Auto merge of #118538 - RalfJung:size-of-val-comments, r=WaffleLapkin

fix dynamic size/align computation logic for packed types with dyn trait tail

This logic was never updated to support `packed(N)` where `N > 1`, and it turns out to be wrong for that case.

Fixes rust-lang/rust#80925

`@bjorn3` I have not looked at cranelift; I assume it basically copied the size-of-val logic and hence could use much the same patch.

Author: bors

tests/pass/packed-struct-dyn-trait.rs

@@ -0,0 +1,21 @@
+// run-pass
+use std::ptr::addr_of;
+
+// When the unsized tail is a `dyn Trait`, its alignments is only dynamically known. This means the
+// packed(2) needs to be applied at runtime: the actual alignment of the field is `min(2,
+// usual_alignment)`. Here we check that we do this right by comparing size, alignment, and field
+// offset before and after unsizing.
+fn main() {
+    #[repr(C, packed(2))]
+    struct Packed<T: ?Sized>(u8, core::mem::ManuallyDrop<T>);
+
+    let p = Packed(0, core::mem::ManuallyDrop::new(1));
+    let p: &Packed<usize> = &p;
+    let sized = (core::mem::size_of_val(p), core::mem::align_of_val(p));
+    let sized_offset = unsafe { addr_of!(p.1).cast::<u8>().offset_from(addr_of!(p.0)) };
+    let p: &Packed<dyn Send> = p;
+    let un_sized = (core::mem::size_of_val(p), core::mem::align_of_val(p));
+    let un_sized_offset = unsafe { addr_of!(p.1).cast::<u8>().offset_from(addr_of!(p.0)) };
+    assert_eq!(sized, un_sized);
+    assert_eq!(sized_offset, un_sized_offset);
+}