Ban non-array SIMD

Author: scottmcm

compiler/rustc_error_codes/src/error_codes/E0074.md

@@ -11,7 +11,7 @@ This will cause an error:
 #![feature(repr_simd)]
 
 #[repr(simd)]
-struct Bad<T>(T, T, T, T);
+struct Bad<T>([T; 4]);
 ```
 
 This will not:
@@ -20,5 +20,5 @@ This will not:
 #![feature(repr_simd)]
 
 #[repr(simd)]
-struct Good(u32, u32, u32, u32);
+struct Good([u32; 4]);
 ```

compiler/rustc_error_codes/src/error_codes/E0075.md

@@ -1,6 +1,6 @@
-A `#[simd]` attribute was applied to an empty tuple struct.
+A `#[simd]` attribute was applied to an empty or multi-field struct.
 
-Erroneous code example:
+Erroneous code examples:
 
 ```compile_fail,E0075
 #![feature(repr_simd)]
@@ -9,15 +9,21 @@ Erroneous code example:
 struct Bad; // error!
 ```
 
-The `#[simd]` attribute can only be applied to non empty tuple structs, because
-it doesn't make sense to try to use SIMD operations when there are no values to
-operate on.
+```compile_fail,E0075
+#![feature(repr_simd)]
+
+#[repr(simd)]
+struct Bad([u32; 1], [u32; 1]); // error!
+```
+
+The `#[simd]` attribute can only be applied to a single-field struct, because
+the one field must be the array of values in the vector.
 
 Fixed example:
 
 ```
 #![feature(repr_simd)]
 
 #[repr(simd)]
-struct Good(u32); // ok!
+struct Good([u32; 2]); // ok!
 ```

compiler/rustc_error_codes/src/error_codes/E0076.md

@@ -1,4 +1,4 @@
-All types in a tuple struct aren't the same when using the `#[simd]`
+The type of the field in a tuple struct isn't an array when using the `#[simd]`
 attribute.
 
 Erroneous code example:
@@ -7,18 +7,18 @@ Erroneous code example:
 #![feature(repr_simd)]
 
 #[repr(simd)]
-struct Bad(u16, u32, u32 u32); // error!
+struct Bad(u16); // error!
 ```
 
 When using the `#[simd]` attribute to automatically use SIMD operations in tuple
-struct, the types in the struct must all be of the same type, or the compiler
-will trigger this error.
+structs, if you want a single-lane vector then the field must be a 1-element
+array, or the compiler will trigger this error.
 
 Fixed example:
 
 ```
 #![feature(repr_simd)]
 
 #[repr(simd)]
-struct Good(u32, u32, u32, u32); // ok!
+struct Good([u16; 1]); // ok!
 ```

compiler/rustc_error_codes/src/error_codes/E0077.md

@@ -7,7 +7,7 @@ Erroneous code example:
 #![feature(repr_simd)]
 
 #[repr(simd)]
-struct Bad(String); // error!
+struct Bad([String; 2]); // error!
 ```
 
 When using the `#[simd]` attribute on a tuple struct, the elements in the tuple
@@ -19,5 +19,5 @@ Fixed example:
 #![feature(repr_simd)]
 
 #[repr(simd)]
-struct Good(u32, u32, u32, u32); // ok!
+struct Good([u32; 4]); // ok!
 ```

compiler/rustc_error_codes/src/error_codes/E0511.md

@@ -23,11 +23,11 @@ The generic type has to be a SIMD type. Example:
 
 #[repr(simd)]
 #[derive(Copy, Clone)]
-struct i32x2(i32, i32);
+struct i32x2([i32; 2]);
 
 extern "rust-intrinsic" {
     fn simd_add<T>(a: T, b: T) -> T;
 }
 
-unsafe { simd_add(i32x2(0, 0), i32x2(1, 2)); } // ok!
+unsafe { simd_add(i32x2([0, 0]), i32x2([1, 2])); } // ok!
 ```

compiler/rustc_hir_analysis/src/check/check.rs

@@ -1064,20 +1064,29 @@ fn check_simd(tcx: TyCtxt<'_>, sp: Span, def_id: LocalDefId) {
             struct_span_code_err!(tcx.dcx(), sp, E0075, "SIMD vector cannot be empty").emit();
             return;
         }
-        let e = fields[FieldIdx::ZERO].ty(tcx, args);
-        if !fields.iter().all(|f| f.ty(tcx, args) == e) {
-            struct_span_code_err!(tcx.dcx(), sp, E0076, "SIMD vector should be homogeneous")
-                .with_span_label(sp, "SIMD elements must have the same type")
+
+        let array_field = &fields[FieldIdx::ZERO];
+        let array_ty = array_field.ty(tcx, args);
+        let ty::Array(element_ty, len_const) = array_ty.kind() else {
+            struct_span_code_err!(
+                tcx.dcx(),
+                sp,
+                E0076,
+                "SIMD vector's only field must be an array"
+            )
+            .with_span_label(tcx.def_span(array_field.did), "not an array")
+            .emit();
+            return;
+        };
+
+        if let Some(second_field) = fields.get(FieldIdx::from_u32(1)) {
+            struct_span_code_err!(tcx.dcx(), sp, E0075, "SIMD vector cannot have multiple fields")
+                .with_span_label(tcx.def_span(second_field.did), "excess field")
                 .emit();
             return;
         }
 
-        let len = if let ty::Array(_ty, c) = e.kind() {
-            c.try_eval_target_usize(tcx, tcx.param_env(def.did()))
-        } else {
-            Some(fields.len() as u64)
-        };
-        if let Some(len) = len {
+        if let Some(len) = len_const.try_eval_target_usize(tcx, tcx.param_env(def.did())) {
             if len == 0 {
                 struct_span_code_err!(tcx.dcx(), sp, E0075, "SIMD vector cannot be empty").emit();
                 return;
@@ -1097,16 +1106,9 @@ fn check_simd(tcx: TyCtxt<'_>, sp: Span, def_id: LocalDefId) {
         // These are scalar types which directly match a "machine" type
         // Yes: Integers, floats, "thin" pointers
         // No: char, "fat" pointers, compound types
-        match e.kind() {
-            ty::Param(_) => (), // pass struct<T>(T, T, T, T) through, let monomorphization catch errors
-            ty::Int(_) | ty::Uint(_) | ty::Float(_) | ty::RawPtr(_, _) => (), // struct(u8, u8, u8, u8) is ok
-            ty::Array(t, _) if matches!(t.kind(), ty::Param(_)) => (), // pass struct<T>([T; N]) through, let monomorphization catch errors
-            ty::Array(t, _clen)
-                if matches!(
-                    t.kind(),
-                    ty::Int(_) | ty::Uint(_) | ty::Float(_) | ty::RawPtr(_, _)
-                ) =>
-            { /* struct([f32; 4]) is ok */ }
+        match element_ty.kind() {
+            ty::Param(_) => (), // pass struct<T>([T; 4]) through, let monomorphization catch errors
+            ty::Int(_) | ty::Uint(_) | ty::Float(_) | ty::RawPtr(_, _) => (), // struct([u8; 4]) is ok
             _ => {
                 struct_span_code_err!(
                     tcx.dcx(),

compiler/rustc_middle/src/ty/sty.rs

@@ -1091,29 +1091,21 @@ impl<'tcx> Ty<'tcx> {
     }
 
     pub fn simd_size_and_type(self, tcx: TyCtxt<'tcx>) -> (u64, Ty<'tcx>) {
-        match self.kind() {
-            Adt(def, args) => {
-                assert!(def.repr().simd(), "`simd_size_and_type` called on non-SIMD type");
-                let variant = def.non_enum_variant();
-                let f0_ty = variant.fields[FieldIdx::ZERO].ty(tcx, args);
-
-                match f0_ty.kind() {
-                    // If the first field is an array, we assume it is the only field and its
-                    // elements are the SIMD components.
-                    Array(f0_elem_ty, f0_len) => {
-                        // FIXME(repr_simd): https://github.com/rust-lang/rust/pull/78863#discussion_r522784112
-                        // The way we evaluate the `N` in `[T; N]` here only works since we use
-                        // `simd_size_and_type` post-monomorphization. It will probably start to ICE
-                        // if we use it in generic code. See the `simd-array-trait` ui test.
-                        (f0_len.eval_target_usize(tcx, ParamEnv::empty()), *f0_elem_ty)
-                    }
-                    // Otherwise, the fields of this Adt are the SIMD components (and we assume they
-                    // all have the same type).
-                    _ => (variant.fields.len() as u64, f0_ty),
-                }
-            }
-            _ => bug!("`simd_size_and_type` called on invalid type"),
-        }
+        let Adt(def, args) = self.kind() else {
+            bug!("`simd_size_and_type` called on invalid type")
+        };
+        assert!(def.repr().simd(), "`simd_size_and_type` called on non-SIMD type");
+        let variant = def.non_enum_variant();
+        assert_eq!(variant.fields.len(), 1);
+        let field_ty = variant.fields[FieldIdx::ZERO].ty(tcx, args);
+        let Array(f0_elem_ty, f0_len) = field_ty.kind() else {
+            bug!("Simd type has non-array field type {field_ty:?}")
+        };
+        // FIXME(repr_simd): https://github.com/rust-lang/rust/pull/78863#discussion_r522784112
+        // The way we evaluate the `N` in `[T; N]` here only works since we use
+        // `simd_size_and_type` post-monomorphization. It will probably start to ICE
+        // if we use it in generic code. See the `simd-array-trait` ui test.
+        (f0_len.eval_target_usize(tcx, ParamEnv::empty()), *f0_elem_ty)
     }
 
     #[inline]

library/alloc/benches/slice.rs

@@ -336,10 +336,10 @@ reverse!(reverse_u32, u32, |x| x as u32);
 reverse!(reverse_u64, u64, |x| x as u64);
 reverse!(reverse_u128, u128, |x| x as u128);
 #[repr(simd)]
-struct F64x4(f64, f64, f64, f64);
+struct F64x4([f64; 4]);
 reverse!(reverse_simd_f64x4, F64x4, |x| {
     let x = x as f64;
-    F64x4(x, x, x, x)
+    F64x4([x, x, x, x])
 });
 
 macro_rules! rotate {

tests/assembly/asm/aarch64-types.rs

@@ -31,36 +31,39 @@ trait Sized {}
 #[lang = "copy"]
 trait Copy {}
 
+// Do we really need to use no_core for this?!?
+impl<T: Copy, const N: usize> Copy for [T; N] {}
+
 type ptr = *mut u8;
 
 #[repr(simd)]
-pub struct i8x8(i8, i8, i8, i8, i8, i8, i8, i8);
+pub struct i8x8([i8; 8]);
 #[repr(simd)]
-pub struct i16x4(i16, i16, i16, i16);
+pub struct i16x4([i16; 4]);
 #[repr(simd)]
-pub struct i32x2(i32, i32);
+pub struct i32x2([i32; 2]);
 #[repr(simd)]
-pub struct i64x1(i64);
+pub struct i64x1([i64; 1]);
 #[repr(simd)]
-pub struct f16x4(f16, f16, f16, f16);
+pub struct f16x4([f16; 4]);
 #[repr(simd)]
-pub struct f32x2(f32, f32);
+pub struct f32x2([f32; 2]);
 #[repr(simd)]
-pub struct f64x1(f64);
+pub struct f64x1([f64; 1]);
 #[repr(simd)]
-pub struct i8x16(i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8);
+pub struct i8x16([i8; 16]);
 #[repr(simd)]
-pub struct i16x8(i16, i16, i16, i16, i16, i16, i16, i16);
+pub struct i16x8([i16; 8]);
 #[repr(simd)]
-pub struct i32x4(i32, i32, i32, i32);
+pub struct i32x4([i32; 4]);
 #[repr(simd)]
-pub struct i64x2(i64, i64);
+pub struct i64x2([i64; 2]);
 #[repr(simd)]
-pub struct f16x8(f16, f16, f16, f16, f16, f16, f16, f16);
+pub struct f16x8([f16; 8]);
 #[repr(simd)]
-pub struct f32x4(f32, f32, f32, f32);
+pub struct f32x4([f32; 4]);
 #[repr(simd)]
-pub struct f64x2(f64, f64);
+pub struct f64x2([f64; 2]);
 
 impl Copy for i8 {}
 impl Copy for i16 {}

tests/assembly/asm/arm-modifiers.rs

@@ -28,8 +28,11 @@ trait Sized {}
 #[lang = "copy"]
 trait Copy {}
 
+// Do we really need to use no_core for this?!?
+impl<T: Copy, const N: usize> Copy for [T; N] {}
+
 #[repr(simd)]
-pub struct f32x4(f32, f32, f32, f32);
+pub struct f32x4([f32; 4]);
 
 impl Copy for i32 {}
 impl Copy for f32 {}

tests/assembly/asm/arm-types.rs

@@ -31,32 +31,35 @@ trait Sized {}
 #[lang = "copy"]
 trait Copy {}
 
+// Do we really need to use no_core for this?!?
+impl<T: Copy, const N: usize> Copy for [T; N] {}
+
 type ptr = *mut u8;
 
 #[repr(simd)]
-pub struct i8x8(i8, i8, i8, i8, i8, i8, i8, i8);
+pub struct i8x8([i8; 8]);
 #[repr(simd)]
-pub struct i16x4(i16, i16, i16, i16);
+pub struct i16x4([i16; 4]);
 #[repr(simd)]
-pub struct i32x2(i32, i32);
+pub struct i32x2([i32; 2]);
 #[repr(simd)]
-pub struct i64x1(i64);
+pub struct i64x1([i64; 1]);
 #[repr(simd)]
-pub struct f16x4(f16, f16, f16, f16);
+pub struct f16x4([f16; 4]);
 #[repr(simd)]
-pub struct f32x2(f32, f32);
+pub struct f32x2([f32; 2]);
 #[repr(simd)]
-pub struct i8x16(i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8, i8);
+pub struct i8x16([i8; 16]);
 #[repr(simd)]
-pub struct i16x8(i16, i16, i16, i16, i16, i16, i16, i16);
+pub struct i16x8([i16; 8]);
 #[repr(simd)]
-pub struct i32x4(i32, i32, i32, i32);
+pub struct i32x4([i32; 4]);
 #[repr(simd)]
-pub struct i64x2(i64, i64);
+pub struct i64x2([i64; 2]);
 #[repr(simd)]
-pub struct f16x8(f16, f16, f16, f16, f16, f16, f16, f16);
+pub struct f16x8([f16; 8]);
 #[repr(simd)]
-pub struct f32x4(f32, f32, f32, f32);
+pub struct f32x4([f32; 4]);
 
 impl Copy for i8 {}
 impl Copy for i16 {}