Override component model Lower::store_list and Lift::load_list for f32/f64 (#9892)

* Override component model Lower::store_list and Lift::load_list for f32/f64

* work with [u8] instead of integers

* Import from `core` instead of `std`

Co-authored-by: Joel Dice <[email protected]>

---------

Co-authored-by: Joel Dice <[email protected]>

Author: kaivol

crates/wasmtime/src/runtime/component/func/typed.rs

@@ -8,6 +8,7 @@ use crate::{AsContextMut, StoreContext, StoreContextMut, ValRaw};
 use alloc::borrow::Cow;
 use alloc::sync::Arc;
 use core::fmt;
+use core::iter;
 use core::marker;
 use core::mem::{self, MaybeUninit};
 use core::ptr::NonNull;
@@ -942,6 +943,39 @@ macro_rules! floats {
                 *ptr = self.to_bits().to_le_bytes();
                 Ok(())
             }
+
+            fn store_list<T>(
+                cx: &mut LowerContext<'_, T>,
+                ty: InterfaceType,
+                offset: usize,
+                items: &[Self],
+            ) -> Result<()> {
+                debug_assert!(matches!(ty, InterfaceType::$ty));
+
+                // Double-check that the CM alignment is at least the host's
+                // alignment for this type which should be true for all
+                // platforms.
+                assert!((Self::ALIGN32 as usize) >= mem::align_of::<Self>());
+
+                // Slice `cx`'s memory to the window that we'll be modifying.
+                // This should all have already been verified in terms of
+                // alignment and sizing meaning that these assertions here are
+                // not truly necessary but are instead double-checks.
+                let dst = &mut cx.as_slice_mut()[offset..][..items.len() * Self::SIZE32];
+                assert!(dst.as_ptr().cast::<Self>().is_aligned());
+
+                // And with all that out of the way perform the copying loop.
+                // This is not a `copy_from_slice` because endianness needs to
+                // be handled here, but LLVM should pretty easily transform this
+                // into a memcpy on little-endian platforms.
+                // TODO use `as_chunks` when https://github.com/rust-lang/rust/issues/74985
+                // is stabilized
+                for (dst, src) in iter::zip(dst.chunks_exact_mut(Self::SIZE32), items) {
+                    let dst: &mut [u8; Self::SIZE32] = dst.try_into().unwrap();
+                    *dst = src.to_le_bytes();
+                }
+                Ok(())
+            }
         }
 
         unsafe impl Lift for $float {
@@ -957,6 +991,27 @@ macro_rules! floats {
                 debug_assert!((bytes.as_ptr() as usize) % Self::SIZE32 == 0);
                 Ok($float::from_le_bytes(bytes.try_into().unwrap()))
             }
+
+            fn load_list(cx: &mut LiftContext<'_>, list: &WasmList<Self>) -> Result<Vec<Self>> where Self: Sized {
+                // See comments in `WasmList::get` for the panicking indexing
+                let byte_size = list.len * mem::size_of::<Self>();
+                let bytes = &cx.memory()[list.ptr..][..byte_size];
+
+                // The canonical ABI requires that everything is aligned to its
+                // own size, so this should be an aligned array.
+                assert!(bytes.as_ptr().cast::<Self>().is_aligned());
+
+                // Copy the resulting slice to a new Vec, handling endianness
+                // in the process
+                // TODO use `as_chunks` when https://github.com/rust-lang/rust/issues/74985
+                // is stabilized
+                Ok(
+                    bytes
+                        .chunks_exact(Self::SIZE32)
+                        .map(|i| $float::from_le_bytes(i.try_into().unwrap()))
+                        .collect()
+                )
+            }
         }
     };)*)
 }