diff --git a/hercules_cg/src/cpu.rs b/hercules_cg/src/cpu.rs
index a19218c7ed459834a605b9fd2adde795b479170d..f90657a4b81313a2dcadd6405bc86635c51606a0 100644
--- a/hercules_cg/src/cpu.rs
+++ b/hercules_cg/src/cpu.rs
@@ -507,11 +507,10 @@ impl<'a> CPUContext<'a> {
                     // load.
                     write!(
                         body,
-                        "  {} = load {}, ptr {}, align {}\n",
+                        "  {} = load {}, ptr {}\n",
                         self.get_value(id, false),
                         self.get_type(self_ty),
                         index_ptr_name,
-                        get_type_alignment(self.types, collect_ty),
                     )?;
                 } else {
                     // If this read doesn't reach a primitive type, just return
@@ -540,10 +539,9 @@ impl<'a> CPUContext<'a> {
                     // perform a single store of the data value.
                     write!(
                         body,
-                        "  store {}, ptr {}, align {}\n",
+                        "  store {}, ptr {}\n",
                         self.get_value(data, true),
                         index_ptr_name,
-                        get_type_alignment(self.types, collect_ty),
                     )?;
                 } else {
                     // If the data item being written is not a primitive type,
diff --git a/hercules_cg/src/lib.rs b/hercules_cg/src/lib.rs
index 15946f72f86c44eaa2bf537450c1d22a57ba6992..af2420d83a550e7a50ff22962302612f21627995 100644
--- a/hercules_cg/src/lib.rs
+++ b/hercules_cg/src/lib.rs
@@ -16,7 +16,7 @@ use std::collections::BTreeMap;
 
 use hercules_ir::*;
 
-pub const LARGEST_ALIGNMENT: usize = 8;
+pub const LARGEST_ALIGNMENT: usize = 32;
 
 /*
  * The alignment of a type does not depend on dynamic constants.
@@ -33,7 +33,8 @@ pub fn get_type_alignment(types: &Vec<Type>, ty: TypeID) -> usize {
             .map(|id| get_type_alignment(types, *id))
             .max()
             .unwrap_or(1),
-        Type::Array(elem, _) => get_type_alignment(types, elem),
+        // Use a large alignment for arrays to generate better vector code.
+        Type::Array(_, _) => LARGEST_ALIGNMENT,
     }
 }
 
diff --git a/hercules_rt/src/lib.rs b/hercules_rt/src/lib.rs
index f8fdf2effa9bfc9bbf900b79915b8716706bde6c..0c3ffd8032cc6873f65d91245ff845576d4e2d2a 100644
--- a/hercules_rt/src/lib.rs
+++ b/hercules_rt/src/lib.rs
@@ -13,8 +13,10 @@ use std::sync::OnceLock;
  * src/rt.rs (the RT backend).
  */
 
+pub const LARGEST_ALIGNMENT: usize = 32;
+
 pub unsafe fn __cpu_alloc(size: usize) -> *mut u8 {
-    let ptr = alloc(Layout::from_size_align(size, 16).unwrap());
+    let ptr = alloc(Layout::from_size_align(size, LARGEST_ALIGNMENT).unwrap());
     if cfg!(feature = "debug") {
         eprintln!("__cpu_alloc: {:?}, {}", ptr, size);
         assert!(!ptr.is_null() || size == 0);
@@ -27,7 +29,7 @@ pub unsafe fn __cpu_dealloc(ptr: *mut u8, size: usize) {
         eprintln!("__cpu_dealloc: {:?}, {}", ptr, size);
         assert!(!ptr.is_null() || size == 0);
     }
-    dealloc(ptr, Layout::from_size_align(size, 16).unwrap())
+    dealloc(ptr, Layout::from_size_align(size, LARGEST_ALIGNMENT).unwrap())
 }
 
 pub unsafe fn __cpu_zero_mem(ptr: *mut u8, size: usize) {