diff --git a/compiler/rustc_const_eval/src/interpret/place.rs b/compiler/rustc_const_eval/src/interpret/place.rs
index 4a86ec3f57a4a..046ff34e3d07b 100644
--- a/compiler/rustc_const_eval/src/interpret/place.rs
+++ b/compiler/rustc_const_eval/src/interpret/place.rs
@@ -499,13 +499,14 @@ where
         &self,
         mplace: &MPlaceTy<'tcx, M::Provenance>,
     ) -> InterpResult<'tcx, (MPlaceTy<'tcx, M::Provenance>, u64)> {
-        // Basically we just transmute this place into an array following simd_size_and_type.
-        // (Transmuting is okay since this is an in-memory place. We also double-check the size
-        // stays the same.)
+        // Basically we want to transmute this place into an array following simd_size_and_type.
         let (len, e_ty) = mplace.layout.ty.simd_size_and_type(*self.tcx);
-        let array = Ty::new_array(self.tcx.tcx, e_ty, len);
-        let layout = self.layout_of(array)?;
-        let mplace = mplace.transmute(layout, self)?;
+        // Some SIMD types have padding, so `len` many `e_ty` does not cover the entire place.
+        // Therefore we cannot transmute, and instead we project at offset 0, which side-steps
+        // the size check.
+        let array_layout = self.layout_of(Ty::new_array(self.tcx.tcx, e_ty, len))?;
+        assert!(array_layout.size <= mplace.layout.size);
+        let mplace = mplace.offset(Size::ZERO, array_layout, self)?;
         Ok((mplace, len))
     }
 
diff --git a/compiler/rustc_const_eval/src/interpret/projection.rs b/compiler/rustc_const_eval/src/interpret/projection.rs
index 0e594914c3a79..09e1a59dfa1a7 100644
--- a/compiler/rustc_const_eval/src/interpret/projection.rs
+++ b/compiler/rustc_const_eval/src/interpret/projection.rs
@@ -81,6 +81,8 @@ pub trait Projectable<'tcx, Prov: Provenance>: Sized + std::fmt::Debug {
         ecx: &InterpCx<'tcx, M>,
     ) -> InterpResult<'tcx, Self> {
         assert!(layout.is_sized());
+        // We sometimes do pointer arithmetic with this function, disregarding the source type.
+        // So we don't check the sizes here.
         self.offset_with_meta(offset, OffsetMode::Inbounds, MemPlaceMeta::None, layout, ecx)
     }
 
diff --git a/src/tools/miri/tests/pass/intrinsics/portable-simd.rs b/src/tools/miri/tests/pass/intrinsics/portable-simd.rs
index c4ec45e4bbef9..03d9fc0a76fe1 100644
--- a/src/tools/miri/tests/pass/intrinsics/portable-simd.rs
+++ b/src/tools/miri/tests/pass/intrinsics/portable-simd.rs
@@ -658,11 +658,32 @@ fn simd_masked_loadstore() {
     assert_eq!(buf, [2, 3, 4]);
 }
 
+fn simd_ops_non_pow2() {
+    // Just a little smoke test for operations on non-power-of-two vectors.
+    #[repr(simd, packed)]
+    #[derive(Copy, Clone)]
+    pub struct SimdPacked<T, const N: usize>([T; N]);
+    #[repr(simd)]
+    #[derive(Copy, Clone)]
+    pub struct SimdPadded<T, const N: usize>([T; N]);
+
+    let x = SimdPacked([1u32; 3]);
+    let y = SimdPacked([2u32; 3]);
+    let z = unsafe { intrinsics::simd_add(x, y) };
+    assert_eq!({ z.0 }, [3u32; 3]);
+
+    let x = SimdPadded([1u32; 3]);
+    let y = SimdPadded([2u32; 3]);
+    let z = unsafe { intrinsics::simd_add(x, y) };
+    assert_eq!(z.0, [3u32; 3]);
+}
+
 fn main() {
     simd_mask();
     simd_ops_f32();
     simd_ops_f64();
     simd_ops_i32();
+    simd_ops_non_pow2();
     simd_cast();
     simd_swizzle();
     simd_gather_scatter();