b_field = b.project_deeper(
+                            &[ProjectionElem::Field(rustc_abi::FieldIdx::ZERO, unique_ty)],
+                            this.tcx,
+                        );
+                        // `ptr` is `b` transmuted to `*mut T`.
+                        let ptr_ty = Ty::new_mut_ptr(this.tcx, t);
+                        let ptr = this.local_decls.push(LocalDecl::new(ptr_ty, expr_span));
+                        this.cfg.push(
+                            block,
+                            Statement::new(source_info, StatementKind::StorageLive(ptr)),
+                        );
+                        // Make sure `StorageDead` gets emitted.
+                        this.schedule_drop_storage_and_value(expr_span, this.local_scope(), ptr);
+                        this.cfg.push_assign(
+                            block,
+                            source_info,
+                            Place::from(ptr),
+                            // Needs to be a `Copy` so that `b` still gets dropped if `val` panics.
+                            Rvalue::Cast(CastKind::Transmute, Operand::Copy(b_field), ptr_ty),
+                        );
+                        // Store `val` into `ptr`.
+                        let ptr_deref =
+                            Place::from(ptr).project_deeper(&[ProjectionElem::Deref], this.tcx);
+                        unpack!(block = this.expr_into_dest(ptr_deref, block, val));
+                        // Return `ptr` transmuted to `Box<T>`.
+                        this.cfg.push_assign(
+                            block,
+                            source_info,
+                            destination,
+                            Rvalue::Cast(
+                                CastKind::Transmute,
+                                // Move from `b` so that does not get dropped any more.
+                                Operand::Move(b),
+                                Ty::new_box(this.tcx, t),
+                            ),
+                        );
+                        block.unit()
+                    }
+                    _ => rustc_middle::bug!(),
+                }
             }
             ExprKind::Call { ty: _, fun, ref args, from_hir_call, fn_span } => {
                 let fun = unpack!(block = this.as_local_operand(block, fun));

compiler/rustc_mir_build/src/thir/cx/expr.rs

@@ -20,7 +20,7 @@ use rustc_middle::ty::{
     self, AdtKind, GenericArgs, InlineConstArgs, InlineConstArgsParts, ScalarInt, Ty, UpvarArgs,
 };
 use rustc_middle::{bug, span_bug};
-use rustc_span::{Span, sym};
+use rustc_span::Span;
 use tracing::{debug, info, instrument, trace};
 
 use crate::errors::*;
@@ -385,24 +385,6 @@ impl<'tcx> ThirBuildCx<'tcx> {
                         from_hir_call: true,
                         fn_span: expr.span,
                     }
-                } else if let ty::FnDef(def_id, _) = self.typeck_results.expr_ty(fun).kind()
-                    && let Some(intrinsic) = self.tcx.intrinsic(def_id)
-                    && intrinsic.name == sym::box_new
-                {
-                    // We don't actually evaluate `fun` here, so make sure that doesn't miss any side-effects.
-                    if !matches!(fun.kind, hir::ExprKind::Path(_)) {
-                        span_bug!(
-                            expr.span,
-                            "`box_new` intrinsic can only be called via path expression"
-                        );
-                    }
-                    let value = &args[0];
-                    return Expr {
-                        temp_lifetime: TempLifetime { temp_lifetime, backwards_incompatible },
-                        ty: expr_ty,
-                        span: expr.span,
-                        kind: ExprKind::Box { value: self.mirror_expr(value) },
-                    };
                 } else {
                     // Tuple-like ADTs are represented as ExprKind::Call. We convert them here.
                     let adt_data = if let hir::ExprKind::Path(ref qpath) = fun.kind

compiler/rustc_span/src/symbol.rs

@@ -1233,6 +1233,7 @@ symbols! {
         infer_static_outlives_requirements,
         inherent_associated_types,
         inherit,
+        init_box_via_move,
         initial,
         inlateout,
         inline,

library/alloc/src/alloc.rs

@@ -342,23 +342,6 @@ unsafe impl Allocator for Global {
     }
 }
 
-/// The allocator for `Box`.
-///
-/// # Safety
-///
-/// `size` and `align` must satisfy the conditions in [`Layout::from_size_align`].
-#[cfg(not(no_global_oom_handling))]
-#[lang = "exchange_malloc"]
-#[inline]
-#[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
-pub(crate) unsafe fn exchange_malloc(size: usize, align: usize) -> *mut u8 {
-    let layout = unsafe { Layout::from_size_align_unchecked(size, align) };
-    match Global.allocate(layout) {
-        Ok(ptr) => ptr.as_mut_ptr(),
-        Err(_) => handle_alloc_error(layout),
-    }
-}
-
 // # Allocation error handler
 
 #[cfg(not(no_global_oom_handling))]

library/alloc/src/boxed.rs

@@ -192,7 +192,7 @@ use core::fmt;
 use core::future::Future;
 use core::hash::{Hash, Hasher};
 use core::marker::{Tuple, Unsize};
-use core::mem::{self, SizedTypeProperties};
+use core::mem::{self, MaybeUninit, SizedTypeProperties};
 use core::ops::{
     AsyncFn, AsyncFnMut, AsyncFnOnce, CoerceUnsized, Coroutine, CoroutineState, Deref, DerefMut,
     DerefPure, DispatchFromDyn, LegacyReceiver,
@@ -203,7 +203,7 @@ use core::task::{Context, Poll};
 
 #[cfg(not(no_global_oom_handling))]
 use crate::alloc::handle_alloc_error;
-use crate::alloc::{AllocError, Allocator, Global, Layout, exchange_malloc};
+use crate::alloc::{AllocError, Allocator, Global, Layout};
 use crate::raw_vec::RawVec;
 #[cfg(not(no_global_oom_handling))]
 use crate::str::from_boxed_utf8_unchecked;
@@ -233,14 +233,39 @@ pub struct Box<
     #[unstable(feature = "allocator_api", issue = "32838")] A: Allocator = Global,
 >(Unique<T>, A);
 
-/// Constructs a `Box<T>` by calling the `exchange_malloc` lang item and moving the argument into
-/// the newly allocated memory. This is an intrinsic to avoid unnecessary copies.
+/// Monomorphic function for allocating an uninit `Box`.
 ///
-/// This is the surface syntax for `box <expr>` expressions.
-#[doc(hidden)]
+/// # Safety
+///
+/// size and align need to be safe for `Layout::from_size_align_unchecked`.
+#[inline]
+#[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
+unsafe fn box_new_uninit(size: usize, align: usize) -> *mut u8 {
+    let layout = unsafe { Layout::from_size_align_unchecked(size, align) };
+    match Global.allocate(layout) {
+        Ok(ptr) => ptr.as_mut_ptr(),
+        Err(_) => handle_alloc_error(layout),
+    }
+}
+
+/// Writes `x` into `b`, then returns `b` at its new type`.
+///
+/// This is needed for `vec!`, which can't afford any extra copies of the argument (or else debug
+/// builds regress), has to be written fully as a call chain without `let` (or else the temporary
+/// lifetimes of the arguments change), and can't use an `unsafe` block as that would then also
+/// include the user-provided `$x`.
 #[rustc_intrinsic]
 #[unstable(feature = "liballoc_internals", issue = "none")]
-pub fn box_new<T>(x: T) -> Box<T>;
+pub fn init_box_via_move<T>(b: Box<MaybeUninit<T>>, x: T) -> Box<T>;
+
+/// Helper for `vec!` to ensure type inferences work correctly (which it wouldn't if we
+/// inlined the `as` cast).
+#[doc(hidden)]
+#[unstable(feature = "liballoc_internals", issue = "none")]
+#[inline(always)]
+pub fn box_array_into_vec<T, const N: usize>(b: Box<[T; N]>) -> crate::vec::Vec<T> {
+    (b as Box<[T]>).into_vec()
+}
 
 impl<T> Box<T> {
     /// Allocates memory on the heap and then places `x` into it.
@@ -259,15 +284,18 @@ impl<T> Box<T> {
     #[rustc_diagnostic_item = "box_new"]
     #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
     pub fn new(x: T) -> Self {
-        // SAFETY: the size and align of a valid type `T` are always valid for `Layout`.
-        let ptr = unsafe {
-            exchange_malloc(<T as SizedTypeProperties>::SIZE, <T as SizedTypeProperties>::ALIGN)
-        } as *mut T;
-        // Nothing below can panic so we do not have to worry about deallocating `ptr`.
-        // SAFETY: we just allocated the box to store `x`.
-        unsafe { core::intrinsics::write_via_move(ptr, x) };
-        // SAFETY: we just initialized `b`.
-        unsafe { mem::transmute(ptr) }
+        // This is `Box::new_uninit` but inlined to avoid build time regressions.
+        // SAFETY:
+        // - The size and align of a valid type `T` are always valid for `Layout`.
+        // - If `allocate` succeeds, the returned pointer exactly matches what `Box` needs.
+        let b = unsafe {
+            mem::transmute(box_new_uninit(
+                <T as SizedTypeProperties>::SIZE,
+                <T as SizedTypeProperties>::ALIGN,
+            ))
+        };
+        // We could do this with `write_via_move`, but may as well use `init_box_via_move`.
+        init_box_via_move(b, x)
     }
 
     /// Constructs a new box with uninitialized contents.
@@ -285,9 +313,21 @@ impl<T> Box<T> {
     #[cfg(not(no_global_oom_handling))]
     #[stable(feature = "new_uninit", since = "1.82.0")]
     #[must_use]
-    #[inline]
+    #[inline(always)]
+    #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
     pub fn new_uninit() -> Box<mem::MaybeUninit<T>> {
-        Self::new_uninit_in(Global)
+        // This is the same as `Self::new_uninit_in(Global)`, but manually inlined (just like
+        // `Box::new`).
+
+        // SAFETY:
+        // - The size and align of a valid type `T` are always valid for `Layout`.
+        // - If `allocate` succeeds, the returned pointer exactly matches what `Box` needs.
+        unsafe {
+            mem::transmute(box_new_uninit(
+                <T as SizedTypeProperties>::SIZE,
+                <T as SizedTypeProperties>::ALIGN,
+            ))
+        }
     }
 
     /// Constructs a new `Box` with uninitialized contents, with the memory