diff --git a/std/assembly/array.ts b/std/assembly/array.ts
index 1771618b..2fcdf6fd 100644
--- a/std/assembly/array.ts
+++ b/std/assembly/array.ts
@@ -35,8 +35,6 @@ export class Array<T> extends ArrayBufferView {
   // length is modified in a way that a null value would exist. Otherwise, the compiler wouldn't be
   // able to guarantee type-safety anymore. For lack of a better word, such an array is "holey".
 
-  // Also note that capacity, not length, indicates the actual retained contents.
-
   private length_: i32;
 
   static isArray<U>(value: U): bool {
@@ -69,14 +67,23 @@ export class Array<T> extends ArrayBufferView {
     return this.length_;
   }
 
-  set length(length: i32) {
+  set length(newLength: i32) {
+    var oldLength = this.length_;
     if (isReference<T>()) {
       if (!isNullable<T>()) {
-        if (<u32>length > <u32>this.length_) throw new Error(E_HOLEYARRAY);
+        if (<u32>newLength > <u32>oldLength) throw new Error(E_HOLEYARRAY);
       }
     }
-    ensureSize(changetype<usize>(this), length, alignof<T>());
-    this.length_ = length;
+    ensureSize(changetype<usize>(this), newLength, alignof<T>());
+    if (isManaged<T>()) { // release no longer used refs
+      if (oldLength > newLength) {
+        let dataStart = this.dataStart;
+        do __release(load<usize>(dataStart + (<usize>--oldLength << alignof<T>())));
+        while (oldLength > newLength);
+        // no need to zero memory on shrink -> is zeroed on grow
+      }
+    }
+    this.length_ = newLength;
   }
 
   every(fn: (value: T, index: i32, array: Array<T>) => bool): bool {
@@ -195,12 +202,7 @@ export class Array<T> extends ArrayBufferView {
     var newLength = length + 1;
     ensureSize(changetype<usize>(this), newLength, alignof<T>());
     if (isManaged<T>()) {
-      let offset = this.dataStart + (<usize>length << alignof<T>());
-      let oldRef: usize = load<usize>(offset);
-      if (changetype<usize>(value) != oldRef) {
-        store<usize>(offset, __retain(changetype<usize>(value)));
-        __release(oldRef);
-      }
+      store<usize>(this.dataStart + (<usize>length << alignof<T>()), __retain(changetype<usize>(value)));
     } else {
       store<T>(this.dataStart + (<usize>length << alignof<T>()), value);
     }
@@ -247,23 +249,35 @@ export class Array<T> extends ArrayBufferView {
     var count = min(last - from, len - to);
 
     if (isManaged<T>()) {
-      // TODO: retain/release + consider intersection, only releasing what's removed
-      ERROR("not implemented");
-    } else {
-      if (from < to && to < (from + count)) {
+      if (from < to && to < (from + count)) { // right to left
         from += count - 1;
         to   += count - 1;
         while (count) {
-          store<T>(dataStart + (<usize>to << alignof<T>()), load<T>(dataStart + (<usize>from << alignof<T>())));
+          let oldRef: usize = load<usize>(dataStart + (<usize>to << alignof<T>()));
+          let newRef: usize = load<usize>(dataStart + (<usize>from << alignof<T>()));
+          if (newRef != oldRef) {
+            store<usize>(dataStart + (<usize>to << alignof<T>()), __retain(newRef));
+            __release(oldRef);
+          }
           --from, --to, --count;
         }
-      } else {
-        memory.copy(
-          dataStart + (<usize>to << alignof<T>()),
-          dataStart + (<usize>from << alignof<T>()),
-          <usize>count << alignof<T>()
-        );
+      } else { // left to right
+        while (count) {
+          let oldRef: usize = load<usize>(dataStart + (<usize>to << alignof<T>()));
+          let newRef: usize = load<usize>(dataStart + (<usize>from << alignof<T>()));
+          if (newRef != oldRef) {
+            store<usize>(dataStart + (<usize>to << alignof<T>()), __retain(newRef));
+            __release(oldRef);
+          }
+          ++from, ++to, --count;
+        }
       }
+    } else {
+      memory.copy( // is memmove
+        dataStart + (<usize>to << alignof<T>()),
+        dataStart + (<usize>from << alignof<T>()),
+        <usize>count << alignof<T>()
+      );
     }
     return this;
   }
@@ -273,7 +287,7 @@ export class Array<T> extends ArrayBufferView {
     if (length < 1) throw new RangeError(E_EMPTYARRAY);
     var element = load<T>(this.dataStart + (<usize>(--length) << alignof<T>()));
     this.length_ = length;
-    return element;
+    return element; // no need to retain -> is moved
   }
 
   forEach(fn: (value: T, index: i32, array: Array<T>) => void): void {
@@ -345,7 +359,7 @@ export class Array<T> extends ArrayBufferView {
       <T>null
     );
     this.length_ = lastIndex;
-    return element;
+    return element; // no need to retain -> is moved
   }
 
   some(fn: (value: T, index: i32, array: Array<T>) => bool): bool {
@@ -408,7 +422,7 @@ export class Array<T> extends ArrayBufferView {
         store<usize>(resultStart + (<usize>i << alignof<T>()),
           load<usize>(thisBase + (<usize>i << alignof<T>()))
         );
-        // element is moved, so refcount doesn't change
+        // no need to retain -> is moved
       }
     } else {
       memory.copy(
@@ -713,7 +727,7 @@ export class Array<T> extends ArrayBufferView {
   @unsafe private __visit_impl(cookie: u32): void {
     if (isManaged<T>()) {
       let cur = this.dataStart;
-      let end = cur + <usize>this.dataLength;
+      let end = cur + (<usize>this.length_ << alignof<T>());
       while (cur < end) {
         let val = load<usize>(cur);
         if (val) __visit(val, cookie);
diff --git a/tests/compiler/assert-nonnull.optimized.wat b/tests/compiler/assert-nonnull.optimized.wat
index 416aa0f0..ed5bd823 100644
--- a/tests/compiler/assert-nonnull.optimized.wat
+++ b/tests/compiler/assert-nonnull.optimized.wat
@@ -64,7 +64,7 @@
   if
    i32.const 24
    i32.const 136
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -78,7 +78,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -105,7 +105,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
diff --git a/tests/compiler/assert-nonnull.untouched.wat b/tests/compiler/assert-nonnull.untouched.wat
index 57d33b37..780007f7 100644
--- a/tests/compiler/assert-nonnull.untouched.wat
+++ b/tests/compiler/assert-nonnull.untouched.wat
@@ -106,7 +106,7 @@
   if
    i32.const 24
    i32.const 136
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -120,7 +120,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -168,7 +168,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
diff --git a/tests/compiler/retain-release-sanity.optimized.wat b/tests/compiler/retain-release-sanity.optimized.wat
index c87cc3af..1aaa5944 100644
--- a/tests/compiler/retain-release-sanity.optimized.wat
+++ b/tests/compiler/retain-release-sanity.optimized.wat
@@ -2015,7 +2015,7 @@
   if
    i32.const 424
    i32.const 376
-   i32.const 273
+   i32.const 287
    i32.const 20
    call $~lib/builtins/abort
    unreachable
@@ -2038,7 +2038,6 @@
  (func $~lib/array/Array<~lib/string/String>#push (; 32 ;) (type $FUNCSIG$vi) (param $0 i32)
   (local $1 i32)
   (local $2 i32)
-  (local $3 i32)
   i32.const 584
   call $~lib/rt/pure/__retain
   drop
@@ -2056,19 +2055,9 @@
   i32.const 2
   i32.shl
   i32.add
-  local.tee $1
-  i32.load
-  local.tee $3
   i32.const 584
-  i32.ne
-  if
-   local.get $1
-   i32.const 584
-   call $~lib/rt/pure/__retain
-   i32.store
-   local.get $3
-   call $~lib/rt/pure/__release
-  end
+  call $~lib/rt/pure/__retain
+  i32.store
   local.get $0
   local.get $2
   i32.store offset=12
@@ -2738,7 +2727,9 @@
   i32.load offset=4
   local.tee $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $0
   loop $continue|0
diff --git a/tests/compiler/retain-release-sanity.untouched.wat b/tests/compiler/retain-release-sanity.untouched.wat
index 4567ba53..529966f6 100644
--- a/tests/compiler/retain-release-sanity.untouched.wat
+++ b/tests/compiler/retain-release-sanity.untouched.wat
@@ -3782,7 +3782,7 @@
   if
    i32.const 424
    i32.const 376
-   i32.const 273
+   i32.const 287
    i32.const 20
    call $~lib/builtins/abort
    unreachable
@@ -3824,7 +3824,7 @@
   if
    i32.const 472
    i32.const 376
-   i32.const 58
+   i32.const 56
    i32.const 20
    call $~lib/builtins/abort
    unreachable
@@ -3855,7 +3855,7 @@
   if
    i32.const 472
    i32.const 376
-   i32.const 58
+   i32.const 56
    i32.const 20
    call $~lib/builtins/abort
    unreachable
@@ -3869,7 +3869,6 @@
   (local $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
   local.get $1
   call $~lib/rt/pure/__retain
   drop
@@ -3890,29 +3889,17 @@
   i32.const 2
   i32.shl
   i32.add
-  local.set $4
-  local.get $4
-  i32.load
-  local.set $5
   local.get $1
-  local.get $5
-  i32.ne
-  if
-   local.get $4
-   local.get $1
-   call $~lib/rt/pure/__retain
-   i32.store
-   local.get $5
-   call $~lib/rt/pure/__release
-  end
+  call $~lib/rt/pure/__retain
+  i32.store
   local.get $0
   local.get $3
   i32.store offset=12
   local.get $3
-  local.set $5
+  local.set $4
   local.get $1
   call $~lib/rt/pure/__release
-  local.get $5
+  local.get $4
  )
  (func $~lib/string/String#get:length (; 37 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
   local.get $0
@@ -4728,7 +4715,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -4764,7 +4753,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
diff --git a/tests/compiler/std/array-access.optimized.wat b/tests/compiler/std/array-access.optimized.wat
index e480a43b..0effe66d 100644
--- a/tests/compiler/std/array-access.optimized.wat
+++ b/tests/compiler/std/array-access.optimized.wat
@@ -26,7 +26,7 @@
   if
    i32.const 24
    i32.const 136
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -40,7 +40,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -63,7 +63,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
diff --git a/tests/compiler/std/array-access.untouched.wat b/tests/compiler/std/array-access.untouched.wat
index cc1f1ae2..ee37d0f1 100644
--- a/tests/compiler/std/array-access.untouched.wat
+++ b/tests/compiler/std/array-access.untouched.wat
@@ -43,7 +43,7 @@
   if
    i32.const 24
    i32.const 136
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -57,7 +57,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -85,7 +85,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -134,7 +134,7 @@
   if
    i32.const 24
    i32.const 136
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -148,7 +148,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -374,7 +374,7 @@
   if
    i32.const 24
    i32.const 136
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -388,7 +388,7 @@
   if
    i32.const 184
    i32.const 136
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
diff --git a/tests/compiler/std/array-literal.optimized.wat b/tests/compiler/std/array-literal.optimized.wat
index 3bdd0d17..f1827972 100644
--- a/tests/compiler/std/array-literal.optimized.wat
+++ b/tests/compiler/std/array-literal.optimized.wat
@@ -47,7 +47,7 @@
   if
    i32.const 136
    i32.const 192
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -68,7 +68,7 @@
   if
    i32.const 136
    i32.const 192
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -2170,7 +2170,9 @@
   i32.load offset=4
   local.tee $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $0
   loop $continue|0
diff --git a/tests/compiler/std/array-literal.untouched.wat b/tests/compiler/std/array-literal.untouched.wat
index a42ca1e6..a5bdca6c 100644
--- a/tests/compiler/std/array-literal.untouched.wat
+++ b/tests/compiler/std/array-literal.untouched.wat
@@ -70,7 +70,7 @@
   if
    i32.const 136
    i32.const 192
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -102,7 +102,7 @@
   if
    i32.const 136
    i32.const 192
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -4009,7 +4009,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -4045,7 +4047,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
diff --git a/tests/compiler/std/array.optimized.wat b/tests/compiler/std/array.optimized.wat
index 921ccc02..6fe95033 100644
--- a/tests/compiler/std/array.optimized.wat
+++ b/tests/compiler/std/array.optimized.wat
@@ -2367,7 +2367,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -2521,7 +2521,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -2858,7 +2858,7 @@
   if
    i32.const 872
    i32.const 488
-   i32.const 273
+   i32.const 287
    i32.const 20
    call $~lib/builtins/abort
    unreachable
@@ -2905,7 +2905,7 @@
    call $~lib/rt/pure/__release
    i32.const 272
    i32.const 488
-   i32.const 215
+   i32.const 217
    i32.const 59
    call $~lib/builtins/abort
    unreachable
@@ -2942,9 +2942,6 @@
  (func $~lib/array/Array<i32>#copyWithin (; 50 ;) (type $FUNCSIG$iiiii) (param $0 i32) (param $1 i32) (param $2 i32) (param $3 i32) (result i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.load offset=4
-  local.set $5
   local.get $3
   local.get $0
   i32.load offset=12
@@ -2954,6 +2951,9 @@
   i32.lt_s
   select
   local.set $3
+  local.get $0
+  i32.load offset=4
+  local.tee $5
   local.get $1
   i32.const 0
   i32.lt_s
@@ -2975,7 +2975,10 @@
    i32.lt_s
    select
   end
-  local.set $1
+  local.tee $1
+  i32.const 2
+  i32.shl
+  i32.add
   local.get $2
   i32.const 0
   i32.lt_s
@@ -2997,7 +3000,11 @@
    i32.lt_s
    select
   end
-  local.set $2
+  local.tee $2
+  i32.const 2
+  i32.shl
+  local.get $5
+  i32.add
   local.get $3
   i32.const 0
   i32.lt_s
@@ -3021,84 +3028,18 @@
   end
   local.get $2
   i32.sub
-  local.tee $3
+  local.tee $2
   local.get $4
   local.get $1
   i32.sub
-  local.tee $4
-  local.get $3
-  local.get $4
-  i32.lt_s
-  select
-  local.set $3
-  local.get $1
-  local.get $2
-  local.get $3
-  i32.add
-  i32.lt_s
-  i32.const 0
+  local.tee $1
   local.get $2
   local.get $1
   i32.lt_s
   select
-  if
-   local.get $3
-   i32.const 1
-   i32.sub
-   local.tee $4
-   local.get $2
-   i32.add
-   local.set $2
-   local.get $1
-   local.get $4
-   i32.add
-   local.set $1
-   loop $continue|0
-    local.get $3
-    if
-     local.get $1
-     i32.const 2
-     i32.shl
-     local.get $5
-     i32.add
-     local.get $2
-     i32.const 2
-     i32.shl
-     local.get $5
-     i32.add
-     i32.load
-     i32.store
-     local.get $2
-     i32.const 1
-     i32.sub
-     local.set $2
-     local.get $1
-     i32.const 1
-     i32.sub
-     local.set $1
-     local.get $3
-     i32.const 1
-     i32.sub
-     local.set $3
-     br $continue|0
-    end
-   end
-  else   
-   local.get $1
-   i32.const 2
-   i32.shl
-   local.get $5
-   i32.add
-   local.get $2
-   i32.const 2
-   i32.shl
-   local.get $5
-   i32.add
-   local.get $3
-   i32.const 2
-   i32.shl
-   call $~lib/memory/memory.copy
-  end
+  i32.const 2
+  i32.shl
+  call $~lib/memory/memory.copy
   local.get $0
   call $~lib/rt/pure/__retain
  )
@@ -3144,7 +3085,7 @@
   if
    i32.const 872
    i32.const 488
-   i32.const 334
+   i32.const 348
    i32.const 20
    call $~lib/builtins/abort
    unreachable
@@ -3985,7 +3926,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -5397,7 +5338,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -5911,7 +5852,7 @@
   if
    i32.const 272
    i32.const 488
-   i32.const 47
+   i32.const 45
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -6187,7 +6128,7 @@
    call $~lib/rt/pure/__release
    i32.const 4040
    i32.const 488
-   i32.const 114
+   i32.const 121
    i32.const 38
    call $~lib/builtins/abort
    unreachable
@@ -6443,7 +6384,7 @@
   if
    i32.const 4040
    i32.const 488
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -6457,7 +6398,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -6636,7 +6577,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -17031,7 +16972,9 @@
   i32.load offset=4
   local.tee $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $0
   loop $continue|0
diff --git a/tests/compiler/std/array.untouched.wat b/tests/compiler/std/array.untouched.wat
index b1e44bc6..aa775d7f 100644
--- a/tests/compiler/std/array.untouched.wat
+++ b/tests/compiler/std/array.untouched.wat
@@ -4243,7 +4243,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -4448,7 +4448,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -4856,7 +4856,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -4877,7 +4877,7 @@
   if
    i32.const 872
    i32.const 488
-   i32.const 273
+   i32.const 287
    i32.const 20
    call $~lib/builtins/abort
    unreachable
@@ -4933,7 +4933,7 @@
    block
     i32.const 272
     i32.const 488
-    i32.const 215
+    i32.const 217
     i32.const 59
     call $~lib/builtins/abort
     unreachable
@@ -5087,79 +5087,20 @@
   i32.lt_s
   select
   local.set $11
-  local.get $9
+  local.get $4
   local.get $8
-  i32.lt_s
-  if (result i32)
-   local.get $8
-   local.get $9
-   local.get $11
-   i32.add
-   i32.lt_s
-  else   
-   i32.const 0
-  end
-  if
-   local.get $9
-   local.get $11
-   i32.const 1
-   i32.sub
-   i32.add
-   local.set $9
-   local.get $8
-   local.get $11
-   i32.const 1
-   i32.sub
-   i32.add
-   local.set $8
-   block $break|0
-    loop $continue|0
-     local.get $11
-     if
-      local.get $4
-      local.get $8
-      i32.const 2
-      i32.shl
-      i32.add
-      local.get $4
-      local.get $9
-      i32.const 2
-      i32.shl
-      i32.add
-      i32.load
-      i32.store
-      local.get $9
-      i32.const 1
-      i32.sub
-      local.set $9
-      local.get $8
-      i32.const 1
-      i32.sub
-      local.set $8
-      local.get $11
-      i32.const 1
-      i32.sub
-      local.set $11
-      br $continue|0
-     end
-    end
-   end
-  else   
-   local.get $4
-   local.get $8
-   i32.const 2
-   i32.shl
-   i32.add
-   local.get $4
-   local.get $9
-   i32.const 2
-   i32.shl
-   i32.add
-   local.get $11
-   i32.const 2
-   i32.shl
-   call $~lib/memory/memory.copy
-  end
+  i32.const 2
+  i32.shl
+  i32.add
+  local.get $4
+  local.get $9
+  i32.const 2
+  i32.shl
+  i32.add
+  local.get $11
+  i32.const 2
+  i32.shl
+  call $~lib/memory/memory.copy
   local.get $0
   call $~lib/rt/pure/__retain
  )
@@ -5296,7 +5237,7 @@
   if
    i32.const 872
    i32.const 488
-   i32.const 334
+   i32.const 348
    i32.const 20
    call $~lib/builtins/abort
    unreachable
@@ -6337,7 +6278,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -8288,7 +8229,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -9390,7 +9331,7 @@
   if
    i32.const 272
    i32.const 488
-   i32.const 47
+   i32.const 45
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -9669,7 +9610,7 @@
   if
    i32.const 272
    i32.const 488
-   i32.const 47
+   i32.const 45
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -9739,7 +9680,7 @@
    block
     i32.const 4040
     i32.const 488
-    i32.const 114
+    i32.const 121
     i32.const 38
     call $~lib/builtins/abort
     unreachable
@@ -10033,7 +9974,7 @@
   if
    i32.const 4040
    i32.const 488
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -10047,7 +9988,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -10161,7 +10102,7 @@
   if
    i32.const 272
    i32.const 488
-   i32.const 47
+   i32.const 45
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -10246,7 +10187,7 @@
    block
     i32.const 4040
     i32.const 488
-    i32.const 114
+    i32.const 121
     i32.const 38
     call $~lib/builtins/abort
     unreachable
@@ -10534,7 +10475,7 @@
   if
    i32.const 4040
    i32.const 488
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -10548,7 +10489,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -10853,7 +10794,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -11356,7 +11297,7 @@
   if
    i32.const 272
    i32.const 488
-   i32.const 47
+   i32.const 45
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -11632,7 +11573,7 @@
    block
     i32.const 4040
     i32.const 488
-    i32.const 114
+    i32.const 121
     i32.const 38
     call $~lib/builtins/abort
     unreachable
@@ -11898,7 +11839,7 @@
   if
    i32.const 4040
    i32.const 488
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -11912,7 +11853,7 @@
   if
    i32.const 176
    i32.const 488
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -24186,7 +24127,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -24222,7 +24165,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -24258,7 +24203,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -24294,7 +24241,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -24339,7 +24288,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -24384,7 +24335,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -24420,7 +24373,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
@@ -24456,7 +24411,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
diff --git a/tests/compiler/std/static-array.optimized.wat b/tests/compiler/std/static-array.optimized.wat
index da87c808..0fc9400f 100644
--- a/tests/compiler/std/static-array.optimized.wat
+++ b/tests/compiler/std/static-array.optimized.wat
@@ -37,7 +37,7 @@
   if
    i32.const 320
    i32.const 376
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -624,7 +624,7 @@
   if
    i32.const 320
    i32.const 376
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -668,7 +668,7 @@
   if
    i32.const 320
    i32.const 376
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -712,7 +712,7 @@
   if
    i32.const 320
    i32.const 376
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
diff --git a/tests/compiler/std/static-array.untouched.wat b/tests/compiler/std/static-array.untouched.wat
index b061cafb..afc869cb 100644
--- a/tests/compiler/std/static-array.untouched.wat
+++ b/tests/compiler/std/static-array.untouched.wat
@@ -59,7 +59,7 @@
   if
    i32.const 320
    i32.const 376
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -1999,7 +1999,7 @@
   if
    i32.const 320
    i32.const 376
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -2067,7 +2067,7 @@
   if
    i32.const 320
    i32.const 376
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -2135,7 +2135,7 @@
   if
    i32.const 320
    i32.const 376
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
diff --git a/tests/compiler/std/string.optimized.wat b/tests/compiler/std/string.optimized.wat
index bcc96242..4b0ca1c2 100644
--- a/tests/compiler/std/string.optimized.wat
+++ b/tests/compiler/std/string.optimized.wat
@@ -3751,7 +3751,6 @@
  (func $~lib/array/Array<~lib/string/String>#push (; 57 ;) (type $FUNCSIG$vii) (param $0 i32) (param $1 i32)
   (local $2 i32)
   (local $3 i32)
-  (local $4 i32)
   local.get $1
   call $~lib/rt/pure/__retain
   drop
@@ -3769,19 +3768,9 @@
   i32.const 2
   i32.shl
   i32.add
-  local.tee $2
-  i32.load
-  local.tee $4
   local.get $1
-  i32.ne
-  if
-   local.get $2
-   local.get $1
-   call $~lib/rt/pure/__retain
-   i32.store
-   local.get $4
-   call $~lib/rt/pure/__release
-  end
+  call $~lib/rt/pure/__retain
+  i32.store
   local.get $0
   local.get $3
   i32.store offset=12
@@ -4018,7 +4007,7 @@
   if
    i32.const 2296
    i32.const 2248
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -4032,7 +4021,7 @@
   if
    i32.const 232
    i32.const 2248
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -9166,7 +9155,9 @@
   i32.load offset=4
   local.tee $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $0
   loop $continue|0
diff --git a/tests/compiler/std/string.untouched.wat b/tests/compiler/std/string.untouched.wat
index fd7f3f21..d9868054 100644
--- a/tests/compiler/std/string.untouched.wat
+++ b/tests/compiler/std/string.untouched.wat
@@ -5943,7 +5943,6 @@
   (local $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
   local.get $1
   call $~lib/rt/pure/__retain
   drop
@@ -5964,29 +5963,17 @@
   i32.const 2
   i32.shl
   i32.add
-  local.set $4
-  local.get $4
-  i32.load
-  local.set $5
   local.get $1
-  local.get $5
-  i32.ne
-  if
-   local.get $4
-   local.get $1
-   call $~lib/rt/pure/__retain
-   i32.store
-   local.get $5
-   call $~lib/rt/pure/__release
-  end
+  call $~lib/rt/pure/__retain
+  i32.store
   local.get $0
   local.get $3
   i32.store offset=12
   local.get $3
-  local.set $5
+  local.set $4
   local.get $1
   call $~lib/rt/pure/__release
-  local.get $5
+  local.get $4
  )
  (func $~lib/string/String#split (; 61 ;) (type $FUNCSIG$iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
   (local $3 i32)
@@ -6331,7 +6318,7 @@
   if
    i32.const 2296
    i32.const 2248
-   i32.const 99
+   i32.const 106
    i32.const 45
    call $~lib/builtins/abort
    unreachable
@@ -6345,7 +6332,7 @@
   if
    i32.const 232
    i32.const 2248
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -12448,7 +12435,9 @@
   local.set $2
   local.get $2
   local.get $0
-  i32.load offset=8
+  i32.load offset=12
+  i32.const 2
+  i32.shl
   i32.add
   local.set $3
   block $break|0
diff --git a/tests/compiler/std/typedarray.optimized.wat b/tests/compiler/std/typedarray.optimized.wat
index 93e7c6aa..931f9eaa 100644
--- a/tests/compiler/std/typedarray.optimized.wat
+++ b/tests/compiler/std/typedarray.optimized.wat
@@ -3354,7 +3354,7 @@
   if
    i32.const 336
    i32.const 512
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -3605,7 +3605,7 @@
   if
    i32.const 336
    i32.const 512
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
diff --git a/tests/compiler/std/typedarray.untouched.wat b/tests/compiler/std/typedarray.untouched.wat
index c24bf736..0e620958 100644
--- a/tests/compiler/std/typedarray.untouched.wat
+++ b/tests/compiler/std/typedarray.untouched.wat
@@ -5439,7 +5439,7 @@
   if
    i32.const 336
    i32.const 512
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable
@@ -5766,7 +5766,7 @@
   if
    i32.const 336
    i32.const 512
-   i32.const 102
+   i32.const 109
    i32.const 61
    call $~lib/builtins/abort
    unreachable