Implement Unsized Rvalues

src/doc/unstable-book/src/language-features/unsized-locals.md

@@ -0,0 +1,180 @@
+# `unsized_locals`
+
+The tracking issue for this feature is: [#48055]
+
+[#48055]: /~https://github.com/rust-lang/rust/issues/48055
+
+------------------------
+
+This implements [RFC1909]. When turned on, you can have unsized arguments and locals:
+
+[RFC1909]: /~https://github.com/rust-lang/rfcs/blob/master/text/1909-coercions.md
+
+```rust
+#![feature(unsized_locals)]
+
+use std::any::Any;
+
+fn main() {
+    let x: Box<dyn Any> = Box::new(42);
+    let x: dyn Any = *x;
+    //  ^ unsized local variable
+    //               ^^ unsized temporary
+    foo(x);
+}
+
+fn foo(_: dyn Any) {}
+//     ^^^^^^ unsized argument
+```
+
+The RFC still forbids the following unsized expressions:
+
+```rust,ignore
+#![feature(unsized_locals)]
+
+use std::any::Any;
+
+struct MyStruct<T: ?Sized> {
+    content: T,
+}
+
+struct MyTupleStruct<T: ?Sized>(T);
+
+fn answer() -> Box<dyn Any> {
+    Box::new(42)
+}
+
+fn main() {
+    // You CANNOT have unsized statics.
+    static X: dyn Any = *answer();  // ERROR
+    const Y: dyn Any = *answer();  // ERROR
+
+    // You CANNOT have struct initialized unsized.
+    MyStruct { content: *answer() };  // ERROR
+    MyTupleStruct(*answer());  // ERROR
+    (42, *answer());  // ERROR
+
+    // You CANNOT have unsized return types.
+    fn my_function() -> dyn Any { *answer() }  // ERROR
+
+    // You CAN have unsized local variables...
+    let mut x: dyn Any = *answer();  // OK
+    // ...but you CANNOT reassign to them.
+    x = *answer();  // ERROR
+
+    // You CANNOT even initialize them separately.
+    let y: dyn Any;  // OK
+    y = *answer();  // ERROR
+
+    // Not mentioned in the RFC, but by-move captured variables are also Sized.
+    let x: dyn Any = *answer();
+    (move || {  // ERROR
+        let y = x;
+    })();
+
+    // You CAN create a closure with unsized arguments,
+    // but you CANNOT call it.
+    // This is an implementation detail and may be changed in the future.
+    let f = |x: dyn Any| {};
+    f(*answer());  // ERROR
+}
+```
+
+However, the current implementation allows `MyTupleStruct(..)` to be unsized. This will be fixed in the future.
+
+## By-value trait objects
+
+With this feature, you can have by-value `self` arguments without `Self: Sized` bounds.
+
+```rust
+#![feature(unsized_locals)]
+
+trait Foo {
+    fn foo(self) {}
+}
+
+impl<T: ?Sized> Foo for T {}
+
+fn main() {
+    let slice: Box<[i32]> = Box::new([1, 2, 3]);
+    <[i32] as Foo>::foo(*slice);
+}
+```
+
+And `Foo` will also be object-safe. However, this object-safety is not yet implemented.
+
+```rust,ignore
+#![feature(unsized_locals)]
+
+trait Foo {
+    fn foo(self) {}
+}
+
+impl<T: ?Sized> Foo for T {}
+
+fn main () {
+    let slice: Box<dyn Foo> = Box::new([1, 2, 3]);
+    // doesn't compile yet
+    <dyn Foo as Foo>::foo(*slice);
+}
+```
+
+Unfortunately, this is not implemented yet.
+
+One of the objectives of this feature is to allow `Box<dyn FnOnce>`, instead of `Box<dyn FnBox>` in the future. See [#28796] for details.
+
+[#28796]: /~https://github.com/rust-lang/rust/issues/28796
+
+## Variable length arrays
+
+The RFC also describes an extension to the array literal syntax: `[e; dyn n]`. In the syntax, `n` isn't necessarily a constant expression. The array is dynamically allocated on the stack and has the type of `[T]`, instead of `[T; n]`.
+
+```rust,ignore
+#![feature(unsized_locals)]
+
+fn mergesort<T: Ord>(a: &mut [T]) {
+    let mut tmp = [T; dyn a.len()];
+    // ...
+}
+
+fn main() {
+    let mut a = [3, 1, 5, 6];
+    mergesort(&mut a);
+    assert_eq!(a, [1, 3, 5, 6]);
+}
+```
+
+VLAs are not implemented yet. The syntax isn't final, either. We may need an alternative syntax for Rust 2015 because, in Rust 2015, expressions like `[e; dyn(1)]` would be ambiguous. One possible alternative proposed in the RFC is `[e; n]`: if `n` captures one or more local variables, then it is considered as `[e; dyn n]`.
+
+## Advisory on stack usage
+
+It's advised not to casually use the `#![feature(unsized_locals)]` feature. Typical use-cases are:
+
+- When you need a by-value trait objects.
+- When you really need a fast allocation of small temporary arrays.
+
+Another pitfall is repetitive allocation and temporaries. Currently the compiler simply extends the stack frame every time it encounters an unsized assignment. So for example, the code
+
+```rust
+#![feature(unsized_locals)]
+
+fn main() {
+    let x: Box<[i32]> = Box::new([1, 2, 3, 4, 5]);
+    let _x = {{{{{{{{{{*x}}}}}}}}}};
+}
+```
+
+and the code
+
+```rust
+#![feature(unsized_locals)]
+
+fn main() {
+    for _ in 0..10 {
+        let x: Box<[i32]> = Box::new([1, 2, 3, 4, 5]);
+        let _x = *x;
+    }
+}
+```
+
+will unnecessarily extend the stack frame.

src/librustc/traits/error_reporting.rs

@@ -1454,6 +1454,15 @@ impl<'a, 'gcx, 'tcx> InferCtxt<'a, 'gcx, 'tcx> {
             }
             ObligationCauseCode::VariableType(_) => {
                 err.note("all local variables must have a statically known size");
+                if !self.tcx.features().unsized_locals {
+                    err.help("unsized locals are gated as an unstable feature");
+                }
+            }
+            ObligationCauseCode::SizedArgumentType => {
+                err.note("all function arguments must have a statically known size");
+                if !self.tcx.features().unsized_locals {
+                    err.help("unsized locals are gated as an unstable feature");
+                }
             }
             ObligationCauseCode::SizedReturnType => {
                 err.note("the return type of a function must have a \

src/librustc/traits/mod.rs

@@ -185,6 +185,8 @@ pub enum ObligationCauseCode<'tcx> {
     StructInitializerSized,
     /// Type of each variable must be Sized
     VariableType(ast::NodeId),
+    /// Argument type must be Sized
+    SizedArgumentType,
     /// Return type must be Sized
     SizedReturnType,
     /// Yield type must be Sized

src/librustc/traits/structural_impls.rs

@@ -203,6 +203,7 @@ impl<'a, 'tcx> Lift<'tcx> for traits::ObligationCauseCode<'a> {
             super::StructInitializerSized => Some(super::StructInitializerSized),
             super::VariableType(id) => Some(super::VariableType(id)),
             super::ReturnType(id) => Some(super::ReturnType(id)),
+            super::SizedArgumentType => Some(super::SizedArgumentType),
             super::SizedReturnType => Some(super::SizedReturnType),
             super::SizedYieldType => Some(super::SizedYieldType),
             super::RepeatVec => Some(super::RepeatVec),

src/librustc_codegen_llvm/abi.rs

@@ -187,8 +187,10 @@ impl ArgTypeExt<'ll, 'tcx> for ArgType<'tcx, Ty<'tcx>> {
             return;
         }
         let cx = bx.cx;
-        if self.is_indirect() {
-            OperandValue::Ref(val, self.layout.align).store(bx, dst)
+        if self.is_sized_indirect() {
+            OperandValue::Ref(val, None, self.layout.align).store(bx, dst)
+        } else if self.is_unsized_indirect() {
+            bug!("unsized ArgType must be handled through store_fn_arg");
         } else if let PassMode::Cast(cast) = self.mode {
             // FIXME(eddyb): Figure out when the simpler Store is safe, clang
             // uses it for i16 -> {i8, i8}, but not for i24 -> {i8, i8, i8}.
@@ -246,7 +248,10 @@ impl ArgTypeExt<'ll, 'tcx> for ArgType<'tcx, Ty<'tcx>> {
             PassMode::Pair(..) => {
                 OperandValue::Pair(next(), next()).store(bx, dst);
             }
-            PassMode::Direct(_) | PassMode::Indirect(_) | PassMode::Cast(_) => {
+            PassMode::Indirect(_, Some(_)) => {
+                OperandValue::Ref(next(), Some(next()), self.layout.align).store(bx, dst);
+            }
+            PassMode::Direct(_) | PassMode::Indirect(_, None) | PassMode::Cast(_) => {
                 self.store(bx, next(), dst);
             }
         }
@@ -302,6 +307,10 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
             // Don't pass the vtable, it's not an argument of the virtual fn.
             // Instead, pass just the (thin pointer) first field of `*dyn Trait`.
             if arg_idx == Some(0) {
+                if layout.is_unsized() {
+                    unimplemented!("by-value trait object is not \
+                                    yet implemented in #![feature(unsized_locals)]");
+                }
                 // FIXME(eddyb) `layout.field(cx, 0)` is not enough because e.g.
                 // `Box<dyn Trait>` has a few newtype wrappers around the raw
                 // pointer, so we'd have to "dig down" to find `*dyn Trait`.
@@ -538,7 +547,7 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
                 }
 
                 let size = arg.layout.size;
-                if size > layout::Pointer.size(cx) {
+                if arg.layout.is_unsized() || size > layout::Pointer.size(cx) {
                     arg.make_indirect();
                 } else {
                     // We want to pass small aggregates as immediates, but using
@@ -554,7 +563,7 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
             for arg in &mut self.args {
                 fixup(arg);
             }
-            if let PassMode::Indirect(ref mut attrs) = self.ret.mode {
+            if let PassMode::Indirect(ref mut attrs, _) = self.ret.mode {
                 attrs.set(ArgAttribute::StructRet);
             }
             return;
@@ -571,7 +580,7 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
             if let PassMode::Pair(_, _) = arg.mode { 2 } else { 1 }
         ).sum();
         let mut llargument_tys = Vec::with_capacity(
-            if let PassMode::Indirect(_) = self.ret.mode { 1 } else { 0 } + args_capacity
+            if let PassMode::Indirect(..) = self.ret.mode { 1 } else { 0 } + args_capacity
         );
 
         let llreturn_ty = match self.ret.mode {
@@ -580,7 +589,7 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
                 self.ret.layout.immediate_llvm_type(cx)
             }
             PassMode::Cast(cast) => cast.llvm_type(cx),
-            PassMode::Indirect(_) => {
+            PassMode::Indirect(..) => {
                 llargument_tys.push(self.ret.memory_ty(cx).ptr_to());
                 Type::void(cx)
             }
@@ -600,8 +609,15 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
                     llargument_tys.push(arg.layout.scalar_pair_element_llvm_type(cx, 1, true));
                     continue;
                 }
+                PassMode::Indirect(_, Some(_)) => {
+                    let ptr_ty = cx.tcx.mk_mut_ptr(arg.layout.ty);
+                    let ptr_layout = cx.layout_of(ptr_ty);
+                    llargument_tys.push(ptr_layout.scalar_pair_element_llvm_type(cx, 0, true));
+                    llargument_tys.push(ptr_layout.scalar_pair_element_llvm_type(cx, 1, true));
+                    continue;
+                }
                 PassMode::Cast(cast) => cast.llvm_type(cx),
-                PassMode::Indirect(_) => arg.memory_ty(cx).ptr_to(),
+                PassMode::Indirect(_, None) => arg.memory_ty(cx).ptr_to(),
             };
             llargument_tys.push(llarg_ty);
         }
@@ -640,7 +656,7 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
             PassMode::Direct(ref attrs) => {
                 attrs.apply_llfn(llvm::AttributePlace::ReturnValue, llfn);
             }
-            PassMode::Indirect(ref attrs) => apply(attrs),
+            PassMode::Indirect(ref attrs, _) => apply(attrs),
             _ => {}
         }
         for arg in &self.args {
@@ -650,7 +666,11 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
             match arg.mode {
                 PassMode::Ignore => {}
                 PassMode::Direct(ref attrs) |
-                PassMode::Indirect(ref attrs) => apply(attrs),
+                PassMode::Indirect(ref attrs, None) => apply(attrs),
+                PassMode::Indirect(ref attrs, Some(ref extra_attrs)) => {
+                    apply(attrs);
+                    apply(extra_attrs);
+                }
                 PassMode::Pair(ref a, ref b) => {
                     apply(a);
                     apply(b);
@@ -670,7 +690,7 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
             PassMode::Direct(ref attrs) => {
                 attrs.apply_callsite(llvm::AttributePlace::ReturnValue, callsite);
             }
-            PassMode::Indirect(ref attrs) => apply(attrs),
+            PassMode::Indirect(ref attrs, _) => apply(attrs),
             _ => {}
         }
         if let layout::Abi::Scalar(ref scalar) = self.ret.layout.abi {
@@ -694,7 +714,11 @@ impl<'tcx> FnTypeExt<'tcx> for FnType<'tcx, Ty<'tcx>> {
             match arg.mode {
                 PassMode::Ignore => {}
                 PassMode::Direct(ref attrs) |
-                PassMode::Indirect(ref attrs) => apply(attrs),
+                PassMode::Indirect(ref attrs, None) => apply(attrs),
+                PassMode::Indirect(ref attrs, Some(ref extra_attrs)) => {
+                    apply(attrs);
+                    apply(extra_attrs);
+                }
                 PassMode::Pair(ref a, ref b) => {
                     apply(a);
                     apply(b);

src/librustc_codegen_llvm/builder.rs

@@ -445,6 +445,25 @@ impl Builder<'a, 'll, 'tcx> {
         }
     }
 
+    pub fn array_alloca(&self,
+                        ty: &'ll Type,
+                        len: &'ll Value,
+                        name: &str,
+                        align: Align) -> &'ll Value {
+        self.count_insn("alloca");
+        unsafe {
+            let alloca = if name.is_empty() {
+                llvm::LLVMBuildArrayAlloca(self.llbuilder, ty, len, noname())
+            } else {
+                let name = SmallCStr::new(name);
+                llvm::LLVMBuildArrayAlloca(self.llbuilder, ty, len,
+                                           name.as_ptr())
+            };
+            llvm::LLVMSetAlignment(alloca, align.abi() as c_uint);
+            alloca
+        }
+    }
+
     pub fn load(&self, ptr: &'ll Value, align: Align) -> &'ll Value {
         self.count_insn("load");
         unsafe {

src/librustc_codegen_llvm/intrinsic.rs

@@ -605,7 +605,7 @@ pub fn codegen_intrinsic_call(
                         // etc.
                         assert!(!bx.cx.type_needs_drop(arg.layout.ty));
                         let (ptr, align) = match arg.val {
-                            OperandValue::Ref(ptr, align) => (ptr, align),
+                            OperandValue::Ref(ptr, None, align) => (ptr, align),
                             _ => bug!()
                         };
                         let arg = PlaceRef::new_sized(ptr, arg.layout, align);

src/librustc_codegen_llvm/llvm/ffi.rs

@@ -875,6 +875,11 @@ extern "C" {
 
     // Memory
     pub fn LLVMBuildAlloca(B: &Builder<'a>, Ty: &'a Type, Name: *const c_char) -> &'a Value;
+    pub fn LLVMBuildArrayAlloca(B: &Builder<'a>,
+                                Ty: &'a Type,
+                                Val: &'a Value,
+                                Name: *const c_char)
+                                -> &'a Value;
     pub fn LLVMBuildLoad(B: &Builder<'a>, PointerVal: &'a Value, Name: *const c_char) -> &'a Value;
 
     pub fn LLVMBuildStore(B: &Builder<'a>, Val: &'a Value, Ptr: &'a Value) -> &'a Value;