Fix -Z print-type-sizes for generators with discriminant field ordered first

compiler/rustc_ty_utils/src/layout.rs

@@ -919,7 +919,7 @@ fn variant_info_for_generator<'tcx>(
     def_id: DefId,
     substs: ty::SubstsRef<'tcx>,
 ) -> (Vec<VariantInfo>, Option<Size>) {
-    let Variants::Multiple { tag, ref tag_encoding, .. } = layout.variants else {
+    let Variants::Multiple { tag, ref tag_encoding, tag_field, .. } = layout.variants else {
         return (vec![], None);
     };
 
@@ -975,12 +975,28 @@ fn variant_info_for_generator<'tcx>(
             if variant_size == Size::ZERO {
                 variant_size = upvars_size;
             }
-            // We need to add the discriminant size back into min_size, since it is subtracted
-            // later during printing.
-            variant_size += match tag_encoding {
-                TagEncoding::Direct => tag.size(cx),
-                _ => Size::ZERO,
-            };
+
+            // This `if` deserves some explanation.
+            //
+            // The layout code has a choice of where to place the discriminant of this generator.
+            // If the discriminant of the generator is placed early in the layout (before the
+            // variant's own fields), then it'll implicitly be counted towards the size of the
+            // variant, since we use the maximum offset to calculate size.
+            //    (side-note: I know this is a bit problematic given upvars placement, etc).
+            //
+            // This is important, since the layout printing code always subtracts this discriminant
+            // size from the variant size if the struct is "enum"-like, so failing to account for it
+            // will either lead to numerical underflow, or an underreported variant size...
+            //
+            // However, if the discriminant is placed past the end of the variant, then we need
+            // to factor in the size of the discriminant manually. This really should be refactored
+            // better, but this "works" for now.
+            if layout.fields.offset(tag_field) >= variant_size {
+                variant_size += match tag_encoding {
+                    TagEncoding::Direct => tag.size(cx),
+                    _ => Size::ZERO,
+                };
+            }
 
             VariantInfo {
                 name: Some(Symbol::intern(&ty::GeneratorSubsts::variant_name(variant_idx))),

src/test/ui/print_type_sizes/async.rs

@@ -1,19 +1,11 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type lib
 // edition:2021
 // build-pass
 // ignore-pass
 
-#![feature(start)]
-
 async fn wait() {}
 
-async fn test(arg: [u8; 8192]) {
+pub async fn test(arg: [u8; 8192]) {
     wait().await;
     drop(arg);
 }
-
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
-    let _ = test([0; 8192]);
-    0
-}

src/test/ui/print_type_sizes/async.stdout

@@ -1,4 +1,4 @@
-print-type-size type: `[async fn body@$DIR/async.rs:10:32: 13:2]`: 16386 bytes, alignment: 1 bytes
+print-type-size type: `[async fn body@$DIR/async.rs:8:36: 11:2]`: 16386 bytes, alignment: 1 bytes
 print-type-size     discriminant: 1 bytes
 print-type-size     variant `Suspend0`: 16385 bytes
 print-type-size         field `.arg`: 8192 bytes, offset: 0 bytes, alignment: 1 bytes
@@ -16,14 +16,14 @@ print-type-size type: `std::mem::MaybeUninit<[u8; 8192]>`: 8192 bytes, alignment
 print-type-size     variant `MaybeUninit`: 8192 bytes
 print-type-size         field `.uninit`: 0 bytes
 print-type-size         field `.value`: 8192 bytes
-print-type-size type: `[async fn body@$DIR/async.rs:8:17: 8:19]`: 1 bytes, alignment: 1 bytes
+print-type-size type: `[async fn body@$DIR/async.rs:6:17: 6:19]`: 1 bytes, alignment: 1 bytes
 print-type-size     discriminant: 1 bytes
 print-type-size     variant `Unresumed`: 0 bytes
 print-type-size     variant `Returned`: 0 bytes
 print-type-size     variant `Panicked`: 0 bytes
-print-type-size type: `std::mem::ManuallyDrop<[async fn body@$DIR/async.rs:8:17: 8:19]>`: 1 bytes, alignment: 1 bytes
+print-type-size type: `std::mem::ManuallyDrop<[async fn body@$DIR/async.rs:6:17: 6:19]>`: 1 bytes, alignment: 1 bytes
 print-type-size     field `.value`: 1 bytes
-print-type-size type: `std::mem::MaybeUninit<[async fn body@$DIR/async.rs:8:17: 8:19]>`: 1 bytes, alignment: 1 bytes
+print-type-size type: `std::mem::MaybeUninit<[async fn body@$DIR/async.rs:6:17: 6:19]>`: 1 bytes, alignment: 1 bytes
 print-type-size     variant `MaybeUninit`: 1 bytes
 print-type-size         field `.uninit`: 0 bytes
 print-type-size         field `.value`: 1 bytes

src/test/ui/print_type_sizes/generator.rs

@@ -1,8 +1,8 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type=lib
 // build-pass
 // ignore-pass
 
-#![feature(start, generators, generator_trait)]
+#![feature(generators, generator_trait)]
 
 use std::ops::Generator;
 
@@ -13,8 +13,6 @@ fn generator<const C: usize>(array: [u8; C]) -> impl Generator<Yield = (), Retur
     }
 }
 
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
+pub fn foo() {
     let _ = generator([0; 8192]);
-    0
 }

src/test/ui/print_type_sizes/generator_discr_placement.rs

@@ -0,0 +1,23 @@
+// compile-flags: -Z print-type-sizes --crate-type lib
+// build-pass
+// ignore-pass
+
+// Tests a generator that has its discriminant as the *final* field.
+
+// Avoid emitting panic handlers, like the rest of these tests...
+#![feature(generators)]
+
+pub fn foo() {
+    let a = || {
+        {
+            let w: i32 = 4;
+            yield;
+            drop(w);
+        }
+        {
+            let z: i32 = 7;
+            yield;
+            drop(z);
+        }
+    };
+}

src/test/ui/print_type_sizes/generator_discr_placement.stdout

@@ -0,0 +1,11 @@
+print-type-size type: `[generator@$DIR/generator_discr_placement.rs:11:13: 11:15]`: 8 bytes, alignment: 4 bytes
+print-type-size     discriminant: 1 bytes
+print-type-size     variant `Suspend0`: 7 bytes
+print-type-size         padding: 3 bytes
+print-type-size         field `.w`: 4 bytes, alignment: 4 bytes
+print-type-size     variant `Suspend1`: 7 bytes
+print-type-size         padding: 3 bytes
+print-type-size         field `.z`: 4 bytes, alignment: 4 bytes
+print-type-size     variant `Unresumed`: 0 bytes
+print-type-size     variant `Returned`: 0 bytes
+print-type-size     variant `Panicked`: 0 bytes

src/test/ui/print_type_sizes/generics.rs

@@ -1,4 +1,4 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type=lib
 // build-pass
 // ignore-pass
 // ^-- needed because `--pass check` does not emit the output needed.
@@ -8,24 +8,6 @@
 // monomorphized, in the MIR of the original function in which they
 // occur, to have their size reported.
 
-#![feature(start)]
-
-// In an ad-hoc attempt to avoid the injection of unwinding code
-// (which clutters the output of `-Z print-type-sizes` with types from
-// `unwind::libunwind`):
-//
-//   * I am not using Default to build values because that seems to
-//     cause the injection of unwinding code. (Instead I just make `fn new`
-//     methods.)
-//
-//   * Pair derive Copy to ensure that we don't inject
-//     unwinding code into generic uses of Pair when T itself is also
-//     Copy.
-//
-//     (I suspect this reflect some naivety within the rust compiler
-//      itself; it should be checking for drop glue, i.e., a destructor
-//      somewhere in the monomorphized types. It should not matter whether
-//      the type is Copy.)
 #[derive(Copy, Clone)]
 pub struct Pair<T> {
     _car: T,
@@ -61,11 +43,9 @@ pub fn f1<T:Copy>(x: T) {
         Pair::new(FiftyBytes::new(), FiftyBytes::new());
 }
 
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
+pub fn start() {
     let _b: Pair<u8> = Pair::new(0, 0);
     let _s: Pair<SevenBytes> = Pair::new(SevenBytes::new(), SevenBytes::new());
     let ref _z: ZeroSized = ZeroSized;
     f1::<SevenBytes>(SevenBytes::new());
-    0
 }

src/test/ui/print_type_sizes/multiple_types.rs

@@ -1,23 +1,13 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type=lib
 // build-pass
 
 // This file illustrates that when multiple structural types occur in
 // a function, every one of them is included in the output.
 
-#![feature(start)]
-
 pub struct SevenBytes([u8;  7]);
 pub struct FiftyBytes([u8; 50]);
 
 pub enum Enum {
     Small(SevenBytes),
     Large(FiftyBytes),
 }
-
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
-    let _e: Enum;
-    let _f: FiftyBytes;
-    let _s: SevenBytes;
-    0
-}

src/test/ui/print_type_sizes/niche-filling.rs

@@ -1,4 +1,4 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type=lib
 // build-pass
 // ignore-pass
 // ^-- needed because `--pass check` does not emit the output needed.
@@ -14,7 +14,6 @@
 // aligned (while on most it is 8-byte aligned) and so the resulting
 // padding and overall computed sizes can be quite different.
 
-#![feature(start)]
 #![feature(rustc_attrs)]
 #![allow(dead_code)]
 
@@ -56,7 +55,7 @@ pub struct NestedNonZero {
 
 impl Default for NestedNonZero {
     fn default() -> Self {
-        NestedNonZero { pre: 0, val: NonZeroU32::new(1).unwrap(), post: 0 }
+        NestedNonZero { pre: 0, val: unsafe { NonZeroU32::new_unchecked(1) }, post: 0 }
     }
 }
 
@@ -76,8 +75,7 @@ pub union Union2<A: Copy, B: Copy> {
     b: B,
 }
 
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
+pub fn test() {
     let _x: MyOption<NonZeroU32> = Default::default();
     let _y: EmbeddedDiscr = Default::default();
     let _z: MyOption<IndirectNonZero> = Default::default();
@@ -96,6 +94,4 @@ fn start(_: isize, _: *const *const u8) -> isize {
     // ...even when theoretically possible.
     let _j: MyOption<Union1<NonZeroU32>> = Default::default();
     let _k: MyOption<Union2<NonZeroU32, NonZeroU32>> = Default::default();
-
-    0
 }

src/test/ui/print_type_sizes/no_duplicates.rs

@@ -1,4 +1,4 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type=lib
 // build-pass
 // ignore-pass
 // ^-- needed because `--pass check` does not emit the output needed.
@@ -8,16 +8,12 @@
 // (even if multiple functions), it is only printed once in the
 // print-type-sizes output.
 
-#![feature(start)]
-
 pub struct SevenBytes([u8; 7]);
 
 pub fn f1() {
     let _s: SevenBytes = SevenBytes([0; 7]);
 }
 
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
+pub fn test() {
     let _s: SevenBytes = SevenBytes([0; 7]);
-    0
 }

src/test/ui/print_type_sizes/packed.rs

@@ -1,4 +1,4 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type=lib
 // build-pass
 // ignore-pass
 // ^-- needed because `--pass check` does not emit the output needed.
@@ -13,11 +13,10 @@
 // padding and overall computed sizes can be quite different.
 
 #![allow(dead_code)]
-#![feature(start)]
 
 #[derive(Default)]
 #[repr(packed)]
-struct Packed1 {
+pub struct Packed1 {
     a: u8,
     b: u8,
     g: i32,
@@ -28,7 +27,7 @@ struct Packed1 {
 
 #[derive(Default)]
 #[repr(packed(2))]
-struct Packed2 {
+pub struct Packed2 {
     a: u8,
     b: u8,
     g: i32,
@@ -40,7 +39,7 @@ struct Packed2 {
 #[derive(Default)]
 #[repr(packed(2))]
 #[repr(C)]
-struct Packed2C {
+pub struct Packed2C {
     a: u8,
     b: u8,
     g: i32,
@@ -50,20 +49,11 @@ struct Packed2C {
 }
 
 #[derive(Default)]
-struct Padded {
+pub struct Padded {
     a: u8,
     b: u8,
     g: i32,
     c: u8,
     h: i16,
     d: u8,
 }
-
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
-    let _c: Packed1 = Default::default();
-    let _d: Packed2 = Default::default();
-    let _e: Packed2C = Default::default();
-    let _f: Padded = Default::default();
-    0
-}

src/test/ui/print_type_sizes/padding.rs

@@ -1,4 +1,4 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type=lib
 // build-pass
 
 // This file illustrates how padding is handled: alignment
@@ -9,7 +9,6 @@
 // aligned (while on most it is 8-byte aligned) and so the resulting
 // padding and overall computed sizes can be quite different.
 
-#![feature(start)]
 #![allow(dead_code)]
 
 struct S {
@@ -27,8 +26,3 @@ enum E2 {
     A(i8, i32),
     B(S),
 }
-
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
-    0
-}

src/test/ui/print_type_sizes/repr-align.rs

@@ -1,4 +1,4 @@
-// compile-flags: -Z print-type-sizes
+// compile-flags: -Z print-type-sizes --crate-type=lib
 // build-pass
 // ignore-pass
 // ^-- needed because `--pass check` does not emit the output needed.
@@ -11,7 +11,7 @@
 // It avoids using u64/i64 because on some targets that is only 4-byte
 // aligned (while on most it is 8-byte aligned) and so the resulting
 // padding and overall computed sizes can be quite different.
-#![feature(start)]
+
 #![allow(dead_code)]
 
 #[repr(align(16))]
@@ -24,15 +24,9 @@ enum E {
 }
 
 #[derive(Default)]
-struct S {
+pub struct S {
     a: i32,
     b: i32,
     c: A,
     d: i8,
 }
-
-#[start]
-fn start(_: isize, _: *const *const u8) -> isize {
-    let _s: S = Default::default();
-    0
-}