update comments: inference seems more tricky than temporary lifetimes

Co-authored-by: Nadrieril <Nadrieril@users.noreply.github.com>

Author: RalfJung

library/alloc/src/intrinsics.rs

@@ -8,8 +8,8 @@ use crate::boxed::Box;
 /// Writes `x` into `b`.
 ///
 /// 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
+/// builds regress), has to be written fully as a call chain without `let` (or else this breaks inference
+/// of e.g. unsizing coercions), and can't use an `unsafe` block as that would then also
 /// include the user-provided `$x`.
 #[rustc_intrinsic]
 pub fn write_box_via_move<T>(b: Box<MaybeUninit<T>>, x: T) -> Box<MaybeUninit<T>>;

library/alloc/src/macros.rs

@@ -50,7 +50,7 @@ macro_rules! vec {
         // Using `write_box_via_move` produces a dramatic improvement in stack usage for unoptimized
         // programs using this code path to construct large Vecs. We can't use `write_via_move`
         // because this entire invocation has to remain a call chain without `let` bindings, or else
-        // the temporary scopes change and things break (see `vec-macro-rvalue-scope` test).
+        // inference changes and things break (see `vec-macro-rvalue-scope` and `autoderef-vec-box-fn-36786` tests).
         // This function isn't actually safe but the way we use it here is. We can't use an
         // unsafe block as that would also wrap `$x`.
         $crate::boxed::box_uninit_array_into_vec_unsafe(