Make repr(packed) vectors work with SIMD intrinsics

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -480,8 +480,60 @@ impl<'ll, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'_, 'll, 'tcx> {
             }
 
             _ if name.as_str().starts_with("simd_") => {
+                // Unpack non-power-of-2 #[repr(packed, simd)] arguments.
+                // This gives them the expected layout of a regular #[repr(simd)] vector.
+                let mut loaded_args = Vec::new();
+                for (ty, arg) in arg_tys.iter().zip(args) {
+                    loaded_args.push(
+                        // #[repr(packed, simd)] vectors are passed like arrays (as references,
+                        // with reduced alignment and no padding) rather than as immediates.
+                        // We can use a vector load to fix the layout and turn the argument
+                        // into an immediate.
+                        if ty.is_simd()
+                            && let OperandValue::Ref(place) = arg.val
+                        {
+                            let (size, elem_ty) = ty.simd_size_and_type(self.tcx());
+                            let elem_ll_ty = match elem_ty.kind() {
+                                ty::Float(f) => self.type_float_from_ty(*f),
+                                ty::Int(i) => self.type_int_from_ty(*i),
+                                ty::Uint(u) => self.type_uint_from_ty(*u),
+                                ty::RawPtr(_, _) => self.type_ptr(),
+                                _ => unreachable!(),
+                            };
+                            let loaded =
+                                self.load_from_place(self.type_vector(elem_ll_ty, size), place);
+                            OperandRef::from_immediate_or_packed_pair(self, loaded, arg.layout)
+                        } else {
+                            *arg
+                        },
+                    );
+                }
+
+                let llret_ty = if ret_ty.is_simd()
+                    && let abi::Abi::Aggregate { .. } = self.layout_of(ret_ty).layout.abi
+                {
+                    let (size, elem_ty) = ret_ty.simd_size_and_type(self.tcx());
+                    let elem_ll_ty = match elem_ty.kind() {
+                        ty::Float(f) => self.type_float_from_ty(*f),
+                        ty::Int(i) => self.type_int_from_ty(*i),
+                        ty::Uint(u) => self.type_uint_from_ty(*u),
+                        ty::RawPtr(_, _) => self.type_ptr(),
+                        _ => unreachable!(),
+                    };
+                    self.type_vector(elem_ll_ty, size)
+                } else {
+                    llret_ty
+                };
+
                 match generic_simd_intrinsic(
-                    self, name, callee_ty, fn_args, args, ret_ty, llret_ty, span,
+                    self,
+                    name,
+                    callee_ty,
+                    fn_args,
+                    &loaded_args,
+                    ret_ty,
+                    llret_ty,
+                    span,
                 ) {
                     Ok(llval) => llval,
                     Err(()) => return Ok(()),

tests/codegen/simd/packed-simd-alignment.rs

@@ -0,0 +1,44 @@
+//@ compile-flags: -Cno-prepopulate-passes
+
+#![crate_type = "lib"]
+#![feature(repr_simd, core_intrinsics)]
+// make sure that codegen emits correctly-aligned loads and stores for repr(packed, simd) types
+// the alignment of a load should be no less than T, and no more than the size of the vector type
+use std::intrinsics::simd as intrinsics;
+
+#[derive(Copy, Clone)]
+#[repr(packed, simd)]
+struct f32x3([f32; 3]);
+
+#[derive(Copy, Clone)]
+#[repr(packed, simd)]
+struct f32x4([f32; 4]);
+
+// CHECK-LABEL: load_f32x3
+#[no_mangle]
+pub fn load_f32x3(floats: &f32x3) -> f32x3 {
+    // FIXME: Is a memcpy really the best we can do?
+    // CHECK: @llvm.memcpy.{{.*}}ptr align 4 {{.*}}ptr align 4
+    *floats
+}
+
+// CHECK-LABEL: load_f32x4
+#[no_mangle]
+pub fn load_f32x4(floats: &f32x4) -> f32x4 {
+    // CHECK: load <4 x float>, ptr %{{[a-z0-9_]*}}, align {{4|8|16}}
+    *floats
+}
+
+// CHECK-LABEL: add_f32x3
+#[no_mangle]
+pub fn add_f32x3(x: f32x3, y: f32x3) -> f32x3 {
+    // CHECK: load <3 x float>, ptr %{{[a-z0-9_]*}}, align 4
+    unsafe { intrinsics::simd_add(x, y) }
+}
+
+// CHECK-LABEL: add_f32x4
+#[no_mangle]
+pub fn add_f32x4(x: f32x4, y: f32x4) -> f32x4 {
+    // CHECK: load <4 x float>, ptr %{{[a-z0-9_]*}}, align {{4|8|16}}
+    unsafe { intrinsics::simd_add(x, y) }
+}

tests/ui/simd/repr_packed.rs

@@ -6,9 +6,6 @@
 #[repr(simd, packed)]
 struct Simd<T, const N: usize>([T; N]);
 
-#[repr(simd)]
-struct FullSimd<T, const N: usize>([T; N]);
-
 fn check_size_align<T, const N: usize>() {
     use std::mem;
     assert_eq!(mem::size_of::<Simd<T, N>>(), mem::size_of::<[T; N]>());
@@ -39,21 +36,15 @@ fn main() {
     check_ty::<f64>();
 
     unsafe {
-        // powers-of-two have no padding and work as usual
+        // powers-of-two have no padding and have the same layout as #[repr(simd)]
         let x: Simd<f64, 4> =
             simd_add(Simd::<f64, 4>([0., 1., 2., 3.]), Simd::<f64, 4>([2., 2., 2., 2.]));
         assert_eq!(std::mem::transmute::<_, [f64; 4]>(x), [2., 3., 4., 5.]);
 
-        // non-powers-of-two have padding and need to be expanded to full vectors
-        fn load<T, const N: usize>(v: Simd<T, N>) -> FullSimd<T, N> {
-            unsafe {
-                let mut tmp = core::mem::MaybeUninit::<FullSimd<T, N>>::uninit();
-                std::ptr::copy_nonoverlapping(&v as *const _, tmp.as_mut_ptr().cast(), 1);
-                tmp.assume_init()
-            }
-        }
-        let x: FullSimd<f64, 3> =
-            simd_add(load(Simd::<f64, 3>([0., 1., 2.])), load(Simd::<f64, 3>([2., 2., 2.])));
-        assert_eq!(x.0, [2., 3., 4.]);
+        // non-powers-of-two should have padding (which is removed by #[repr(packed)]),
+        // but the intrinsic handles it
+        let x: Simd<f64, 3> = simd_add(Simd::<f64, 3>([0., 1., 2.]), Simd::<f64, 3>([2., 2., 2.]));
+        let arr: [f64; 3] = x.0;
+        assert_eq!(arr, [2., 3., 4.]);
     }
 }