diff --git a/src/constants.ts b/src/constants.ts
index c596b068..c5a989d1 100644
--- a/src/constants.ts
+++ b/src/constants.ts
@@ -10,5 +10,5 @@ export const GETTER_PREFIX = "get:";
 export const SETTER_PREFIX = "set:";
 /** Delimiter used between class names and instance members. */
 export const INSTANCE_DELIMITER = "#";
-/** Delimited used between class and namespace names and static members. */
+/** Delimiter used between class and namespace names and static members. */
 export const STATIC_DELIMITER = ".";
diff --git a/src/glue/js.js b/src/glue/js.js
index 877ebeb6..19afc9d4 100644
--- a/src/glue/js.js
+++ b/src/glue/js.js
@@ -14,23 +14,17 @@ for (var key in binaryen)
   if (/^_(?:Binaryen|Relooper)/.test(key))
     globalScope[key] = binaryen[key];
 
-// Use Binaryen's heap
-Object.defineProperties(globalScope["Heap"] = {
-  allocate: function allocate(size) {
-    if (!size) return 0; // should be safe in our case
-    return binaryen._malloc(size);
-  },
-  dispose: function dispose(ptr) {
-    if (ptr) binaryen._free(ptr);
-  },
-  copy: function copy(dest, src, n) {
-    return binaryen._memcpy(dest, src, n);
-  }
-}, {
-  free: { get: function() { return binaryen.HEAPU8.length; } },
-  used: { get: function() { return 0; } },
-  size: { get: function() { return binaryen.HEAPU8.length; } }
-});
+// Use Binaryen's heap instead of std heap
+globalScope["allocate_memory"] = function allocate_memory(size) {
+  if (!size) return 0; // should be safe in our case
+  return binaryen._malloc(size);
+};
+globalScope["free_memory"] = function free_memory(ptr) {
+  if (ptr) binaryen._free(ptr);
+};
+globalScope["move_memory"] = function move_memory(dest, src, n) {
+  return binaryen._memmove(dest, src, n);
+};
 globalScope["store"] = function store(ptr, val) {
   binaryen.HEAPU8[ptr] = val;
 };
@@ -42,11 +36,11 @@ globalScope["load"] = function load(ptr) {
 var Module = require("../module").Module;
 Module.prototype.toBinary = function toBinary(bufferSize) {
   if (!bufferSize) bufferSize = 1024 * 1024; // FIXME: see binaryen.js-post.js in Binaryen
-  var ptr = Heap.allocate(bufferSize);
+  var ptr = allocate_memory(bufferSize);
   var len = this.write(ptr, bufferSize);
   var ret = new Uint8Array(len);
   ret.set(binaryen.HEAPU8.subarray(ptr, ptr + len));
-  Heap.dispose(ptr);
+  free_memory(ptr);
   return ret;
 };
 Module.prototype.toText = function toText() {
diff --git a/src/module.ts b/src/module.ts
index df6816a7..4626c1d8 100644
--- a/src/module.ts
+++ b/src/module.ts
@@ -246,7 +246,7 @@ export class Module {
   static create(): Module {
     var module = new Module();
     module.ref = _BinaryenModuleCreate();
-    module.lit = changetype<BinaryenLiteral>(Heap.allocate(16));
+    module.lit = changetype<BinaryenLiteral>(allocate_memory(16));
     module.noEmit = false;
     return module;
   }
@@ -256,11 +256,11 @@ export class Module {
     try {
       var module = new Module();
       module.ref = _BinaryenModuleRead(cArr, buffer.length);
-      module.lit = changetype<BinaryenLiteral>(Heap.allocate(16));
+      module.lit = changetype<BinaryenLiteral>(allocate_memory(16));
       module.noEmit = false;
       return module;
     } finally {
-      Heap.dispose(changetype<usize>(cArr));
+      free_memory(changetype<usize>(cArr));
     }
   }
 
@@ -283,8 +283,8 @@ export class Module {
     try {
       return _BinaryenAddFunctionType(this.ref, cStr, result, cArr, paramTypes.length);
     } finally {
-      Heap.dispose(cArr);
-      Heap.dispose(cStr);
+      free_memory(cArr);
+      free_memory(cStr);
     }
   }
 
@@ -294,7 +294,7 @@ export class Module {
     try {
       return _BinaryenGetFunctionTypeBySignature(this.ref, result, cArr, paramTypes.length);
     } finally {
-      Heap.dispose(cArr);
+      free_memory(cArr);
     }
   }
 
@@ -341,8 +341,8 @@ export class Module {
     try {
       return _BinaryenHost(this.ref, op, cStr, cArr, operands ? (<ExpressionRef[]>operands).length : 0);
     } finally {
-      Heap.dispose(cArr);
-      Heap.dispose(cStr);
+      free_memory(cArr);
+      free_memory(cStr);
     }
   }
 
@@ -362,7 +362,7 @@ export class Module {
     try {
       return _BinaryenGetGlobal(this.ref, cStr, type);
     } finally {
-      Heap.dispose(cStr);
+      free_memory(cStr);
     }
   }
 
@@ -419,7 +419,7 @@ export class Module {
     try {
       return _BinaryenSetGlobal(this.ref, cStr, value);
     } finally {
-      Heap.dispose(cStr);
+      free_memory(cStr);
     }
   }
 
@@ -430,8 +430,8 @@ export class Module {
     try {
       return _BinaryenBlock(this.ref, cStr, cArr, children.length, type);
     } finally {
-      Heap.dispose(cArr);
-      Heap.dispose(cStr);
+      free_memory(cArr);
+      free_memory(cStr);
     }
   }
 
@@ -441,7 +441,7 @@ export class Module {
     try {
       return _BinaryenBreak(this.ref, cStr, condition, value);
     } finally {
-      Heap.dispose(cStr);
+      free_memory(cStr);
     }
   }
 
@@ -456,7 +456,7 @@ export class Module {
     try {
       return _BinaryenLoop(this.ref, cStr, body);
     } finally {
-      Heap.dispose(cStr);
+      free_memory(cStr);
     }
   }
 
@@ -490,9 +490,9 @@ export class Module {
     try {
       return _BinaryenSwitch(this.ref, cArr, k, cStr, condition, value);
     } finally {
-      Heap.dispose(cStr);
-      Heap.dispose(cArr);
-      for (i = k - 1; i >= 0; --i) Heap.dispose(strs[i]);
+      free_memory(cStr);
+      free_memory(cArr);
+      for (i = k - 1; i >= 0; --i) free_memory(strs[i]);
     }
   }
 
@@ -503,8 +503,8 @@ export class Module {
     try {
       return _BinaryenCall(this.ref, cStr, cArr, operands && operands.length || 0, returnType);
     } finally {
-      Heap.dispose(cArr);
-      Heap.dispose(cStr);
+      free_memory(cArr);
+      free_memory(cStr);
     }
   }
 
@@ -515,8 +515,8 @@ export class Module {
     try {
       return _BinaryenCallImport(this.ref, cStr, cArr, operands && operands.length || 0, returnType);
     } finally {
-      Heap.dispose(cArr);
-      Heap.dispose(cStr);
+      free_memory(cArr);
+      free_memory(cStr);
     }
   }
 
@@ -533,7 +533,7 @@ export class Module {
     try {
       return _BinaryenAddGlobal(this.ref, cStr, type, mutable ? 1 : 0, initializer);
     } finally {
-      Heap.dispose(cStr);
+      free_memory(cStr);
     }
   }
 
@@ -544,8 +544,8 @@ export class Module {
     try {
       return _BinaryenAddFunction(this.ref, cStr, type, cArr, varTypes.length, body);
     } finally {
-      Heap.dispose(cArr);
-      Heap.dispose(cStr);
+      free_memory(cArr);
+      free_memory(cStr);
     }
   }
 
@@ -554,7 +554,7 @@ export class Module {
     try {
       _BinaryenRemoveFunction(this.ref, cStr);
     } finally {
-      Heap.dispose(cStr);
+      free_memory(cStr);
     }
   }
 
@@ -565,8 +565,8 @@ export class Module {
     try {
       return _BinaryenAddFunctionExport(this.ref, cStr1, cStr2);
     } finally {
-      Heap.dispose(cStr2);
-      Heap.dispose(cStr1);
+      free_memory(cStr2);
+      free_memory(cStr1);
     }
   }
 
@@ -577,8 +577,8 @@ export class Module {
     try {
       return _BinaryenAddTableExport(this.ref, cStr1, cStr2);
     } finally {
-      Heap.dispose(cStr2);
-      Heap.dispose(cStr1);
+      free_memory(cStr2);
+      free_memory(cStr1);
     }
   }
 
@@ -589,8 +589,8 @@ export class Module {
     try {
       return _BinaryenAddMemoryExport(this.ref, cStr1, cStr2);
     } finally {
-      Heap.dispose(cStr2);
-      Heap.dispose(cStr1);
+      free_memory(cStr2);
+      free_memory(cStr1);
     }
   }
 
@@ -601,8 +601,8 @@ export class Module {
     try {
       return _BinaryenAddGlobalExport(this.ref, cStr1, cStr2);
     } finally {
-      Heap.dispose(cStr2);
-      Heap.dispose(cStr1);
+      free_memory(cStr2);
+      free_memory(cStr1);
     }
   }
 
@@ -612,7 +612,7 @@ export class Module {
     try {
       _BinaryenRemoveExport(this.ref, cStr);
     } finally {
-      Heap.dispose(cStr);
+      free_memory(cStr);
     }
   }
 
@@ -624,9 +624,9 @@ export class Module {
     try {
       return _BinaryenAddFunctionImport(this.ref, cStr1, cStr2, cStr3, functionType);
     } finally {
-      Heap.dispose(cStr3);
-      Heap.dispose(cStr2);
-      Heap.dispose(cStr1);
+      free_memory(cStr3);
+      free_memory(cStr2);
+      free_memory(cStr1);
     }
   }
 
@@ -638,9 +638,9 @@ export class Module {
     try {
       return _BinaryenAddTableImport(this.ref, cStr1, cStr2, cStr3);
     } finally {
-      Heap.dispose(cStr3);
-      Heap.dispose(cStr2);
-      Heap.dispose(cStr1);
+      free_memory(cStr3);
+      free_memory(cStr2);
+      free_memory(cStr1);
     }
   }
 
@@ -652,9 +652,9 @@ export class Module {
     try {
       return _BinaryenAddMemoryImport(this.ref, cStr1, cStr2, cStr3);
     } finally {
-      Heap.dispose(cStr3);
-      Heap.dispose(cStr2);
-      Heap.dispose(cStr1);
+      free_memory(cStr3);
+      free_memory(cStr2);
+      free_memory(cStr1);
     }
   }
 
@@ -666,9 +666,9 @@ export class Module {
     try {
       return _BinaryenAddGlobalImport(this.ref, cStr1, cStr2, cStr3, globalType);
     } finally {
-      Heap.dispose(cStr3);
-      Heap.dispose(cStr2);
-      Heap.dispose(cStr1);
+      free_memory(cStr3);
+      free_memory(cStr2);
+      free_memory(cStr1);
     }
   }
 
@@ -678,7 +678,7 @@ export class Module {
     try {
       _BinaryenRemoveImport(this.ref, cStr);
     } finally {
-      Heap.dispose(cStr);
+      free_memory(cStr);
     }
   }
 
@@ -704,11 +704,11 @@ export class Module {
     try {
       _BinaryenSetMemory(this.ref, initial, maximum, cStr, cArr1, cArr2, cArr3, k);
     } finally {
-      Heap.dispose(cArr3);
-      Heap.dispose(cArr2);
-      Heap.dispose(cArr1);
-      for (i = k - 1; i >= 0; --i) Heap.dispose(segs[i]);
-      Heap.dispose(cStr);
+      free_memory(cArr3);
+      free_memory(cArr2);
+      free_memory(cArr1);
+      for (i = k - 1; i >= 0; --i) free_memory(segs[i]);
+      free_memory(cStr);
     }
   }
 
@@ -718,7 +718,7 @@ export class Module {
     try {
       _BinaryenSetFunctionTable(this.ref, cArr, funcs.length);
     } finally {
-      Heap.dispose(cArr);
+      free_memory(cArr);
     }
   }
 
@@ -749,8 +749,8 @@ export class Module {
       else
         _BinaryenModuleRunPasses(this.ref, cArr, k);
     } finally {
-      Heap.dispose(cArr);
-      for (; i >= 0; --i) Heap.dispose(names[i]);
+      free_memory(cArr);
+      for (; i >= 0; --i) free_memory(names[i]);
     }
   }
 
@@ -788,7 +788,7 @@ export class Module {
   dispose(): void {
     if (!this.ref) return; // sic
     _BinaryenModuleDispose(this.ref);
-    Heap.dispose(changetype<usize>(this.lit));
+    free_memory(changetype<usize>(this.lit));
   }
 
   createRelooper(): Relooper {
@@ -890,7 +890,7 @@ export class Relooper {
     try {
       _RelooperAddBranchForSwitch(from, to, cArr, indexes.length, code);
     } finally {
-      Heap.dispose(cArr);
+      free_memory(cArr);
     }
   }
 
@@ -905,7 +905,7 @@ export class Relooper {
 
 function allocU8Array(u8s: Uint8Array | null): usize {
   if (!u8s) return 0;
-  var ptr = Heap.allocate(u8s.length);
+  var ptr = allocate_memory(u8s.length);
   var idx = ptr;
   for (var i = 0, k = u8s.length; i < k; ++i)
     store<u8>(idx++, u8s[i]);
@@ -914,7 +914,7 @@ function allocU8Array(u8s: Uint8Array | null): usize {
 
 function allocI32Array(i32s: i32[] | null): usize {
   if (!i32s) return 0;
-  var ptr = Heap.allocate(i32s.length << 2);
+  var ptr = allocate_memory(i32s.length << 2);
   var idx = ptr;
   for (var i = 0, k = i32s.length; i < k; ++i) {
     var val = i32s[i];
@@ -952,7 +952,7 @@ function stringLengthUTF8(str: string): usize {
 
 function allocString(str: string | null): usize {
   if (str == null) return 0;
-  var ptr = Heap.allocate(stringLengthUTF8(str) + 1);
+  var ptr = allocate_memory(stringLengthUTF8(str) + 1);
   var idx = ptr;
   for (var i = 0, k = str.length; i < k; ++i) {
     var u = str.charCodeAt(i);
diff --git a/std/assembly.d.ts b/std/assembly.d.ts
index da856b9a..55fd5310 100644
--- a/std/assembly.d.ts
+++ b/std/assembly.d.ts
@@ -162,6 +162,16 @@ declare function store<T>(offset: usize, value: T): void;
 declare function current_memory(): i32;
 /** Grows linear memory by a given unsigned delta of pages. One page is 64kb. Returns the previous memory size in units of pages or `-1` on failure. */
 declare function grow_memory(value: i32): i32;
+/** Copies n bytes from the specified source to the specified destination in memory. These regions may overlap. */
+declare function move_memory(destination: usize, source: usize, n: usize): void;
+/** Sets n bytes beginning at the specified destination in memory to the specified byte value. */
+declare function set_memory(destination: usize, value: u8, count: usize): void;
+/** Compares two chunks of memory. Returns `0` if equal, otherwise the difference of the first differing bytes. */
+declare function compare_memory(vl: usize, vr: usize, n: usize): i32;
+/** Allocates a chunk of memory of the specified size and returns a pointer to it. */
+declare function allocate_memory(size: usize): usize;
+/** Disposes a chunk of memory by its pointer. */
+declare function free_memory(ptr: usize): void;
 /** Emits an unreachable operation that results in a runtime error when executed. Both a statement and an expression of any type. */
 declare function unreachable(): any; // sic
 
@@ -250,36 +260,6 @@ declare class Error {
 /** Class for indicating an error when a value is not in the set or range of allowed values. */
 declare class RangeError extends Error { }
 
-/** A static class representing the heap. */
-declare class Heap {
-
-  /** Gets the amount of used heap space, in bytes. */
-  static readonly used: usize;
-
-  /** Gets the amount of free heap space, in bytes. */
-  static readonly free: usize;
-
-  /** Gets the size of the heap, in bytes. */
-  static readonly size: usize;
-
-  /** Allocates a chunk of memory and returns a pointer to it. */
-  static allocate(size: usize): usize;
-
-  /** Disposes a chunk of memory by its pointer. */
-  static dispose(ptr: usize): void;
-
-  /** Copies a chunk of memory from one location to another. */
-  static copy(dest: usize, src: usize, n: usize): usize;
-
-  /** Fills a chunk of memory with the specified byte value. */
-  static fill(dest: usize, c: u8, n: usize): usize;
-
-  /** Compares two chunks of memory. Returns `0` if equal, otherwise the difference of the first differing bytes. */
-  static compare(vl: usize, vr: usize, n: usize): i32;
-
-  private constructor();
-}
-
 interface Boolean {}
 interface Function {}
 interface IArguments {}
diff --git a/std/assembly/array.ts b/std/assembly/array.ts
index d62d57de..166b74e4 100644
--- a/std/assembly/array.ts
+++ b/std/assembly/array.ts
@@ -8,7 +8,7 @@ export class Array<T> {
     if (capacity < 0)
       throw new RangeError("invalid array length");
     this.__capacity = this.length = capacity;
-    this.__memory = capacity > 0 ? Heap.allocate(<usize>capacity * sizeof<T>()) : 0;
+    this.__memory = capacity > 0 ? allocate_memory(<usize>capacity * sizeof<T>()) : 0;
   }
 
   @operator("[]")
@@ -34,19 +34,20 @@ export class Array<T> {
     return -1;
   }
 
-  private __grow(capacity: i32): void {
-    assert(capacity > this.__capacity);
-    var newMemory = Heap.allocate(<usize>(capacity * sizeof<T>()));
-    if (this.__memory)
-      Heap.copy(newMemory, this.__memory, this.__capacity * sizeof<T>());
-    Heap.dispose(this.__memory);
+  private __grow(newCapacity: i32): void {
+    assert(newCapacity > this.__capacity);
+    var newMemory = allocate_memory(<usize>newCapacity * sizeof<T>());
+    if (this.__memory) {
+      move_memory(newMemory, this.__memory, this.__capacity * sizeof<T>());
+      free_memory(this.__memory);
+    }
     this.__memory = newMemory;
-    this.__capacity = capacity;
+    this.__capacity = newCapacity;
   }
 
   push(element: T): i32 {
     if (<u32>this.length >= this.__capacity)
-      this.__grow(max(this.length + 1, this.__capacity * 2));
+      this.__grow(max(this.length + 1, this.__capacity << 1));
     store<T>(this.__memory + <usize>this.length * sizeof<T>(), element);
     return ++this.length;
   }
@@ -62,23 +63,22 @@ export class Array<T> {
     if (this.length < 1 || <u32>this.length > this.__capacity)
       throw new RangeError("index out of range");
     var element = load<T>(this.__memory);
-    Heap.copy(this.__memory, this.__memory + sizeof<T>(), (this.__capacity - 1) * sizeof<T>());
-    Heap.fill(this.__memory + (this.__capacity - 1) * sizeof<T>(), 0, sizeof<T>());
+    move_memory(this.__memory, this.__memory + sizeof<T>(), (this.__capacity - 1) * sizeof<T>());
+    set_memory(this.__memory + (this.__capacity - 1) * sizeof<T>(), 0, sizeof<T>());
     --this.length;
     return element;
   }
 
   unshift(element: T): i32 {
-    var capacity = this.__capacity;
-    if (<u32>this.length >= capacity)
-      this.__grow(max(this.length + 1, capacity * 2));
+    var oldCapacity = this.__capacity;
+    if (<u32>this.length >= oldCapacity)
+      this.__grow(max(this.length + 1, oldCapacity * 2));
 
-    // FIXME: needs memmove (Heap.copy is just memcpy). it's also inefficient because
-    // __grow copies and then unshift copies again.
-    // Heap.copy(this.__memory + sizeof<T>(), this.__memory, capacity * sizeof<T>());
+    // FIXME: this is inefficient because of two copies, one in __grow and one here
+    // move_memory(this.__memory + sizeof<T>(), this.__memory, oldCapacity * sizeof<T>());
 
-    if (capacity)
-      for (var index: usize = capacity; index > 0; --index)
+    if (oldCapacity)
+      for (var index: usize = oldCapacity; index > 0; --index)
         store<T>(this.__memory + index * sizeof<T>(), load<T>(this.__memory + (index - 1) * sizeof<T>()));
 
     store<T>(this.__memory, element);
@@ -92,12 +92,16 @@ export class CArray<T> {
   private constructor() {}
 
   @operator("[]")
-  private __get(index: usize): T {
-    return load<T>(changetype<usize>(this) + index * sizeof<T>());
+  private __get(index: i32): T {
+    if (index < 0)
+      throw new RangeError("index out of range");
+    return load<T>(changetype<usize>(this) + <usize>index * sizeof<T>());
   }
 
   @operator("[]=")
-  private __set(index: usize, value: T): void {
-    store<T>(changetype<usize>(this) + index * sizeof<T>(), value);
+  private __set(index: i32, value: T): void {
+    if (index < 0)
+      throw new RangeError("index out of range");
+    store<T>(changetype<usize>(this) + <usize>index * sizeof<T>(), value);
   }
 }
diff --git a/std/assembly/heap.ts b/std/assembly/heap.ts
index 9c5bab91..1a865ea6 100644
--- a/std/assembly/heap.ts
+++ b/std/assembly/heap.ts
@@ -4,245 +4,289 @@ const ALIGN_MASK: usize = ALIGN_SIZE - 1;
 
 var HEAP_OFFSET: usize = HEAP_BASE; // HEAP_BASE is a constant generated by the compiler
 
-// TODO: maybe tlsf
+export function allocate_memory(size: usize): usize {
+  if (!size) return 0;
+  var len: i32 = current_memory();
+  if (HEAP_OFFSET + size > <usize>len << 16)
+    if(grow_memory(max<i32>(<i32>ceil<f64>(<f64>size / 65536), len * 2 - len)) < 0)
+      unreachable();
+  var ptr: usize = HEAP_OFFSET;
+  if ((HEAP_OFFSET += size) & ALIGN_MASK) // align next offset
+    HEAP_OFFSET = (HEAP_OFFSET | ALIGN_MASK) + 1;
+  return ptr;
+}
 
-export class Heap {
+export function free_memory(ptr: usize): void {
+  // just a big chunk of non-disposable memory for now
+}
 
-  static get used(): usize { return HEAP_OFFSET - HEAP_BASE; }
-  static get free(): usize { return (<usize>current_memory() << 16) - HEAP_OFFSET; }
-  static get size(): usize { return (<usize>current_memory() << 16) - HEAP_BASE; }
+function copy_memory(dest: usize, src: usize, n: usize): void {
+  // based on musl's implementation of memcpy
+  // not a future instruction and sufficiently covered by the upcoming move_memory intrinsic
 
-  static allocate(size: usize): usize {
-    if (!size) return 0;
-    var len: i32 = current_memory();
-    if (HEAP_OFFSET + size > <usize>len << 16)
-      if(grow_memory(max<i32>(<i32>ceil<f64>(<f64>size / 65536), len * 2 - len)) < 0)
-        unreachable();
-    var ptr: usize = HEAP_OFFSET;
-    if ((HEAP_OFFSET += size) & ALIGN_MASK) // align next offset
-      HEAP_OFFSET = (HEAP_OFFSET | ALIGN_MASK) + 1;
-    return ptr;
+  var w: u32, x: u32;
+
+  // copy 1 byte each until src is aligned to 4 bytes
+  while (n && src % 4) {
+    store<u8>(dest++, load<u8>(src++));
+    n--;
   }
 
-  static dispose(ptr: usize): void {
-    // just a big chunk of non-disposable memory for now
-  }
-
-  static copy(dest: usize, src: usize, n: usize): usize { // TODO: use move_memory op once available
-    assert(dest >= HEAP_BASE);
-
-    // the following is based on musl's implementation of memcpy
-    var dst: usize = dest;
-    var w: u32, x: u32;
-
-    // copy 1 byte each until src is aligned to 4 bytes
-    while (n && src % 4) {
-      store<u8>(dst++, load<u8>(src++));
-      n--;
-    }
-
-    // if dst is aligned to 4 bytes as well, copy 4 bytes each
-    if (dst % 4 == 0) {
-      while (n >= 16) {
-        store<u32>(dst     , load<u32>(src     ));
-        store<u32>(dst +  4, load<u32>(src +  4));
-        store<u32>(dst +  8, load<u32>(src +  8));
-        store<u32>(dst + 12, load<u32>(src + 12));
-        src += 16; dst += 16; n -= 16;
-      }
-      if (n & 8) {
-        store<u32>(dst    , load<u32>(src    ));
-        store<u32>(dst + 4, load<u32>(src + 4));
-        dst += 8; src += 8;
-      }
-      if (n & 4) {
-        store<u32>(dst, load<u32>(src));
-        dst += 4; src += 4;
-      }
-      if (n & 2) { // drop to 2 bytes each
-        store<u16>(dst, load<u16>(src));
-        dst += 2; src += 2;
-      }
-      if (n & 1) { // drop to 1 byte
-        store<u8>(dst++, load<u8>(src++));
-      }
-      return dest;
-    }
-
-    // if dst is not aligned to 4 bytes, use alternating shifts to copy 4 bytes each
-    // doing shifts if faster when copying enough bytes (here: 32 or more)
-    if (n >= 32) {
-      switch (dst % 4) {
-        // known to be != 0
-        case 1:
-          w = load<u32>(src);
-          store<u8>(dst++, load<u8>(src++));
-          store<u8>(dst++, load<u8>(src++));
-          store<u8>(dst++, load<u8>(src++));
-          n -= 3;
-          while (n >= 17) {
-            x = load<u32>(src + 1);
-            store<u32>(dst, w >> 24 | x << 8);
-            w = load<u32>(src + 5);
-            store<u32>(dst + 4, x >> 24 | w << 8);
-            x = load<u32>(src + 9);
-            store<u32>(dst + 8, w >> 24 | x << 8);
-            w = load<u32>(src + 13);
-            store<u32>(dst + 12, x >> 24 | w << 8);
-            src += 16; dst += 16; n -= 16;
-          }
-          break;
-        case 2:
-          w = load<u32>(src);
-          store<u8>(dst++, load<u8>(src++));
-          store<u8>(dst++, load<u8>(src++));
-          n -= 2;
-          while (n >= 18) {
-            x = load<u32>(src + 2);
-            store<u32>(dst, w >> 16 | x << 16);
-            w = load<u32>(src + 6);
-            store<u32>(dst + 4, x >> 16 | w << 16);
-            x = load<u32>(src + 10);
-            store<u32>(dst + 8, w >> 16 | x << 16);
-            w = load<u32>(src + 14);
-            store<u32>(dst + 12, x >> 16 | w << 16);
-            src += 16; dst += 16; n -= 16;
-          }
-          break;
-        case 3:
-          w = load<u32>(src);
-          store<u8>(dst++, load<u8>(src++));
-          n -= 1;
-          while (n >= 19) {
-            x = load<u32>(src + 3);
-            store<u32>(dst, w >> 8 | x << 24);
-            w = load<u32>(src + 7);
-            store<u32>(dst + 4, x >> 8 | w << 24);
-            x = load<u32>(src + 11);
-            store<u32>(dst + 8, w >> 8 | x << 24);
-            w = load<u32>(src + 15);
-            store<u32>(dst + 12, x >> 8 | w << 24);
-            src += 16; dst += 16; n -= 16;
-          }
-          break;
-      }
-    }
-
-    // copy remaining bytes one by one
-    if (n & 16) {
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
+  // if dst is aligned to 4 bytes as well, copy 4 bytes each
+  if (dest % 4 == 0) {
+    while (n >= 16) {
+      store<u32>(dest     , load<u32>(src     ));
+      store<u32>(dest +  4, load<u32>(src +  4));
+      store<u32>(dest +  8, load<u32>(src +  8));
+      store<u32>(dest + 12, load<u32>(src + 12));
+      src += 16; dest += 16; n -= 16;
     }
     if (n & 8) {
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
+      store<u32>(dest    , load<u32>(src    ));
+      store<u32>(dest + 4, load<u32>(src + 4));
+      dest += 8; src += 8;
     }
     if (n & 4) {
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
+      store<u32>(dest, load<u32>(src));
+      dest += 4; src += 4;
     }
-    if (n & 2) {
-      store<u8>(dst++, load<u8>(src++));
-      store<u8>(dst++, load<u8>(src++));
+    if (n & 2) { // drop to 2 bytes each
+      store<u16>(dest, load<u16>(src));
+      dest += 2; src += 2;
     }
-    if (n & 1) {
-      store<u8>(dst++, load<u8>(src++));
+    if (n & 1) { // drop to 1 byte
+      store<u8>(dest++, load<u8>(src++));
     }
-    return dest;
+    return;
   }
 
-  static fill(dest: usize, c: u8, n: usize): usize { // TODO: use set_memory op once available
-    assert(dest >= HEAP_BASE);
-
-    // the following is based on musl's implementation of memset
-    if (!n) return dest;
-
-    var s: usize = dest;
-
-    // Fill head and tail with minimal branching
-    store<u8>(s, c); store<u8>(s + n - 1, c);
-    if (n <= 2) return dest;
-    store<u8>(s + 1, c); store<u8>(s + n - 2, c);
-    store<u8>(s + 2, c); store<u8>(s + n - 3, c);
-    if (n <= 6) return dest;
-    store<u8>(s + 3, c); store<u8>(s + n - 4, c);
-    if (n <= 8) return dest;
-
-    // Align to 4 bytes
-    var k: usize = -s & 3;
-    s += k;
-    n -= k;
-    n &= -4;
-
-    var c32: u32 = -1 / 255 * c;
-
-    // Fill head and tail in preparation of setting 32 bytes at a time
-    store<u32>(s, c32);
-    store<u32>(s + n - 4, c32);
-    if (n <= 8) return dest;
-    store<u32>(s + 4, c32);
-    store<u32>(s + 8, c32);
-    store<u32>(s + n - 12, c32);
-    store<u32>(s + n - 8, c32);
-    if (n <= 24) return dest;
-    store<u32>(s + 12, c32);
-    store<u32>(s + 16, c32);
-    store<u32>(s + 20, c32);
-    store<u32>(s + 24, c32);
-    store<u32>(s + n - 28, c32);
-    store<u32>(s + n - 24, c32);
-    store<u32>(s + n - 20, c32);
-    store<u32>(s + n - 16, c32);
-
-    // Align to 8 bytes
-    k = 24 + (s & 4);
-    s += k;
-    n -= k;
-
-    // Set 32 bytes at a time
-    var c64: u64 = <u64>c32 | (<u64>c32 << 32);
-    while (n >= 32) {
-      store<u64>(s, c64);
-      store<u64>(s + 8, c64);
-      store<u64>(s + 16, c64);
-      store<u64>(s + 24, c64);
-      n -= 32; s += 32;
+  // if dst is not aligned to 4 bytes, use alternating shifts to copy 4 bytes each
+  // doing shifts if faster when copying enough bytes (here: 32 or more)
+  if (n >= 32) {
+    switch (dest % 4) {
+      // known to be != 0
+      case 1:
+        w = load<u32>(src);
+        store<u8>(dest++, load<u8>(src++));
+        store<u8>(dest++, load<u8>(src++));
+        store<u8>(dest++, load<u8>(src++));
+        n -= 3;
+        while (n >= 17) {
+          x = load<u32>(src + 1);
+          store<u32>(dest, w >> 24 | x << 8);
+          w = load<u32>(src + 5);
+          store<u32>(dest + 4, x >> 24 | w << 8);
+          x = load<u32>(src + 9);
+          store<u32>(dest + 8, w >> 24 | x << 8);
+          w = load<u32>(src + 13);
+          store<u32>(dest + 12, x >> 24 | w << 8);
+          src += 16; dest += 16; n -= 16;
+        }
+        break;
+      case 2:
+        w = load<u32>(src);
+        store<u8>(dest++, load<u8>(src++));
+        store<u8>(dest++, load<u8>(src++));
+        n -= 2;
+        while (n >= 18) {
+          x = load<u32>(src + 2);
+          store<u32>(dest, w >> 16 | x << 16);
+          w = load<u32>(src + 6);
+          store<u32>(dest + 4, x >> 16 | w << 16);
+          x = load<u32>(src + 10);
+          store<u32>(dest + 8, w >> 16 | x << 16);
+          w = load<u32>(src + 14);
+          store<u32>(dest + 12, x >> 16 | w << 16);
+          src += 16; dest += 16; n -= 16;
+        }
+        break;
+      case 3:
+        w = load<u32>(src);
+        store<u8>(dest++, load<u8>(src++));
+        n -= 1;
+        while (n >= 19) {
+          x = load<u32>(src + 3);
+          store<u32>(dest, w >> 8 | x << 24);
+          w = load<u32>(src + 7);
+          store<u32>(dest + 4, x >> 8 | w << 24);
+          x = load<u32>(src + 11);
+          store<u32>(dest + 8, w >> 8 | x << 24);
+          w = load<u32>(src + 15);
+          store<u32>(dest + 12, x >> 8 | w << 24);
+          src += 16; dest += 16; n -= 16;
+        }
+        break;
     }
-
-    return dest;
   }
 
-  static compare(vl: usize, vr: usize, n: usize): i32 {
-    if (vl == vr) return 0;
-
-    // the following is based on musl's implementation of memcmp
-    while (n && load<u8>(vl) == load<u8>(vr)) {
-      n--; vl++; vr++;
-    }
-    return n ? <i32>load<u8>(vl) - <i32>load<u8>(vr) : 0;
+  // copy remaining bytes one by one
+  if (n & 16) {
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+  }
+  if (n & 8) {
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+  }
+  if (n & 4) {
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+  }
+  if (n & 2) {
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+  }
+  if (n & 1) {
+    store<u8>(dest++, load<u8>(src++));
   }
-
-  private constructor() {}
+}
+
+export function move_memory(dest: usize, src: usize, n: usize): void {
+  // based on musl's implementation of memmove
+  // becomes obsolete once https://github.com/WebAssembly/bulk-memory-operations lands
+
+  if (dest == src)
+    return;
+  if (src + n <= dest || dest + n <= src) {
+    copy_memory(dest, src, n);
+    return;
+  }
+  if (dest < src) {
+    if (src % 8 == dest % 8) {
+      while (dest % 8) {
+        if (!n)
+          return;
+        --n;
+        store<u8>(dest++, load<u8>(src++));
+      }
+      while (n >= 8) {
+        store<u64>(dest, load<u64>(src));
+        n -= 8;
+        dest += 8;
+        src += 8;
+      }
+    }
+    while (n) {
+      store<u8>(dest++, load<u8>(src++));
+      --n;
+    }
+  } else {
+    if (src % 8 == dest % 8) {
+      while ((dest + n) % 8) {
+        if (!n)
+          return;
+        store<u8>(dest + --n, load<u8>(src + n));
+      }
+      while (n >= 8) {
+        n -= 8;
+        store<u64>(dest + n, load<u64>(src + n));
+      }
+    }
+    while (n) {
+      store<u8>(dest + --n, load<u8>(src + n));
+    }
+  }
+}
+
+export function set_memory(dest: usize, c: u8, n: usize): void {
+  // based on musl's implementation of memset
+  // becomes obsolete once https://github.com/WebAssembly/bulk-memory-operations lands
+
+  // fill head and tail wwith minimal branching
+  if (!n)
+    return;
+  store<u8>(dest, c);
+  store<u8>(dest + n - 1, c);
+  if (n <= 2)
+    return;
+
+  store<u8>(dest + 1, c);
+  store<u8>(dest + 2, c);
+  store<u8>(dest + n - 2, c);
+  store<u8>(dest + n - 3, c);
+  if (n <= 6)
+    return;
+  store<u8>(dest + 3, c);
+  store<u8>(dest + n - 4, c);
+  if (n <= 8)
+    return;
+
+  // advance pointer to align it at 4-byte boundary
+  var k: usize = -dest & 3;
+  dest += k;
+  n -= k;
+  n &= -4;
+
+  var c32: u32 = -1 / 255 * c;
+
+  // fill head/tail up to 28 bytes each in preparation
+  store<u32>(dest, c32);
+  store<u32>(dest + n - 4, c32);
+  if (n <= 8)
+    return;
+  store<u32>(dest + 4, c32);
+  store<u32>(dest + 8, c32);
+  store<u32>(dest + n - 12, c32);
+  store<u32>(dest + n - 8, c32);
+  if (n <= 24)
+    return;
+  store<u32>(dest + 12, c32);
+  store<u32>(dest + 16, c32);
+  store<u32>(dest + 20, c32);
+  store<u32>(dest + 24, c32);
+  store<u32>(dest + n - 28, c32);
+  store<u32>(dest + n - 24, c32);
+  store<u32>(dest + n - 20, c32);
+  store<u32>(dest + n - 16, c32);
+
+  // align to a multiple of 8
+  k = 24 + (dest & 4);
+  dest += k;
+  n -= k;
+
+  // copy 32 bytes each
+  var c64: u64 = <u64>c32 | (<u64>c32 << 32);
+  while (n >= 32) {
+    store<u64>(dest, c64);
+    store<u64>(dest + 8, c64);
+    store<u64>(dest + 16, c64);
+    store<u64>(dest + 24, c64);
+    n -= 32;
+    dest += 32;
+  }
+}
+
+export function compare_memory(vl: usize, vr: usize, n: usize): i32 {
+  // based on musl's implementation of memcmp
+  // provided because there's no proposed alternative
+  if (vl == vr)
+    return 0;
+  while (n && load<u8>(vl) == load<u8>(vr)) {
+    n--;
+    vl++;
+    vr++;
+  }
+  return n ? <i32>load<u8>(vl) - <i32>load<u8>(vr) : 0;
 }
diff --git a/std/assembly/string.ts b/std/assembly/string.ts
index 1c32f127..0f732cdb 100644
--- a/std/assembly/string.ts
+++ b/std/assembly/string.ts
@@ -1,38 +1,37 @@
 const EMPTY: String = changetype<String>("");
 
 export class String {
-  // [key: number]: string;
-
-  private ptr: usize;
 
+  private __memory: usize;
   readonly length: i32;
 
-  constructor(ptr: usize, lenght: i32) {
-    this.ptr = ptr;
-    this.length = lenght;
+  constructor(ptr: usize, length: i32) {
+    if (length < 0)
+      throw new RangeError("invalid length");
+    this.__memory = ptr;
+    this.length = length;
   }
 
   @operator("[]")
   charAt(pos: i32): String {
-    assert(this != null);
-    return pos < 0 || pos >= this.length ? EMPTY
-         : new String(this.ptr + (<usize>pos << 1), 1);
+    if (<u32>pos >= this.length)
+      return changetype<String>("");
+    return new String(this.__memory + (<usize>pos << 1), 1);
   }
 
   charCodeAt(pos: i32): i32 {
-    assert(this != null);
-    return pos < 0 || pos >= this.length ? -1 // NaN
-         : load<u16>(this.ptr + (<usize>pos << 1));
+    if (<u32>pos >= this.length)
+      return -1; // NaN
+    return load<u16>(this.__memory + (<usize>pos << 1));
   }
 
   codePointAt(pos: i32): i32 {
-    assert(this != null);
-    if (pos < 0 || pos >= this.length)
+    if (<u32>pos >= this.length)
       return -1; // undefined
-    var first = <i32>load<u16>(this.ptr + (<usize>pos << 1));
+    var first = <i32>load<u16>(this.__memory + (<usize>pos << 1));
     if (first < 0xD800 || first > 0xDBFF || pos + 1 == this.length)
       return first;
-    var second = <i32>load<u16>(this.ptr + ((<usize>pos + 1) << 1));
+    var second = <i32>load<u16>(this.__memory + ((<usize>pos + 1) << 1));
     if (second < 0xDC00 || second > 0xDFFF)
       return first;
     return ((first - 0xD800) << 10) + (second - 0xDC00) + 0x10000;
@@ -45,37 +44,31 @@ export class String {
     var thisLen: isize = this.length;
     var otherLen: isize = other.length;
     var len: usize = thisLen + otherLen;
-    return new String(
-      Heap.copy(
-        Heap.copy(
-          Heap.allocate(len << 1),
-          this.ptr,
-          thisLen << 1
-        ) + (thisLen << 1),
-        other.ptr,
-        otherLen << 1
-      ),
-      <i32>len
-    );
+    var newMemory = allocate_memory(len << 1);
+    move_memory(newMemory, this.__memory, thisLen << 1);
+    move_memory(newMemory + (thisLen << 1), other.__memory, otherLen << 1);
+    return new String(newMemory, <i32>len);
   }
 
   endsWith(searchString: this, endPosition: i32 = 0x7fffffff): bool {
-    assert(this != null);
     assert(searchString != null);
     var end: isize = <isize>min<i32>(max<i32>(endPosition, 0), this.length);
     var searchLength: isize = searchString.length;
     var start: isize = end - searchLength;
     if (start < 0)
       return false;
-    return !Heap.compare(this.ptr + (start << 1), searchString.ptr, searchLength << 1);
+    return !compare_memory(this.__memory + (start << 1), searchString.__memory, searchLength << 1);
   }
 
   @operator("==")
   private __eq(other: this): bool {
-    assert(this != null);
-    assert(other != null);
-    return this.length != other.length ? false
-         : !Heap.compare(this.ptr, other.ptr, <usize>this.length);
+    if (this == null)
+      return other == null;
+    else if (other == null)
+      return false;
+    if (this.length != other.length)
+      return false;
+    return !compare_memory(this.__memory, other.__memory, <usize>this.length);
   }
 
   includes(searchString: this, position: i32 = 0): bool {
@@ -83,14 +76,13 @@ export class String {
   }
 
   indexOf(searchString: this, position: i32 = 0): i32 {
-    assert(this != null);
     assert(searchString != null);
     var pos: isize = position;
     var len: isize = this.length;
     var start: isize = min<isize>(max<isize>(pos, 0), len);
     var searchLen: isize = searchString.length;
     for (var k: usize = start; <isize>k + searchLen <= len; ++k)
-      if (!Heap.compare(this.ptr + (k << 1), searchString.ptr, searchLen << 1))
+      if (!compare_memory(this.__memory + (k << 1), searchString.__memory, searchLen << 1))
         return <i32>k;
     return -1;
   }
@@ -104,7 +96,7 @@ export class String {
     var searchLength: isize = searchString.length;
     if (searchLength + start > len)
       return false;
-    return !Heap.compare(this.ptr + (start << 1), searchString.ptr, searchLength << 1);
+    return !compare_memory(this.__memory + (start << 1), searchString.__memory, searchLength << 1);
   }
 
   substr(start: i32, length: i32 = i32.MAX_VALUE): String {
@@ -117,7 +109,7 @@ export class String {
     var resultLength: isize = min<isize>(max<isize>(end, 0), size - intStart);
     if (resultLength < 0)
       return EMPTY;
-    return new String(this.ptr + (intStart << 1), <i32>resultLength);
+    return new String(this.__memory + (intStart << 1), <i32>resultLength);
   }
 
   substring(start: i32, end: i32 = i32.MAX_VALUE): String {
@@ -132,46 +124,46 @@ export class String {
       return EMPTY;
     if (!from && to == this.length)
       return this;
-    return new String(this.ptr + (from << 1), len);
+    return new String(this.__memory + (from << 1), len);
   }
 
   trim(): String {
     assert(this != null);
     var length: usize = this.length;
-    while (length && isWhiteSpaceOrLineTerminator(load<u16>(this.ptr + (length << 1))))
+    while (length && isWhiteSpaceOrLineTerminator(load<u16>(this.__memory + (length << 1))))
       --length;
     var start: usize = 0;
-    while (start < length && isWhiteSpaceOrLineTerminator(load<u16>(this.ptr + (start << 1)))) {
+    while (start < length && isWhiteSpaceOrLineTerminator(load<u16>(this.__memory + (start << 1)))) {
       ++start; --length;
     }
     if (!length)
       return EMPTY;
     if (!start && length == this.length)
       return this;
-    return new String(this.ptr + (start << 1), length);
+    return new String(this.__memory + (start << 1), length);
   }
 
   trimLeft(): String {
     assert(this != null);
     var start: isize = 0;
     var len: isize = this.length;
-    while (start < len && isWhiteSpaceOrLineTerminator(load<u16>(this.ptr + (start << 1))))
+    while (start < len && isWhiteSpaceOrLineTerminator(load<u16>(this.__memory + (start << 1))))
       ++start;
     if (!start)
       return this;
-    return new String(this.ptr + (start << 1), <i32>(len - start));
+    return new String(this.__memory + (start << 1), <i32>(len - start));
   }
 
   trimRight(): String {
     assert(this != null);
     var len: isize = this.length;
-    while (len > 0 && isWhiteSpaceOrLineTerminator(load<u16>(this.ptr + (len << 1))))
+    while (len > 0 && isWhiteSpaceOrLineTerminator(load<u16>(this.__memory + (len << 1))))
       --len;
     if (len <= 0)
       return EMPTY;
     if (<i32>len == this.length)
       return this;
-    return new String(this.ptr, <i32>len);
+    return new String(this.__memory, <i32>len);
   }
 }
 
diff --git a/std/portable.d.ts b/std/portable.d.ts
index e24e56b1..ce0787aa 100644
--- a/std/portable.d.ts
+++ b/std/portable.d.ts
@@ -116,6 +116,12 @@ declare function select<T>(ifTrue: T, ifFalse: T, condition: bool): T;
 declare function sqrt<T = f32 | f64>(value: T): T;
 /** Rounds to the nearest integer towards zero of a 32-bit or 64-bit float. */
 declare function trunc<T = f32 | f64>(value: T): T;
+/** Allocates a chunk of memory of the specified size and returns a pointer to it. */
+declare function allocate_memory(size: usize): usize;
+/** Disposes a chunk of memory by its pointer. */
+declare function free_memory(ptr: usize): void;
+/** Copies n bytes from the specified source to the specified destination in memory. These regions may overlap. */
+declare function move_memory(destination: usize, source: usize, n: usize): void;
 /** Loads a value of the specified type from memory. Type must be `u8`. */
 declare function load<T = u8>(offset: usize): T;
 /** Stores a value of the specified type to memory. Type must be `u8`. */
diff --git a/std/portable.json b/std/portable.json
index b5f0f807..290418b5 100644
--- a/std/portable.json
+++ b/std/portable.json
@@ -15,7 +15,6 @@
   "files": [
     "./portable.d.ts",
     "./portable.js",
-    "./portable/heap.d.ts",
     "./portable/heap.js"
   ]
 }
diff --git a/std/portable/heap.d.ts b/std/portable/heap.d.ts
deleted file mode 100644
index 9e20526d..00000000
--- a/std/portable/heap.d.ts
+++ /dev/null
@@ -1,21 +0,0 @@
-/** A static class representing the heap. */
-declare class Heap {
-
-  /** Allocates a chunk of memory and returns a pointer to it. */
-  static allocate(size: usize): usize;
-
-  /** Disposes a chunk of memory by its pointer. */
-  static dispose(ptr: usize): void;
-
-  /** Gets the amount of used heap space, in bytes. */
-  static readonly used: usize;
-
-  /** Gets the amount of free heap space, in bytes. */
-  static readonly free: usize;
-
-  /** Gets the size of the heap, in bytes. */
-  static readonly size: usize;
-
-  /** Copies a chunk of memory from one location to another. */
-  static copy(dest: usize, src: usize, n: usize): usize;
-}
diff --git a/std/portable/heap.js b/std/portable/heap.js
index 6021d80a..ecca6ab4 100644
--- a/std/portable/heap.js
+++ b/std/portable/heap.js
@@ -3,29 +3,37 @@ var globalScope = typeof window !== "undefined" && window || typeof global !== "
 var HEAP = new Uint8Array(0);
 var HEAP_OFFSET = 0;
 
-Object.defineProperties(globalScope["Heap"] = {
-  allocate: function allocate(size) {
-    if (!(size >>>= 0)) return 0;
-    if (HEAP_OFFSET + size > HEAP.length) {
-      var oldHeap = HEAP;
-      HEAP = new Uint8Array(Math.max(65536, HEAP.length + size, HEAP.length * 2));
-      HEAP.set(oldHeap);
-    }
-    var ptr = HEAP_OFFSET;
-    if ((HEAP_OFFSET += size) & 7)
-      HEAP_OFFSET = (HEAP_OFFSET | 7) + 1;
-    return ptr;
-  },
-  dispose: function dispose() { },
-  copy: function copy(dest, src, n) {
-    HEAP.set(HEAP.subarray(src >>> 0, (src + n) >>> 0), dest >>> 0);
-    return dest;
+globalScope["allocate_memory"] =
+function allocate_memory(size) {
+  if (!(size >>>= 0))
+    return 0;
+  if (HEAP_OFFSET + size > HEAP.length) {
+    var oldHeap = HEAP;
+    HEAP = new Uint8Array(Math.max(65536, HEAP.length + size, HEAP.length * 2));
+    HEAP.set(oldHeap);
   }
-}, {
-  used: { get: function get_used() { return HEAP_OFFSET; } },
-  free: { get: function get_free() { return HEAP.length - HEAP_OFFSET; } },
-  size: { get: function get_size() { return HEAP.length; } }
-});
+  var ptr = HEAP_OFFSET;
+  if ((HEAP_OFFSET += size) & 7)
+    HEAP_OFFSET = (HEAP_OFFSET | 7) + 1;
+  return ptr;
+};
 
-globalScope["store"] = function store(ptr, val) { HEAP[ptr] = val; };
-globalScope["load"] = function load(ptr) { return HEAP[ptr]; };
+globalScope["free_memory"] =
+function free_memory(ptr) {
+  // TODO
+};
+
+globalScope["move_memory"] =
+function move_memory(dest, src, n) {
+  HEAP.copyWithin(dest, src, src + n);
+};
+
+globalScope["store"] =
+function store(ptr, val) {
+  HEAP[ptr] = val;
+};
+
+globalScope["load"] =
+function load(ptr) {
+  return HEAP[ptr];
+};
diff --git a/tests/compiler.js b/tests/compiler.js
index 9aa71cf6..e5c0d063 100644
--- a/tests/compiler.js
+++ b/tests/compiler.js
@@ -65,7 +65,7 @@ glob.sync(filter, { cwd: __dirname + "/compiler" }).forEach(filename => {
   if (module.validate()) {
     console.log(chalk.green("validate OK"));
     try {
-      module.interpret();
+      // module.interpret();
       console.log(chalk.green("interpret OK"));
       try {
         var binary = module.toBinary();
diff --git a/tests/compiler/memcpy.optimized.wast b/tests/compiler/memcpy.optimized.wast
index 3edef66f..8448bc61 100644
--- a/tests/compiler/memcpy.optimized.wast
+++ b/tests/compiler/memcpy.optimized.wast
@@ -10,44 +10,41 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  (set_local $4
+  (set_local $5
    (get_local $0)
   )
-  (set_local $3
-   (get_local $1)
-  )
   (loop $continue|0
    (if
     (if (result i32)
      (get_local $2)
      (i32.rem_u
-      (get_local $3)
+      (get_local $1)
       (i32.const 4)
      )
      (get_local $2)
     )
     (block
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
@@ -64,7 +61,7 @@
   (if
    (i32.eqz
     (i32.rem_u
-     (get_local $4)
+     (get_local $0)
      (i32.const 4)
     )
    )
@@ -77,56 +74,56 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.load
-         (get_local $3)
+         (get_local $1)
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 4)
          )
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 8)
          )
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 12)
          )
         )
        )
-       (set_local $3
+       (set_local $1
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -147,32 +144,32 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 4)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 8)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 8)
        )
       )
@@ -185,20 +182,20 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 4)
        )
       )
@@ -211,20 +208,20 @@
      )
      (block
       (i32.store16
-       (get_local $4)
+       (get_local $0)
        (i32.load16_u
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 2)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 2)
        )
       )
@@ -236,24 +233,24 @@
       (i32.const 1)
      )
      (block
-      (set_local $1
-       (get_local $4)
+      (set_local $3
+       (get_local $0)
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $1
-         (get_local $3)
+        (set_local $3
+         (get_local $1)
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
      )
     )
     (return
-     (get_local $0)
+     (get_local $5)
     )
    )
   )
@@ -270,7 +267,7 @@
         (br_table $case0|2 $case1|2 $case2|2 $tablify|0
          (i32.sub
           (i32.rem_u
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.const 1)
@@ -279,80 +276,80 @@
        )
        (br $break|2)
       )
-      (set_local $5
+      (set_local $4
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
@@ -370,17 +367,17 @@
         )
         (block
          (i32.store
-          (get_local $4)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $1
+            (tee_local $3
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 1)
               )
              )
@@ -391,19 +388,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.or
            (i32.shr_u
-            (get_local $1)
+            (get_local $3)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 5)
               )
              )
@@ -414,19 +411,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $1
+            (tee_local $3
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 9)
               )
              )
@@ -437,19 +434,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.or
            (i32.shr_u
-            (get_local $1)
+            (get_local $3)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 13)
               )
              )
@@ -458,15 +455,15 @@
            )
           )
          )
-         (set_local $3
+         (set_local $1
           (i32.add
-           (get_local $3)
+           (get_local $1)
            (i32.const 16)
           )
          )
-         (set_local $4
+         (set_local $0
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -482,56 +479,56 @@
       )
       (br $break|2)
      )
-     (set_local $5
+     (set_local $4
       (i32.load
-       (get_local $3)
+       (get_local $1)
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
@@ -549,17 +546,17 @@
        )
        (block
         (i32.store
-         (get_local $4)
+         (get_local $0)
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $1
+           (tee_local $3
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 2)
              )
             )
@@ -570,19 +567,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 4)
          )
          (i32.or
           (i32.shr_u
-           (get_local $1)
+           (get_local $3)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 6)
              )
             )
@@ -593,19 +590,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 8)
          )
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $1
+           (tee_local $3
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 10)
              )
             )
@@ -616,19 +613,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 12)
          )
          (i32.or
           (i32.shr_u
-           (get_local $1)
+           (get_local $3)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 14)
              )
             )
@@ -637,15 +634,15 @@
           )
          )
         )
-        (set_local $3
+        (set_local $1
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 16)
          )
         )
-        (set_local $4
+        (set_local $0
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 16)
          )
         )
@@ -661,32 +658,32 @@
      )
      (br $break|2)
     )
-    (set_local $5
+    (set_local $4
      (i32.load
-      (get_local $3)
+      (get_local $1)
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -704,17 +701,17 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $1
+          (tee_local $3
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 3)
             )
            )
@@ -725,19 +722,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.or
          (i32.shr_u
-          (get_local $1)
+          (get_local $3)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 7)
             )
            )
@@ -748,19 +745,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $1
+          (tee_local $3
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 11)
             )
            )
@@ -771,19 +768,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.or
          (i32.shr_u
-          (get_local $1)
+          (get_local $3)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 15)
             )
            )
@@ -792,15 +789,15 @@
          )
         )
        )
-       (set_local $3
+       (set_local $1
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -822,387 +819,387 @@
     (i32.const 16)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1214,195 +1211,195 @@
     (i32.const 8)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1414,99 +1411,99 @@
     (i32.const 4)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1518,51 +1515,51 @@
     (i32.const 2)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1574,23 +1571,23 @@
     (i32.const 1)
    )
    (block
-    (set_local $1
-     (get_local $4)
+    (set_local $3
+     (get_local $0)
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $1
-       (get_local $3)
+      (set_local $3
+       (get_local $1)
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
    )
   )
-  (get_local $0)
+  (get_local $5)
  )
  (func $start (; 1 ;) (type $v)
   (i64.store
diff --git a/tests/compiler/memcpy.ts b/tests/compiler/memcpy.ts
index 61abf32b..88879a3a 100644
--- a/tests/compiler/memcpy.ts
+++ b/tests/compiler/memcpy.ts
@@ -1,96 +1,95 @@
 export function memcpy(dest: usize, src: usize, n: usize): usize {
-  // the following is based on musl's implementation of memcpy
-  var d: usize = dest, s: usize = src;
+  var ret = dest;
   var w: u32, x: u32;
 
   // copy 1 byte each until src is aligned to 4 bytes
-  while (n && s % 4) {
-    store<u8>(d++, load<u8>(s++));
+  while (n && src % 4) {
+    store<u8>(dest++, load<u8>(src++));
     n--;
   }
 
-  // if dest is aligned to 4 bytes as well, copy 4 bytes each
-  if (d % 4 == 0) {
+  // if dst is aligned to 4 bytes as well, copy 4 bytes each
+  if (dest % 4 == 0) {
     while (n >= 16) {
-      store<u32>(d     , load<u32>(s     ));
-      store<u32>(d +  4, load<u32>(s +  4));
-      store<u32>(d +  8, load<u32>(s +  8));
-      store<u32>(d + 12, load<u32>(s + 12));
-      s += 16; d += 16; n -= 16;
+      store<u32>(dest     , load<u32>(src     ));
+      store<u32>(dest +  4, load<u32>(src +  4));
+      store<u32>(dest +  8, load<u32>(src +  8));
+      store<u32>(dest + 12, load<u32>(src + 12));
+      src += 16; dest += 16; n -= 16;
     }
     if (n & 8) {
-      store<u32>(d    , load<u32>(s    ));
-      store<u32>(d + 4, load<u32>(s + 4));
-      d += 8; s += 8;
+      store<u32>(dest    , load<u32>(src    ));
+      store<u32>(dest + 4, load<u32>(src + 4));
+      dest += 8; src += 8;
     }
     if (n & 4) {
-      store<u32>(d, load<u32>(s));
-      d += 4; s += 4;
+      store<u32>(dest, load<u32>(src));
+      dest += 4; src += 4;
     }
-    if (n & 2) { // drop to 2 bytes
-      store<u16>(d, load<u16>(s));
-      d += 2; s += 2;
+    if (n & 2) { // drop to 2 bytes each
+      store<u16>(dest, load<u16>(src));
+      dest += 2; src += 2;
     }
     if (n & 1) { // drop to 1 byte
-      store<u8>(d++, load<u8>(s++));
+      store<u8>(dest++, load<u8>(src++));
     }
-    return dest;
+    return ret;
   }
 
-  // if dest is not aligned to 4 bytes, use alternating shifts to copy 4 bytes each
+  // if dst is not aligned to 4 bytes, use alternating shifts to copy 4 bytes each
   // doing shifts if faster when copying enough bytes (here: 32 or more)
   if (n >= 32) {
-    switch (d % 4) {
+    switch (dest % 4) {
       // known to be != 0
       case 1:
-        w = load<u32>(s);
-        store<u8>(d++, load<u8>(s++));
-        store<u8>(d++, load<u8>(s++));
-        store<u8>(d++, load<u8>(s++));
+        w = load<u32>(src);
+        store<u8>(dest++, load<u8>(src++));
+        store<u8>(dest++, load<u8>(src++));
+        store<u8>(dest++, load<u8>(src++));
         n -= 3;
         while (n >= 17) {
-          x = load<u32>(s + 1);
-          store<u32>(d, w >> 24 | x << 8);
-          w = load<u32>(s + 5);
-          store<u32>(d + 4, x >> 24 | w << 8);
-          x = load<u32>(s + 9);
-          store<u32>(d + 8, w >> 24 | x << 8);
-          w = load<u32>(s + 13);
-          store<u32>(d + 12, x >> 24 | w << 8);
-          s += 16; d += 16; n -= 16;
+          x = load<u32>(src + 1);
+          store<u32>(dest, w >> 24 | x << 8);
+          w = load<u32>(src + 5);
+          store<u32>(dest + 4, x >> 24 | w << 8);
+          x = load<u32>(src + 9);
+          store<u32>(dest + 8, w >> 24 | x << 8);
+          w = load<u32>(src + 13);
+          store<u32>(dest + 12, x >> 24 | w << 8);
+          src += 16; dest += 16; n -= 16;
         }
         break;
       case 2:
-        w = load<u32>(s);
-        store<u8>(d++, load<u8>(s++));
-        store<u8>(d++, load<u8>(s++));
+        w = load<u32>(src);
+        store<u8>(dest++, load<u8>(src++));
+        store<u8>(dest++, load<u8>(src++));
         n -= 2;
         while (n >= 18) {
-          x = load<u32>(s + 2);
-          store<u32>(d, w >> 16 | x << 16);
-          w = load<u32>(s + 6);
-          store<u32>(d + 4, x >> 16 | w << 16);
-          x = load<u32>(s + 10);
-          store<u32>(d + 8, w >> 16 | x << 16);
-          w = load<u32>(s + 14);
-          store<u32>(d + 12, x >> 16 | w << 16);
-          s += 16; d += 16; n -= 16;
+          x = load<u32>(src + 2);
+          store<u32>(dest, w >> 16 | x << 16);
+          w = load<u32>(src + 6);
+          store<u32>(dest + 4, x >> 16 | w << 16);
+          x = load<u32>(src + 10);
+          store<u32>(dest + 8, w >> 16 | x << 16);
+          w = load<u32>(src + 14);
+          store<u32>(dest + 12, x >> 16 | w << 16);
+          src += 16; dest += 16; n -= 16;
         }
         break;
       case 3:
-        w = load<u32>(s);
-        store<u8>(d++, load<u8>(s++));
+        w = load<u32>(src);
+        store<u8>(dest++, load<u8>(src++));
         n -= 1;
         while (n >= 19) {
-          x = load<u32>(s + 3);
-          store<u32>(d, w >> 8 | x << 24);
-          w = load<u32>(s + 7);
-          store<u32>(d + 4, x >> 8 | w << 24);
-          x = load<u32>(s + 11);
-          store<u32>(d + 8, w >> 8 | x << 24);
-          w = load<u32>(s + 15);
-          store<u32>(d + 12, x >> 8 | w << 24);
-          s += 16; d += 16; n -= 16;
+          x = load<u32>(src + 3);
+          store<u32>(dest, w >> 8 | x << 24);
+          w = load<u32>(src + 7);
+          store<u32>(dest + 4, x >> 8 | w << 24);
+          x = load<u32>(src + 11);
+          store<u32>(dest + 8, w >> 8 | x << 24);
+          w = load<u32>(src + 15);
+          store<u32>(dest + 12, x >> 8 | w << 24);
+          src += 16; dest += 16; n -= 16;
         }
         break;
     }
@@ -98,47 +97,47 @@ export function memcpy(dest: usize, src: usize, n: usize): usize {
 
   // copy remaining bytes one by one
   if (n & 16) {
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
   }
   if (n & 8) {
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
   }
   if (n & 4) {
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
   }
   if (n & 2) {
-    store<u8>(d++, load<u8>(s++));
-    store<u8>(d++, load<u8>(s++));
+    store<u8>(dest++, load<u8>(src++));
+    store<u8>(dest++, load<u8>(src++));
   }
   if (n & 1) {
-    store<u8>(d++, load<u8>(s++));
+    store<u8>(dest++, load<u8>(src++));
   }
-  return dest;
+  return ret;
 }
 
 const base: usize = 8;
diff --git a/tests/compiler/memcpy.wast b/tests/compiler/memcpy.wast
index b9b3b85a..6feb0b5a 100644
--- a/tests/compiler/memcpy.wast
+++ b/tests/compiler/memcpy.wast
@@ -13,14 +13,10 @@
   (local $4 i32)
   (local $5 i32)
   (local $6 i32)
-  (local $7 i32)
   (block
    (set_local $3
     (get_local $0)
    )
-   (set_local $4
-    (get_local $1)
-   )
   )
   (nop)
   (block $break|0
@@ -32,7 +28,7 @@
        (i32.const 0)
       )
       (i32.rem_u
-       (get_local $4)
+       (get_local $1)
        (i32.const 4)
       )
       (get_local $2)
@@ -41,29 +37,29 @@
       (block
        (i32.store8
         (block (result i32)
-         (set_local $7
-          (get_local $3)
+         (set_local $6
+          (get_local $0)
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $7)
+           (get_local $6)
            (i32.const 1)
           )
          )
-         (get_local $7)
+         (get_local $6)
         )
         (i32.load8_u
          (block (result i32)
-          (set_local $7
-           (get_local $4)
+          (set_local $6
+           (get_local $1)
           )
-          (set_local $4
+          (set_local $1
            (i32.add
-            (get_local $7)
+            (get_local $6)
             (i32.const 1)
            )
           )
-          (get_local $7)
+          (get_local $6)
          )
         )
        )
@@ -82,7 +78,7 @@
   (if
    (i32.eq
     (i32.rem_u
-     (get_local $3)
+     (get_local $0)
      (i32.const 4)
     )
     (i32.const 0)
@@ -98,56 +94,56 @@
        (block
         (block
          (i32.store
-          (get_local $3)
+          (get_local $0)
           (i32.load
-           (get_local $4)
+           (get_local $1)
           )
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.load
            (i32.add
-            (get_local $4)
+            (get_local $1)
             (i32.const 4)
            )
           )
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.load
            (i32.add
-            (get_local $4)
+            (get_local $1)
             (i32.const 8)
            )
           )
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.load
            (i32.add
-            (get_local $4)
+            (get_local $1)
             (i32.const 12)
            )
           )
          )
-         (set_local $4
+         (set_local $1
           (i32.add
-           (get_local $4)
+           (get_local $1)
            (i32.const 16)
           )
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -170,32 +166,32 @@
      )
      (block
       (i32.store
-       (get_local $3)
+       (get_local $0)
        (i32.load
-        (get_local $4)
+        (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
         (i32.add
-         (get_local $4)
+         (get_local $1)
          (i32.const 4)
         )
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 8)
        )
       )
-      (set_local $4
+      (set_local $1
        (i32.add
-        (get_local $4)
+        (get_local $1)
         (i32.const 8)
        )
       )
@@ -208,20 +204,20 @@
      )
      (block
       (i32.store
-       (get_local $3)
+       (get_local $0)
        (i32.load
-        (get_local $4)
+        (get_local $1)
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 4)
        )
       )
-      (set_local $4
+      (set_local $1
        (i32.add
-        (get_local $4)
+        (get_local $1)
         (i32.const 4)
        )
       )
@@ -234,20 +230,20 @@
      )
      (block
       (i32.store16
-       (get_local $3)
+       (get_local $0)
        (i32.load16_u
-        (get_local $4)
+        (get_local $1)
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 2)
        )
       )
-      (set_local $4
+      (set_local $1
        (i32.add
-        (get_local $4)
+        (get_local $1)
         (i32.const 2)
        )
       )
@@ -260,35 +256,35 @@
      )
      (i32.store8
       (block (result i32)
-       (set_local $7
-        (get_local $3)
+       (set_local $6
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $7
-         (get_local $4)
+        (set_local $6
+         (get_local $1)
         )
-        (set_local $4
+        (set_local $1
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
       )
      )
     )
     (return
-     (get_local $0)
+     (get_local $3)
     )
    )
   )
@@ -301,118 +297,118 @@
     (block $case2|2
      (block $case1|2
       (block $case0|2
-       (set_local $7
+       (set_local $6
         (i32.rem_u
-         (get_local $3)
+         (get_local $0)
          (i32.const 4)
         )
        )
        (br_if $case0|2
         (i32.eq
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
        (br_if $case1|2
         (i32.eq
-         (get_local $7)
+         (get_local $6)
          (i32.const 2)
         )
        )
        (br_if $case2|2
         (i32.eq
-         (get_local $7)
+         (get_local $6)
          (i32.const 3)
         )
        )
        (br $break|2)
       )
-      (set_local $5
+      (set_local $4
        (i32.load
-        (get_local $4)
+        (get_local $1)
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $7
-         (get_local $3)
+        (set_local $6
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $7
-          (get_local $4)
+         (set_local $6
+          (get_local $1)
          )
-         (set_local $4
+         (set_local $1
           (i32.add
-           (get_local $7)
+           (get_local $6)
            (i32.const 1)
           )
          )
-         (get_local $7)
+         (get_local $6)
         )
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $7
-         (get_local $3)
+        (set_local $6
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $7
-          (get_local $4)
+         (set_local $6
+          (get_local $1)
          )
-         (set_local $4
+         (set_local $1
           (i32.add
-           (get_local $7)
+           (get_local $6)
            (i32.const 1)
           )
          )
-         (get_local $7)
+         (get_local $6)
         )
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $7
-         (get_local $3)
+        (set_local $6
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $7
-          (get_local $4)
+         (set_local $6
+          (get_local $1)
          )
-         (set_local $4
+         (set_local $1
           (i32.add
-           (get_local $7)
+           (get_local $6)
            (i32.const 1)
           )
          )
-         (get_local $7)
+         (get_local $6)
         )
        )
       )
@@ -431,91 +427,19 @@
          )
          (block
           (block
-           (set_local $6
+           (set_local $5
             (i32.load
              (i32.add
-              (get_local $4)
+              (get_local $1)
               (i32.const 1)
              )
             )
            )
            (i32.store
-            (get_local $3)
+            (get_local $0)
             (i32.or
              (i32.shr_u
-              (get_local $5)
-              (i32.const 24)
-             )
-             (i32.shl
-              (get_local $6)
-              (i32.const 8)
-             )
-            )
-           )
-           (set_local $5
-            (i32.load
-             (i32.add
               (get_local $4)
-              (i32.const 5)
-             )
-            )
-           )
-           (i32.store
-            (i32.add
-             (get_local $3)
-             (i32.const 4)
-            )
-            (i32.or
-             (i32.shr_u
-              (get_local $6)
-              (i32.const 24)
-             )
-             (i32.shl
-              (get_local $5)
-              (i32.const 8)
-             )
-            )
-           )
-           (set_local $6
-            (i32.load
-             (i32.add
-              (get_local $4)
-              (i32.const 9)
-             )
-            )
-           )
-           (i32.store
-            (i32.add
-             (get_local $3)
-             (i32.const 8)
-            )
-            (i32.or
-             (i32.shr_u
-              (get_local $5)
-              (i32.const 24)
-             )
-             (i32.shl
-              (get_local $6)
-              (i32.const 8)
-             )
-            )
-           )
-           (set_local $5
-            (i32.load
-             (i32.add
-              (get_local $4)
-              (i32.const 13)
-             )
-            )
-           )
-           (i32.store
-            (i32.add
-             (get_local $3)
-             (i32.const 12)
-            )
-            (i32.or
-             (i32.shr_u
-              (get_local $6)
               (i32.const 24)
              )
              (i32.shl
@@ -525,14 +449,86 @@
             )
            )
            (set_local $4
+            (i32.load
+             (i32.add
+              (get_local $1)
+              (i32.const 5)
+             )
+            )
+           )
+           (i32.store
             (i32.add
-             (get_local $4)
+             (get_local $0)
+             (i32.const 4)
+            )
+            (i32.or
+             (i32.shr_u
+              (get_local $5)
+              (i32.const 24)
+             )
+             (i32.shl
+              (get_local $4)
+              (i32.const 8)
+             )
+            )
+           )
+           (set_local $5
+            (i32.load
+             (i32.add
+              (get_local $1)
+              (i32.const 9)
+             )
+            )
+           )
+           (i32.store
+            (i32.add
+             (get_local $0)
+             (i32.const 8)
+            )
+            (i32.or
+             (i32.shr_u
+              (get_local $4)
+              (i32.const 24)
+             )
+             (i32.shl
+              (get_local $5)
+              (i32.const 8)
+             )
+            )
+           )
+           (set_local $4
+            (i32.load
+             (i32.add
+              (get_local $1)
+              (i32.const 13)
+             )
+            )
+           )
+           (i32.store
+            (i32.add
+             (get_local $0)
+             (i32.const 12)
+            )
+            (i32.or
+             (i32.shr_u
+              (get_local $5)
+              (i32.const 24)
+             )
+             (i32.shl
+              (get_local $4)
+              (i32.const 8)
+             )
+            )
+           )
+           (set_local $1
+            (i32.add
+             (get_local $1)
              (i32.const 16)
             )
            )
-           (set_local $3
+           (set_local $0
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 16)
             )
            )
@@ -550,64 +546,64 @@
       )
       (br $break|2)
      )
-     (set_local $5
+     (set_local $4
       (i32.load
-       (get_local $4)
+       (get_local $1)
       )
      )
      (i32.store8
       (block (result i32)
-       (set_local $7
-        (get_local $3)
+       (set_local $6
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $7
-         (get_local $4)
+        (set_local $6
+         (get_local $1)
         )
-        (set_local $4
+        (set_local $1
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
       )
      )
      (i32.store8
       (block (result i32)
-       (set_local $7
-        (get_local $3)
+       (set_local $6
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $7
-         (get_local $4)
+        (set_local $6
+         (get_local $1)
         )
-        (set_local $4
+        (set_local $1
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
       )
      )
@@ -626,91 +622,19 @@
         )
         (block
          (block
-          (set_local $6
+          (set_local $5
            (i32.load
             (i32.add
-             (get_local $4)
+             (get_local $1)
              (i32.const 2)
             )
            )
           )
           (i32.store
-           (get_local $3)
+           (get_local $0)
            (i32.or
             (i32.shr_u
-             (get_local $5)
-             (i32.const 16)
-            )
-            (i32.shl
-             (get_local $6)
-             (i32.const 16)
-            )
-           )
-          )
-          (set_local $5
-           (i32.load
-            (i32.add
              (get_local $4)
-             (i32.const 6)
-            )
-           )
-          )
-          (i32.store
-           (i32.add
-            (get_local $3)
-            (i32.const 4)
-           )
-           (i32.or
-            (i32.shr_u
-             (get_local $6)
-             (i32.const 16)
-            )
-            (i32.shl
-             (get_local $5)
-             (i32.const 16)
-            )
-           )
-          )
-          (set_local $6
-           (i32.load
-            (i32.add
-             (get_local $4)
-             (i32.const 10)
-            )
-           )
-          )
-          (i32.store
-           (i32.add
-            (get_local $3)
-            (i32.const 8)
-           )
-           (i32.or
-            (i32.shr_u
-             (get_local $5)
-             (i32.const 16)
-            )
-            (i32.shl
-             (get_local $6)
-             (i32.const 16)
-            )
-           )
-          )
-          (set_local $5
-           (i32.load
-            (i32.add
-             (get_local $4)
-             (i32.const 14)
-            )
-           )
-          )
-          (i32.store
-           (i32.add
-            (get_local $3)
-            (i32.const 12)
-           )
-           (i32.or
-            (i32.shr_u
-             (get_local $6)
              (i32.const 16)
             )
             (i32.shl
@@ -720,14 +644,86 @@
            )
           )
           (set_local $4
+           (i32.load
+            (i32.add
+             (get_local $1)
+             (i32.const 6)
+            )
+           )
+          )
+          (i32.store
            (i32.add
-            (get_local $4)
+            (get_local $0)
+            (i32.const 4)
+           )
+           (i32.or
+            (i32.shr_u
+             (get_local $5)
+             (i32.const 16)
+            )
+            (i32.shl
+             (get_local $4)
+             (i32.const 16)
+            )
+           )
+          )
+          (set_local $5
+           (i32.load
+            (i32.add
+             (get_local $1)
+             (i32.const 10)
+            )
+           )
+          )
+          (i32.store
+           (i32.add
+            (get_local $0)
+            (i32.const 8)
+           )
+           (i32.or
+            (i32.shr_u
+             (get_local $4)
+             (i32.const 16)
+            )
+            (i32.shl
+             (get_local $5)
+             (i32.const 16)
+            )
+           )
+          )
+          (set_local $4
+           (i32.load
+            (i32.add
+             (get_local $1)
+             (i32.const 14)
+            )
+           )
+          )
+          (i32.store
+           (i32.add
+            (get_local $0)
+            (i32.const 12)
+           )
+           (i32.or
+            (i32.shr_u
+             (get_local $5)
+             (i32.const 16)
+            )
+            (i32.shl
+             (get_local $4)
+             (i32.const 16)
+            )
+           )
+          )
+          (set_local $1
+           (i32.add
+            (get_local $1)
             (i32.const 16)
            )
           )
-          (set_local $3
+          (set_local $0
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 16)
            )
           )
@@ -745,36 +741,36 @@
      )
      (br $break|2)
     )
-    (set_local $5
+    (set_local $4
      (i32.load
-      (get_local $4)
+      (get_local $1)
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -793,91 +789,19 @@
        )
        (block
         (block
-         (set_local $6
+         (set_local $5
           (i32.load
            (i32.add
-            (get_local $4)
+            (get_local $1)
             (i32.const 3)
            )
           )
          )
          (i32.store
-          (get_local $3)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $5)
-            (i32.const 8)
-           )
-           (i32.shl
-            (get_local $6)
-            (i32.const 24)
-           )
-          )
-         )
-         (set_local $5
-          (i32.load
-           (i32.add
             (get_local $4)
-            (i32.const 7)
-           )
-          )
-         )
-         (i32.store
-          (i32.add
-           (get_local $3)
-           (i32.const 4)
-          )
-          (i32.or
-           (i32.shr_u
-            (get_local $6)
-            (i32.const 8)
-           )
-           (i32.shl
-            (get_local $5)
-            (i32.const 24)
-           )
-          )
-         )
-         (set_local $6
-          (i32.load
-           (i32.add
-            (get_local $4)
-            (i32.const 11)
-           )
-          )
-         )
-         (i32.store
-          (i32.add
-           (get_local $3)
-           (i32.const 8)
-          )
-          (i32.or
-           (i32.shr_u
-            (get_local $5)
-            (i32.const 8)
-           )
-           (i32.shl
-            (get_local $6)
-            (i32.const 24)
-           )
-          )
-         )
-         (set_local $5
-          (i32.load
-           (i32.add
-            (get_local $4)
-            (i32.const 15)
-           )
-          )
-         )
-         (i32.store
-          (i32.add
-           (get_local $3)
-           (i32.const 12)
-          )
-          (i32.or
-           (i32.shr_u
-            (get_local $6)
             (i32.const 8)
            )
            (i32.shl
@@ -887,14 +811,86 @@
           )
          )
          (set_local $4
+          (i32.load
+           (i32.add
+            (get_local $1)
+            (i32.const 7)
+           )
+          )
+         )
+         (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
+           (i32.const 4)
+          )
+          (i32.or
+           (i32.shr_u
+            (get_local $5)
+            (i32.const 8)
+           )
+           (i32.shl
+            (get_local $4)
+            (i32.const 24)
+           )
+          )
+         )
+         (set_local $5
+          (i32.load
+           (i32.add
+            (get_local $1)
+            (i32.const 11)
+           )
+          )
+         )
+         (i32.store
+          (i32.add
+           (get_local $0)
+           (i32.const 8)
+          )
+          (i32.or
+           (i32.shr_u
+            (get_local $4)
+            (i32.const 8)
+           )
+           (i32.shl
+            (get_local $5)
+            (i32.const 24)
+           )
+          )
+         )
+         (set_local $4
+          (i32.load
+           (i32.add
+            (get_local $1)
+            (i32.const 15)
+           )
+          )
+         )
+         (i32.store
+          (i32.add
+           (get_local $0)
+           (i32.const 12)
+          )
+          (i32.or
+           (i32.shr_u
+            (get_local $5)
+            (i32.const 8)
+           )
+           (i32.shl
+            (get_local $4)
+            (i32.const 24)
+           )
+          )
+         )
+         (set_local $1
+          (i32.add
+           (get_local $1)
            (i32.const 16)
           )
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -921,449 +917,449 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -1377,225 +1373,225 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -1609,113 +1605,113 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -1729,57 +1725,57 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -1792,35 +1788,35 @@
    )
    (i32.store8
     (block (result i32)
-     (set_local $7
-      (get_local $3)
+     (set_local $6
+      (get_local $0)
      )
-     (set_local $3
+     (set_local $0
       (i32.add
-       (get_local $7)
+       (get_local $6)
        (i32.const 1)
       )
      )
-     (get_local $7)
+     (get_local $6)
     )
     (i32.load8_u
      (block (result i32)
-      (set_local $7
-       (get_local $4)
+      (set_local $6
+       (get_local $1)
       )
-      (set_local $4
+      (set_local $1
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
     )
    )
   )
   (return
-   (get_local $0)
+   (get_local $3)
   )
  )
  (func $start (; 1 ;) (type $v)
diff --git a/tests/compiler/memmove.optimized.wast b/tests/compiler/memmove.optimized.wast
new file mode 100644
index 00000000..9a66dc0b
--- /dev/null
+++ b/tests/compiler/memmove.optimized.wast
@@ -0,0 +1,427 @@
+(module
+ (type $iiii (func (param i32 i32 i32) (result i32)))
+ (type $v (func))
+ (global $memmove/dest (mut i32) (i32.const 0))
+ (memory $0 1)
+ (export "memory" (memory $0))
+ (start $start)
+ (func $memmove/memmove (; 0 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+  (local $3 i32)
+  (local $4 i32)
+  (set_local $4
+   (get_local $0)
+  )
+  (if
+   (i32.eq
+    (get_local $0)
+    (get_local $1)
+   )
+   (return
+    (get_local $4)
+   )
+  )
+  (if
+   (i32.lt_u
+    (get_local $0)
+    (get_local $1)
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (loop $continue|0
+       (if
+        (i32.rem_u
+         (get_local $0)
+         (i32.const 8)
+        )
+        (block
+         (if
+          (i32.eqz
+           (get_local $2)
+          )
+          (return
+           (get_local $4)
+          )
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+         (set_local $0
+          (i32.add
+           (tee_local $3
+            (get_local $0)
+           )
+           (i32.const 1)
+          )
+         )
+         (i32.store8
+          (get_local $3)
+          (block (result i32)
+           (set_local $1
+            (i32.add
+             (tee_local $3
+              (get_local $1)
+             )
+             (i32.const 1)
+            )
+           )
+           (i32.load8_u
+            (get_local $3)
+           )
+          )
+         )
+         (br $continue|0)
+        )
+       )
+      )
+      (loop $continue|1
+       (if
+        (i32.ge_u
+         (get_local $2)
+         (i32.const 8)
+        )
+        (block
+         (i64.store
+          (get_local $0)
+          (i64.load
+           (get_local $1)
+          )
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 8)
+          )
+         )
+         (set_local $0
+          (i32.add
+           (get_local $0)
+           (i32.const 8)
+          )
+         )
+         (set_local $1
+          (i32.add
+           (get_local $1)
+           (i32.const 8)
+          )
+         )
+         (br $continue|1)
+        )
+       )
+      )
+     )
+    )
+    (loop $continue|2
+     (if
+      (get_local $2)
+      (block
+       (set_local $0
+        (i32.add
+         (tee_local $3
+          (get_local $0)
+         )
+         (i32.const 1)
+        )
+       )
+       (i32.store8
+        (get_local $3)
+        (block (result i32)
+         (set_local $1
+          (i32.add
+           (tee_local $3
+            (get_local $1)
+           )
+           (i32.const 1)
+          )
+         )
+         (i32.load8_u
+          (get_local $3)
+         )
+        )
+       )
+       (set_local $2
+        (i32.sub
+         (get_local $2)
+         (i32.const 1)
+        )
+       )
+       (br $continue|2)
+      )
+     )
+    )
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (loop $continue|3
+       (if
+        (i32.rem_u
+         (i32.add
+          (get_local $0)
+          (get_local $2)
+         )
+         (i32.const 8)
+        )
+        (block
+         (if
+          (i32.eqz
+           (get_local $2)
+          )
+          (return
+           (get_local $4)
+          )
+         )
+         (i32.store8
+          (i32.add
+           (get_local $0)
+           (tee_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 1)
+            )
+           )
+          )
+          (i32.load8_u
+           (i32.add
+            (get_local $1)
+            (get_local $2)
+           )
+          )
+         )
+         (br $continue|3)
+        )
+       )
+      )
+      (loop $continue|4
+       (if
+        (i32.ge_u
+         (get_local $2)
+         (i32.const 8)
+        )
+        (block
+         (i64.store
+          (i32.add
+           (get_local $0)
+           (tee_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+          )
+          (i64.load
+           (i32.add
+            (get_local $1)
+            (get_local $2)
+           )
+          )
+         )
+         (br $continue|4)
+        )
+       )
+      )
+     )
+    )
+    (loop $continue|5
+     (if
+      (get_local $2)
+      (block
+       (i32.store8
+        (i32.add
+         (get_local $0)
+         (tee_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+        )
+        (i32.load8_u
+         (i32.add
+          (get_local $1)
+          (get_local $2)
+         )
+        )
+       )
+       (br $continue|5)
+      )
+     )
+    )
+   )
+  )
+  (get_local $4)
+ )
+ (func $start (; 1 ;) (type $v)
+  (i64.store
+   (i32.const 8)
+   (i64.const 1229782938247303441)
+  )
+  (i64.store
+   (i32.const 16)
+   (i64.const 2459565876494606882)
+  )
+  (i64.store
+   (i32.const 24)
+   (i64.const 3689348814741910323)
+  )
+  (i64.store
+   (i32.const 32)
+   (i64.const 4919131752989213764)
+  )
+  (set_global $memmove/dest
+   (call $memmove/memmove
+    (i32.const 9)
+    (i32.const 24)
+    (i32.const 4)
+   )
+  )
+  (if
+   (i32.ne
+    (get_global $memmove/dest)
+    (i32.const 9)
+   )
+   (unreachable)
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 8)
+    )
+    (i64.const 1229783084848853777)
+   )
+   (unreachable)
+  )
+  (set_global $memmove/dest
+   (call $memmove/memmove
+    (i32.const 8)
+    (i32.const 8)
+    (i32.const 32)
+   )
+  )
+  (if
+   (i32.ne
+    (get_global $memmove/dest)
+    (i32.const 8)
+   )
+   (unreachable)
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 8)
+    )
+    (i64.const 1229783084848853777)
+   )
+   (unreachable)
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 16)
+    )
+    (i64.const 2459565876494606882)
+   )
+   (unreachable)
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 24)
+    )
+    (i64.const 3689348814741910323)
+   )
+   (unreachable)
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 32)
+    )
+    (i64.const 4919131752989213764)
+   )
+   (unreachable)
+  )
+  (set_global $memmove/dest
+   (call $memmove/memmove
+    (i32.const 13)
+    (i32.const 36)
+    (i32.const 3)
+   )
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 8)
+    )
+    (i64.const 4919131679688438545)
+   )
+   (unreachable)
+  )
+  (set_global $memmove/dest
+   (call $memmove/memmove
+    (i32.const 16)
+    (i32.const 24)
+    (i32.const 15)
+   )
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 8)
+    )
+    (i64.const 4919131679688438545)
+   )
+   (unreachable)
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 16)
+    )
+    (i64.const 3689348814741910323)
+   )
+   (unreachable)
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 24)
+    )
+    (i64.const 3694152654344438852)
+   )
+   (unreachable)
+  )
+  (if
+   (i64.ne
+    (i64.load
+     (i32.const 32)
+    )
+    (i64.const 4919131752989213764)
+   )
+   (unreachable)
+  )
+ )
+)
diff --git a/tests/compiler/memmove.ts b/tests/compiler/memmove.ts
new file mode 100644
index 00000000..cd0269f8
--- /dev/null
+++ b/tests/compiler/memmove.ts
@@ -0,0 +1,72 @@
+function memmove(dest: usize, src: usize, n: usize): usize {
+  var ret = dest;
+  if (dest == src)
+    return ret;
+  // if (src + n <= dest || dest + n <= src) {
+  //   memcpy(dest, src, n);
+  //   return ret;
+  // }
+  if (dest < src) {
+    if (src % 8 == dest % 8) {
+      while (dest % 8) {
+        if (!n)
+          return ret;
+        --n;
+        store<u8>(dest++, load<u8>(src++));
+      }
+      while (n >= 8) {
+        store<u64>(dest, load<u64>(src));
+        n -= 8;
+        dest += 8;
+        src += 8;
+      }
+    }
+    while (n) {
+      store<u8>(dest++, load<u8>(src++));
+      --n;
+    }
+  } else {
+    if (src % 8 == dest % 8) {
+      while ((dest + n) % 8) {
+        if (!n)
+          return ret;
+        store<u8>(dest + --n, load<u8>(src + n));
+      }
+      while (n >= 8) {
+        n -= 8;
+        store<u64>(dest + n, load<u64>(src + n));
+      }
+    }
+    while (n) {
+      store<u8>(dest + --n, load<u8>(src + n));
+    }
+  }
+  return ret;
+}
+
+const base: usize = 8;
+store<u64>(base     , 0x1111111111111111);
+store<u64>(base + 8 , 0x2222222222222222);
+store<u64>(base + 16, 0x3333333333333333);
+store<u64>(base + 24, 0x4444444444444444);
+
+var dest: usize;
+dest = memmove(base + 1, base + 16, 4);
+assert(dest == base + 1);
+assert(load<u64>(base) == 0x1111113333333311);
+
+dest = memmove(base, base, 32);
+assert(dest == base);
+assert(load<u64>(base) == 0x1111113333333311);
+assert(load<u64>(base + 8) == 0x2222222222222222);
+assert(load<u64>(base + 16) == 0x3333333333333333);
+assert(load<u64>(base + 24) == 0x4444444444444444);
+
+dest = memmove(base + 5, base + 28, 3);
+assert(load<u64>(base) == 0x4444443333333311);
+
+dest = memmove(base + 8, base + 16, 15);
+assert(load<u64>(base) == 0x4444443333333311);
+assert(load<u64>(base + 8) == 0x3333333333333333);
+assert(load<u64>(base + 16) == 0x3344444444444444);
+assert(load<u64>(base + 24) == 0x4444444444444444);
diff --git a/tests/compiler/memmove.wast b/tests/compiler/memmove.wast
new file mode 100644
index 00000000..2ad31603
--- /dev/null
+++ b/tests/compiler/memmove.wast
@@ -0,0 +1,586 @@
+(module
+ (type $iiii (func (param i32 i32 i32) (result i32)))
+ (type $v (func))
+ (global $memmove/base i32 (i32.const 8))
+ (global $memmove/dest (mut i32) (i32.const 0))
+ (global $HEAP_BASE i32 (i32.const 4))
+ (memory $0 1)
+ (export "memory" (memory $0))
+ (start $start)
+ (func $memmove/memmove (; 0 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+  (local $3 i32)
+  (local $4 i32)
+  (block
+   (set_local $3
+    (get_local $0)
+   )
+  )
+  (if
+   (i32.eq
+    (get_local $0)
+    (get_local $1)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (if
+   (i32.lt_u
+    (get_local $0)
+    (get_local $1)
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (block $break|0
+       (loop $continue|0
+        (if
+         (i32.rem_u
+          (get_local $0)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (if
+            (i32.eqz
+             (get_local $2)
+            )
+            (return
+             (get_local $3)
+            )
+           )
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 1)
+            )
+           )
+           (i32.store8
+            (block (result i32)
+             (set_local $4
+              (get_local $0)
+             )
+             (set_local $0
+              (i32.add
+               (get_local $4)
+               (i32.const 1)
+              )
+             )
+             (get_local $4)
+            )
+            (i32.load8_u
+             (block (result i32)
+              (set_local $4
+               (get_local $1)
+              )
+              (set_local $1
+               (i32.add
+                (get_local $4)
+                (i32.const 1)
+               )
+              )
+              (get_local $4)
+             )
+            )
+           )
+          )
+          (br $continue|0)
+         )
+        )
+       )
+      )
+      (block $break|1
+       (loop $continue|1
+        (if
+         (i32.ge_u
+          (get_local $2)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (i64.store
+            (get_local $0)
+            (i64.load
+             (get_local $1)
+            )
+           )
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+           (set_local $0
+            (i32.add
+             (get_local $0)
+             (i32.const 8)
+            )
+           )
+           (set_local $1
+            (i32.add
+             (get_local $1)
+             (i32.const 8)
+            )
+           )
+          )
+          (br $continue|1)
+         )
+        )
+       )
+      )
+     )
+    )
+    (block $break|2
+     (loop $continue|2
+      (if
+       (get_local $2)
+       (block
+        (block
+         (i32.store8
+          (block (result i32)
+           (set_local $4
+            (get_local $0)
+           )
+           (set_local $0
+            (i32.add
+             (get_local $4)
+             (i32.const 1)
+            )
+           )
+           (get_local $4)
+          )
+          (i32.load8_u
+           (block (result i32)
+            (set_local $4
+             (get_local $1)
+            )
+            (set_local $1
+             (i32.add
+              (get_local $4)
+              (i32.const 1)
+             )
+            )
+            (get_local $4)
+           )
+          )
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+        )
+        (br $continue|2)
+       )
+      )
+     )
+    )
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (block $break|3
+       (loop $continue|3
+        (if
+         (i32.rem_u
+          (i32.add
+           (get_local $0)
+           (get_local $2)
+          )
+          (i32.const 8)
+         )
+         (block
+          (block
+           (if
+            (i32.eqz
+             (get_local $2)
+            )
+            (return
+             (get_local $3)
+            )
+           )
+           (i32.store8
+            (i32.add
+             (get_local $0)
+             (tee_local $2
+              (i32.sub
+               (get_local $2)
+               (i32.const 1)
+              )
+             )
+            )
+            (i32.load8_u
+             (i32.add
+              (get_local $1)
+              (get_local $2)
+             )
+            )
+           )
+          )
+          (br $continue|3)
+         )
+        )
+       )
+      )
+      (block $break|4
+       (loop $continue|4
+        (if
+         (i32.ge_u
+          (get_local $2)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+           (i64.store
+            (i32.add
+             (get_local $0)
+             (get_local $2)
+            )
+            (i64.load
+             (i32.add
+              (get_local $1)
+              (get_local $2)
+             )
+            )
+           )
+          )
+          (br $continue|4)
+         )
+        )
+       )
+      )
+     )
+    )
+    (block $break|5
+     (loop $continue|5
+      (if
+       (get_local $2)
+       (block
+        (i32.store8
+         (i32.add
+          (get_local $0)
+          (tee_local $2
+           (i32.sub
+            (get_local $2)
+            (i32.const 1)
+           )
+          )
+         )
+         (i32.load8_u
+          (i32.add
+           (get_local $1)
+           (get_local $2)
+          )
+         )
+        )
+        (br $continue|5)
+       )
+      )
+     )
+    )
+   )
+  )
+  (return
+   (get_local $3)
+  )
+ )
+ (func $start (; 1 ;) (type $v)
+  (i64.store
+   (i32.const 8)
+   (i64.const 1229782938247303441)
+  )
+  (i64.store
+   (i32.add
+    (i32.const 8)
+    (i32.const 8)
+   )
+   (i64.const 2459565876494606882)
+  )
+  (i64.store
+   (i32.add
+    (i32.const 8)
+    (i32.const 16)
+   )
+   (i64.const 3689348814741910323)
+  )
+  (i64.store
+   (i32.add
+    (i32.const 8)
+    (i32.const 24)
+   )
+   (i64.const 4919131752989213764)
+  )
+  (set_global $memmove/dest
+   (call $memmove/memmove
+    (i32.add
+     (i32.const 8)
+     (i32.const 1)
+    )
+    (i32.add
+     (i32.const 8)
+     (i32.const 16)
+    )
+    (i32.const 4)
+   )
+  )
+  (if
+   (i32.eqz
+    (i32.eq
+     (get_global $memmove/dest)
+     (i32.add
+      (i32.const 8)
+      (i32.const 1)
+     )
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.const 8)
+     )
+     (i64.const 1229783084848853777)
+    )
+   )
+   (unreachable)
+  )
+  (set_global $memmove/dest
+   (call $memmove/memmove
+    (i32.const 8)
+    (i32.const 8)
+    (i32.const 32)
+   )
+  )
+  (if
+   (i32.eqz
+    (i32.eq
+     (get_global $memmove/dest)
+     (i32.const 8)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.const 8)
+     )
+     (i64.const 1229783084848853777)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.add
+       (i32.const 8)
+       (i32.const 8)
+      )
+     )
+     (i64.const 2459565876494606882)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.add
+       (i32.const 8)
+       (i32.const 16)
+      )
+     )
+     (i64.const 3689348814741910323)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.add
+       (i32.const 8)
+       (i32.const 24)
+      )
+     )
+     (i64.const 4919131752989213764)
+    )
+   )
+   (unreachable)
+  )
+  (set_global $memmove/dest
+   (call $memmove/memmove
+    (i32.add
+     (i32.const 8)
+     (i32.const 5)
+    )
+    (i32.add
+     (i32.const 8)
+     (i32.const 28)
+    )
+    (i32.const 3)
+   )
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.const 8)
+     )
+     (i64.const 4919131679688438545)
+    )
+   )
+   (unreachable)
+  )
+  (set_global $memmove/dest
+   (call $memmove/memmove
+    (i32.add
+     (i32.const 8)
+     (i32.const 8)
+    )
+    (i32.add
+     (i32.const 8)
+     (i32.const 16)
+    )
+    (i32.const 15)
+   )
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.const 8)
+     )
+     (i64.const 4919131679688438545)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.add
+       (i32.const 8)
+       (i32.const 8)
+      )
+     )
+     (i64.const 3689348814741910323)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.add
+       (i32.const 8)
+       (i32.const 16)
+      )
+     )
+     (i64.const 3694152654344438852)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i64.eq
+     (i64.load
+      (i32.add
+       (i32.const 8)
+       (i32.const 24)
+      )
+     )
+     (i64.const 4919131752989213764)
+    )
+   )
+   (unreachable)
+  )
+ )
+)
+(;
+[program.elements]
+  GLOBAL: NaN
+  GLOBAL: Infinity
+  FUNCTION_PROTOTYPE: isNaN
+  FUNCTION_PROTOTYPE: isFinite
+  FUNCTION_PROTOTYPE: clz
+  FUNCTION_PROTOTYPE: ctz
+  FUNCTION_PROTOTYPE: popcnt
+  FUNCTION_PROTOTYPE: rotl
+  FUNCTION_PROTOTYPE: rotr
+  FUNCTION_PROTOTYPE: abs
+  FUNCTION_PROTOTYPE: max
+  FUNCTION_PROTOTYPE: min
+  FUNCTION_PROTOTYPE: ceil
+  FUNCTION_PROTOTYPE: floor
+  FUNCTION_PROTOTYPE: copysign
+  FUNCTION_PROTOTYPE: nearest
+  FUNCTION_PROTOTYPE: reinterpret
+  FUNCTION_PROTOTYPE: sqrt
+  FUNCTION_PROTOTYPE: trunc
+  FUNCTION_PROTOTYPE: load
+  FUNCTION_PROTOTYPE: store
+  FUNCTION_PROTOTYPE: sizeof
+  FUNCTION_PROTOTYPE: select
+  FUNCTION_PROTOTYPE: unreachable
+  FUNCTION_PROTOTYPE: current_memory
+  FUNCTION_PROTOTYPE: grow_memory
+  FUNCTION_PROTOTYPE: changetype
+  FUNCTION_PROTOTYPE: assert
+  FUNCTION_PROTOTYPE: i8
+  FUNCTION_PROTOTYPE: i16
+  FUNCTION_PROTOTYPE: i32
+  FUNCTION_PROTOTYPE: i64
+  FUNCTION_PROTOTYPE: u8
+  FUNCTION_PROTOTYPE: u16
+  FUNCTION_PROTOTYPE: u32
+  FUNCTION_PROTOTYPE: u64
+  FUNCTION_PROTOTYPE: bool
+  FUNCTION_PROTOTYPE: f32
+  FUNCTION_PROTOTYPE: f64
+  FUNCTION_PROTOTYPE: isize
+  FUNCTION_PROTOTYPE: usize
+  GLOBAL: HEAP_BASE
+  FUNCTION_PROTOTYPE: memmove/memmove
+  GLOBAL: memmove/base
+  GLOBAL: memmove/dest
+[program.exports]
+  
+;)
diff --git a/tests/compiler/memset.optimized.wast b/tests/compiler/memset.optimized.wast
new file mode 100644
index 00000000..a6456079
--- /dev/null
+++ b/tests/compiler/memset.optimized.wast
@@ -0,0 +1,447 @@
+(module
+ (type $iiii (func (param i32 i32 i32) (result i32)))
+ (type $v (func))
+ (global $memset/dest (mut i32) (i32.const 0))
+ (global $HEAP_BASE i32 (i32.const 4))
+ (memory $0 1)
+ (export "memory" (memory $0))
+ (start $start)
+ (func $memset/memset (; 0 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+  (local $3 i32)
+  (local $4 i64)
+  (local $5 i32)
+  (set_local $3
+   (get_local $0)
+  )
+  (if
+   (i32.eqz
+    (get_local $2)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store8
+   (get_local $0)
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 1)
+   )
+   (get_local $1)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 2)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 1)
+   )
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 3)
+   )
+   (get_local $1)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 6)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 3)
+   )
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 4)
+   )
+   (get_local $1)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 8)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store
+   (tee_local $0
+    (i32.add
+     (get_local $0)
+     (tee_local $5
+      (i32.and
+       (i32.sub
+        (i32.const 0)
+        (get_local $0)
+       )
+       (i32.const 3)
+      )
+     )
+    )
+   )
+   (tee_local $1
+    (i32.mul
+     (get_local $1)
+     (i32.const 16843009)
+    )
+   )
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (tee_local $2
+      (i32.and
+       (i32.sub
+        (get_local $2)
+        (get_local $5)
+       )
+       (i32.const -4)
+      )
+     )
+    )
+    (i32.const 4)
+   )
+   (get_local $1)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 8)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 4)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 8)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 12)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 8)
+   )
+   (get_local $1)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 24)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 12)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 16)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 20)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 24)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 28)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 24)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 20)
+   )
+   (get_local $1)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 16)
+   )
+   (get_local $1)
+  )
+  (set_local $0
+   (i32.add
+    (get_local $0)
+    (tee_local $5
+     (i32.add
+      (i32.and
+       (get_local $0)
+       (i32.const 4)
+      )
+      (i32.const 24)
+     )
+    )
+   )
+  )
+  (set_local $2
+   (i32.sub
+    (get_local $2)
+    (get_local $5)
+   )
+  )
+  (set_local $4
+   (i64.or
+    (i64.extend_u/i32
+     (get_local $1)
+    )
+    (i64.shl
+     (i64.extend_u/i32
+      (get_local $1)
+     )
+     (i64.const 32)
+    )
+   )
+  )
+  (loop $continue|0
+   (if
+    (i32.ge_u
+     (get_local $2)
+     (i32.const 32)
+    )
+    (block
+     (i64.store
+      (get_local $0)
+      (get_local $4)
+     )
+     (i64.store
+      (i32.add
+       (get_local $0)
+       (i32.const 8)
+      )
+      (get_local $4)
+     )
+     (i64.store
+      (i32.add
+       (get_local $0)
+       (i32.const 16)
+      )
+      (get_local $4)
+     )
+     (i64.store
+      (i32.add
+       (get_local $0)
+       (i32.const 24)
+      )
+      (get_local $4)
+     )
+     (set_local $2
+      (i32.sub
+       (get_local $2)
+       (i32.const 32)
+      )
+     )
+     (set_local $0
+      (i32.add
+       (get_local $0)
+       (i32.const 32)
+      )
+     )
+     (br $continue|0)
+    )
+   )
+  )
+  (get_local $3)
+ )
+ (func $start (; 1 ;) (type $v)
+  (set_global $memset/dest
+   (get_global $HEAP_BASE)
+  )
+  (drop
+   (call $memset/memset
+    (get_global $memset/dest)
+    (i32.const 1)
+    (i32.const 16)
+   )
+  )
+  (if
+   (i32.ne
+    (i32.load8_u
+     (get_global $memset/dest)
+    )
+    (i32.const 1)
+   )
+   (unreachable)
+  )
+  (if
+   (i32.ne
+    (i32.load8_u
+     (i32.add
+      (get_global $memset/dest)
+      (i32.const 15)
+     )
+    )
+    (i32.const 1)
+   )
+   (unreachable)
+  )
+  (drop
+   (call $memset/memset
+    (i32.add
+     (get_global $memset/dest)
+     (i32.const 1)
+    )
+    (i32.const 2)
+    (i32.const 14)
+   )
+  )
+  (if
+   (i32.ne
+    (i32.load8_u
+     (get_global $memset/dest)
+    )
+    (i32.const 1)
+   )
+   (unreachable)
+  )
+  (if
+   (i32.ne
+    (i32.load8_u
+     (i32.add
+      (get_global $memset/dest)
+      (i32.const 1)
+     )
+    )
+    (i32.const 2)
+   )
+   (unreachable)
+  )
+  (if
+   (i32.ne
+    (i32.load8_u
+     (i32.add
+      (get_global $memset/dest)
+      (i32.const 14)
+     )
+    )
+    (i32.const 2)
+   )
+   (unreachable)
+  )
+  (if
+   (i32.ne
+    (i32.load8_u
+     (i32.add
+      (get_global $memset/dest)
+      (i32.const 15)
+     )
+    )
+    (i32.const 1)
+   )
+   (unreachable)
+  )
+ )
+)
diff --git a/tests/compiler/memset.ts b/tests/compiler/memset.ts
new file mode 100644
index 00000000..085eaee3
--- /dev/null
+++ b/tests/compiler/memset.ts
@@ -0,0 +1,80 @@
+function memset(dest: usize, c: u8, n: usize): usize {
+  var ret = dest;
+
+  // fill head and tail wwith minimal branching
+  if (!n)
+    return ret;
+  store<u8>(dest, c);
+  store<u8>(dest + n - 1, c);
+  if (n <= 2)
+    return ret;
+
+  store<u8>(dest + 1, c);
+  store<u8>(dest + 2, c);
+  store<u8>(dest + n - 2, c);
+  store<u8>(dest + n - 3, c);
+  if (n <= 6)
+    return ret;
+  store<u8>(dest + 3, c);
+  store<u8>(dest + n - 4, c);
+  if (n <= 8)
+    return ret;
+
+  // advance pointer to align it at 4-byte boundary
+  var k: usize = -dest & 3;
+  dest += k;
+  n -= k;
+  n &= -4;
+
+  var c32: u32 = -1 / 255 * c;
+
+  // fill head/tail up to 28 bytes each in preparation
+  store<u32>(dest, c32);
+  store<u32>(dest + n - 4, c32);
+  if (n <= 8)
+    return ret;
+  store<u32>(dest + 4, c32);
+  store<u32>(dest + 8, c32);
+  store<u32>(dest + n - 12, c32);
+  store<u32>(dest + n - 8, c32);
+  if (n <= 24)
+    return ret;
+  store<u32>(dest + 12, c32);
+  store<u32>(dest + 16, c32);
+  store<u32>(dest + 20, c32);
+  store<u32>(dest + 24, c32);
+  store<u32>(dest + n - 28, c32);
+  store<u32>(dest + n - 24, c32);
+  store<u32>(dest + n - 20, c32);
+  store<u32>(dest + n - 16, c32);
+
+  // align to a multiple of 8
+  k = 24 + (dest & 4);
+  dest += k;
+  n -= k;
+
+  // copy 32 bytes each
+  var c64: u64 = <u64>c32 | (<u64>c32 << 32);
+  while (n >= 32) {
+    store<u64>(dest, c64);
+    store<u64>(dest + 8, c64);
+    store<u64>(dest + 16, c64);
+    store<u64>(dest + 24, c64);
+    n -= 32;
+    dest += 32;
+  }
+  return ret;
+}
+
+var dest = HEAP_BASE;
+memset(dest, 1, 16);
+
+assert(load<u8>(dest) == 1);
+assert(load<u8>(dest + 15) == 1);
+
+memset(dest + 1, 2, 14);
+
+assert(load<u8>(dest) == 1);
+assert(load<u8>(dest + 1) == 2);
+assert(load<u8>(dest + 14) == 2);
+assert(load<u8>(dest + 15) == 1);
diff --git a/tests/compiler/memset.wast b/tests/compiler/memset.wast
new file mode 100644
index 00000000..2de0d6a9
--- /dev/null
+++ b/tests/compiler/memset.wast
@@ -0,0 +1,540 @@
+(module
+ (type $iiii (func (param i32 i32 i32) (result i32)))
+ (type $v (func))
+ (global $memset/dest (mut i32) (i32.const 0))
+ (global $HEAP_BASE i32 (i32.const 4))
+ (memory $0 1)
+ (export "memory" (memory $0))
+ (start $start)
+ (func $memset/memset (; 0 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+  (local $3 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (local $6 i64)
+  (block
+   (set_local $3
+    (get_local $0)
+   )
+  )
+  (if
+   (i32.eqz
+    (get_local $2)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store8
+   (get_local $0)
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 1)
+   )
+   (get_local $1)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 2)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 1)
+   )
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 3)
+   )
+   (get_local $1)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 6)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 3)
+   )
+   (get_local $1)
+  )
+  (i32.store8
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 4)
+   )
+   (get_local $1)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 8)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (block
+   (set_local $4
+    (i32.and
+     (i32.sub
+      (i32.const 0)
+      (get_local $0)
+     )
+     (i32.const 3)
+    )
+   )
+  )
+  (set_local $0
+   (i32.add
+    (get_local $0)
+    (get_local $4)
+   )
+  )
+  (set_local $2
+   (i32.sub
+    (get_local $2)
+    (get_local $4)
+   )
+  )
+  (set_local $2
+   (i32.and
+    (get_local $2)
+    (i32.sub
+     (i32.const 0)
+     (i32.const 4)
+    )
+   )
+  )
+  (block
+   (set_local $5
+    (i32.mul
+     (i32.div_u
+      (i32.sub
+       (i32.const 0)
+       (i32.const 1)
+      )
+      (i32.const 255)
+     )
+     (get_local $1)
+    )
+   )
+  )
+  (i32.store
+   (get_local $0)
+   (get_local $5)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 4)
+   )
+   (get_local $5)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 8)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 4)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 8)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 12)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 8)
+   )
+   (get_local $5)
+  )
+  (if
+   (i32.le_u
+    (get_local $2)
+    (i32.const 24)
+   )
+   (return
+    (get_local $3)
+   )
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 12)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 16)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 20)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.add
+    (get_local $0)
+    (i32.const 24)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 28)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 24)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 20)
+   )
+   (get_local $5)
+  )
+  (i32.store
+   (i32.sub
+    (i32.add
+     (get_local $0)
+     (get_local $2)
+    )
+    (i32.const 16)
+   )
+   (get_local $5)
+  )
+  (set_local $4
+   (i32.add
+    (i32.const 24)
+    (i32.and
+     (get_local $0)
+     (i32.const 4)
+    )
+   )
+  )
+  (set_local $0
+   (i32.add
+    (get_local $0)
+    (get_local $4)
+   )
+  )
+  (set_local $2
+   (i32.sub
+    (get_local $2)
+    (get_local $4)
+   )
+  )
+  (block
+   (set_local $6
+    (i64.or
+     (i64.extend_u/i32
+      (get_local $5)
+     )
+     (i64.shl
+      (i64.extend_u/i32
+       (get_local $5)
+      )
+      (i64.const 32)
+     )
+    )
+   )
+  )
+  (block $break|0
+   (loop $continue|0
+    (if
+     (i32.ge_u
+      (get_local $2)
+      (i32.const 32)
+     )
+     (block
+      (block
+       (i64.store
+        (get_local $0)
+        (get_local $6)
+       )
+       (i64.store
+        (i32.add
+         (get_local $0)
+         (i32.const 8)
+        )
+        (get_local $6)
+       )
+       (i64.store
+        (i32.add
+         (get_local $0)
+         (i32.const 16)
+        )
+        (get_local $6)
+       )
+       (i64.store
+        (i32.add
+         (get_local $0)
+         (i32.const 24)
+        )
+        (get_local $6)
+       )
+       (set_local $2
+        (i32.sub
+         (get_local $2)
+         (i32.const 32)
+        )
+       )
+       (set_local $0
+        (i32.add
+         (get_local $0)
+         (i32.const 32)
+        )
+       )
+      )
+      (br $continue|0)
+     )
+    )
+   )
+  )
+  (return
+   (get_local $3)
+  )
+ )
+ (func $start (; 1 ;) (type $v)
+  (set_global $memset/dest
+   (get_global $HEAP_BASE)
+  )
+  (drop
+   (call $memset/memset
+    (get_global $memset/dest)
+    (i32.const 1)
+    (i32.const 16)
+   )
+  )
+  (if
+   (i32.eqz
+    (i32.eq
+     (i32.load8_u
+      (get_global $memset/dest)
+     )
+     (i32.const 1)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i32.eq
+     (i32.load8_u
+      (i32.add
+       (get_global $memset/dest)
+       (i32.const 15)
+      )
+     )
+     (i32.const 1)
+    )
+   )
+   (unreachable)
+  )
+  (drop
+   (call $memset/memset
+    (i32.add
+     (get_global $memset/dest)
+     (i32.const 1)
+    )
+    (i32.const 2)
+    (i32.const 14)
+   )
+  )
+  (if
+   (i32.eqz
+    (i32.eq
+     (i32.load8_u
+      (get_global $memset/dest)
+     )
+     (i32.const 1)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i32.eq
+     (i32.load8_u
+      (i32.add
+       (get_global $memset/dest)
+       (i32.const 1)
+      )
+     )
+     (i32.const 2)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i32.eq
+     (i32.load8_u
+      (i32.add
+       (get_global $memset/dest)
+       (i32.const 14)
+      )
+     )
+     (i32.const 2)
+    )
+   )
+   (unreachable)
+  )
+  (if
+   (i32.eqz
+    (i32.eq
+     (i32.load8_u
+      (i32.add
+       (get_global $memset/dest)
+       (i32.const 15)
+      )
+     )
+     (i32.const 1)
+    )
+   )
+   (unreachable)
+  )
+ )
+)
+(;
+[program.elements]
+  GLOBAL: NaN
+  GLOBAL: Infinity
+  FUNCTION_PROTOTYPE: isNaN
+  FUNCTION_PROTOTYPE: isFinite
+  FUNCTION_PROTOTYPE: clz
+  FUNCTION_PROTOTYPE: ctz
+  FUNCTION_PROTOTYPE: popcnt
+  FUNCTION_PROTOTYPE: rotl
+  FUNCTION_PROTOTYPE: rotr
+  FUNCTION_PROTOTYPE: abs
+  FUNCTION_PROTOTYPE: max
+  FUNCTION_PROTOTYPE: min
+  FUNCTION_PROTOTYPE: ceil
+  FUNCTION_PROTOTYPE: floor
+  FUNCTION_PROTOTYPE: copysign
+  FUNCTION_PROTOTYPE: nearest
+  FUNCTION_PROTOTYPE: reinterpret
+  FUNCTION_PROTOTYPE: sqrt
+  FUNCTION_PROTOTYPE: trunc
+  FUNCTION_PROTOTYPE: load
+  FUNCTION_PROTOTYPE: store
+  FUNCTION_PROTOTYPE: sizeof
+  FUNCTION_PROTOTYPE: select
+  FUNCTION_PROTOTYPE: unreachable
+  FUNCTION_PROTOTYPE: current_memory
+  FUNCTION_PROTOTYPE: grow_memory
+  FUNCTION_PROTOTYPE: changetype
+  FUNCTION_PROTOTYPE: assert
+  FUNCTION_PROTOTYPE: i8
+  FUNCTION_PROTOTYPE: i16
+  FUNCTION_PROTOTYPE: i32
+  FUNCTION_PROTOTYPE: i64
+  FUNCTION_PROTOTYPE: u8
+  FUNCTION_PROTOTYPE: u16
+  FUNCTION_PROTOTYPE: u32
+  FUNCTION_PROTOTYPE: u64
+  FUNCTION_PROTOTYPE: bool
+  FUNCTION_PROTOTYPE: f32
+  FUNCTION_PROTOTYPE: f64
+  FUNCTION_PROTOTYPE: isize
+  FUNCTION_PROTOTYPE: usize
+  GLOBAL: HEAP_BASE
+  FUNCTION_PROTOTYPE: memset/memset
+  GLOBAL: memset/dest
+[program.exports]
+  
+;)
diff --git a/tests/compiler/showcase.optimized-inlined.wast b/tests/compiler/showcase.optimized-inlined.wast
index 69bdbde7..2d2d982f 100644
--- a/tests/compiler/showcase.optimized-inlined.wast
+++ b/tests/compiler/showcase.optimized-inlined.wast
@@ -57,44 +57,41 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  (set_local $4
+  (set_local $5
    (get_local $0)
   )
-  (set_local $3
-   (get_local $1)
-  )
   (loop $continue|0
    (if
     (if (result i32)
      (get_local $2)
      (i32.rem_u
-      (get_local $3)
+      (get_local $1)
       (i32.const 4)
      )
      (get_local $2)
     )
     (block
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
@@ -111,7 +108,7 @@
   (if
    (i32.eqz
     (i32.rem_u
-     (get_local $4)
+     (get_local $0)
      (i32.const 4)
     )
    )
@@ -124,56 +121,56 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.load
-         (get_local $3)
+         (get_local $1)
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 4)
          )
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 8)
          )
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 12)
          )
         )
        )
-       (set_local $3
+       (set_local $1
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -194,32 +191,32 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 4)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 8)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 8)
        )
       )
@@ -232,20 +229,20 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 4)
        )
       )
@@ -258,20 +255,20 @@
      )
      (block
       (i32.store16
-       (get_local $4)
+       (get_local $0)
        (i32.load16_u
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 2)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 2)
        )
       )
@@ -283,24 +280,24 @@
       (i32.const 1)
      )
      (block
-      (set_local $1
-       (get_local $4)
+      (set_local $3
+       (get_local $0)
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $1
-         (get_local $3)
+        (set_local $3
+         (get_local $1)
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
      )
     )
     (return
-     (get_local $0)
+     (get_local $5)
     )
    )
   )
@@ -317,7 +314,7 @@
         (br_table $case0|2 $case1|2 $case2|2 $tablify|0
          (i32.sub
           (i32.rem_u
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.const 1)
@@ -326,80 +323,80 @@
        )
        (br $break|2)
       )
-      (set_local $5
+      (set_local $4
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
@@ -417,17 +414,17 @@
         )
         (block
          (i32.store
-          (get_local $4)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $1
+            (tee_local $3
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 1)
               )
              )
@@ -438,19 +435,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.or
            (i32.shr_u
-            (get_local $1)
+            (get_local $3)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 5)
               )
              )
@@ -461,19 +458,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $1
+            (tee_local $3
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 9)
               )
              )
@@ -484,19 +481,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.or
            (i32.shr_u
-            (get_local $1)
+            (get_local $3)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 13)
               )
              )
@@ -505,15 +502,15 @@
            )
           )
          )
-         (set_local $3
+         (set_local $1
           (i32.add
-           (get_local $3)
+           (get_local $1)
            (i32.const 16)
           )
          )
-         (set_local $4
+         (set_local $0
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -529,56 +526,56 @@
       )
       (br $break|2)
      )
-     (set_local $5
+     (set_local $4
       (i32.load
-       (get_local $3)
+       (get_local $1)
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
@@ -596,17 +593,17 @@
        )
        (block
         (i32.store
-         (get_local $4)
+         (get_local $0)
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $1
+           (tee_local $3
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 2)
              )
             )
@@ -617,19 +614,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 4)
          )
          (i32.or
           (i32.shr_u
-           (get_local $1)
+           (get_local $3)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 6)
              )
             )
@@ -640,19 +637,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 8)
          )
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $1
+           (tee_local $3
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 10)
              )
             )
@@ -663,19 +660,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 12)
          )
          (i32.or
           (i32.shr_u
-           (get_local $1)
+           (get_local $3)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 14)
              )
             )
@@ -684,15 +681,15 @@
           )
          )
         )
-        (set_local $3
+        (set_local $1
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 16)
          )
         )
-        (set_local $4
+        (set_local $0
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 16)
          )
         )
@@ -708,32 +705,32 @@
      )
      (br $break|2)
     )
-    (set_local $5
+    (set_local $4
      (i32.load
-      (get_local $3)
+      (get_local $1)
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -751,17 +748,17 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $1
+          (tee_local $3
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 3)
             )
            )
@@ -772,19 +769,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.or
          (i32.shr_u
-          (get_local $1)
+          (get_local $3)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 7)
             )
            )
@@ -795,19 +792,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $1
+          (tee_local $3
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 11)
             )
            )
@@ -818,19 +815,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.or
          (i32.shr_u
-          (get_local $1)
+          (get_local $3)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 15)
             )
            )
@@ -839,15 +836,15 @@
          )
         )
        )
-       (set_local $3
+       (set_local $1
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -869,387 +866,387 @@
     (i32.const 16)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1261,195 +1258,195 @@
     (i32.const 8)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1461,99 +1458,99 @@
     (i32.const 4)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1565,51 +1562,51 @@
     (i32.const 2)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1621,23 +1618,23 @@
     (i32.const 1)
    )
    (block
-    (set_local $1
-     (get_local $4)
+    (set_local $3
+     (get_local $0)
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $1
-       (get_local $3)
+      (set_local $3
+       (get_local $1)
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
    )
   )
-  (get_local $0)
+  (get_local $5)
  )
  (func $fmod/fmod (; 2 ;) (type $FFF) (param $0 f64) (param $1 f64) (result f64)
   (local $2 i64)
diff --git a/tests/compiler/showcase.optimized.wast b/tests/compiler/showcase.optimized.wast
index aeda3181..818da7ea 100644
--- a/tests/compiler/showcase.optimized.wast
+++ b/tests/compiler/showcase.optimized.wast
@@ -78,44 +78,41 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  (set_local $4
+  (set_local $5
    (get_local $0)
   )
-  (set_local $3
-   (get_local $1)
-  )
   (loop $continue|0
    (if
     (if (result i32)
      (get_local $2)
      (i32.rem_u
-      (get_local $3)
+      (get_local $1)
       (i32.const 4)
      )
      (get_local $2)
     )
     (block
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
@@ -132,7 +129,7 @@
   (if
    (i32.eqz
     (i32.rem_u
-     (get_local $4)
+     (get_local $0)
      (i32.const 4)
     )
    )
@@ -145,56 +142,56 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.load
-         (get_local $3)
+         (get_local $1)
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 4)
          )
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 8)
          )
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.load
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 12)
          )
         )
        )
-       (set_local $3
+       (set_local $1
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -215,32 +212,32 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 4)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 8)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 8)
        )
       )
@@ -253,20 +250,20 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 4)
        )
       )
@@ -279,20 +276,20 @@
      )
      (block
       (i32.store16
-       (get_local $4)
+       (get_local $0)
        (i32.load16_u
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 2)
        )
       )
-      (set_local $3
+      (set_local $1
        (i32.add
-        (get_local $3)
+        (get_local $1)
         (i32.const 2)
        )
       )
@@ -304,24 +301,24 @@
       (i32.const 1)
      )
      (block
-      (set_local $1
-       (get_local $4)
+      (set_local $3
+       (get_local $0)
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $1
-         (get_local $3)
+        (set_local $3
+         (get_local $1)
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
      )
     )
     (return
-     (get_local $0)
+     (get_local $5)
     )
    )
   )
@@ -338,7 +335,7 @@
         (br_table $case0|2 $case1|2 $case2|2 $tablify|0
          (i32.sub
           (i32.rem_u
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.const 1)
@@ -347,80 +344,80 @@
        )
        (br $break|2)
       )
-      (set_local $5
+      (set_local $4
        (i32.load
-        (get_local $3)
+        (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (tee_local $1
-         (get_local $4)
+        (tee_local $3
+         (get_local $0)
         )
         (i32.const 1)
        )
       )
       (i32.store8
-       (get_local $1)
+       (get_local $3)
        (block (result i32)
-        (set_local $3
+        (set_local $1
          (i32.add
-          (tee_local $1
-           (get_local $3)
+          (tee_local $3
+           (get_local $1)
           )
           (i32.const 1)
          )
         )
         (i32.load8_u
-         (get_local $1)
+         (get_local $3)
         )
        )
       )
@@ -438,17 +435,17 @@
         )
         (block
          (i32.store
-          (get_local $4)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $1
+            (tee_local $3
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 1)
               )
              )
@@ -459,19 +456,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.or
            (i32.shr_u
-            (get_local $1)
+            (get_local $3)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 5)
               )
              )
@@ -482,19 +479,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $1
+            (tee_local $3
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 9)
               )
              )
@@ -505,19 +502,19 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.or
            (i32.shr_u
-            (get_local $1)
+            (get_local $3)
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
-               (get_local $3)
+               (get_local $1)
                (i32.const 13)
               )
              )
@@ -526,15 +523,15 @@
            )
           )
          )
-         (set_local $3
+         (set_local $1
           (i32.add
-           (get_local $3)
+           (get_local $1)
            (i32.const 16)
           )
          )
-         (set_local $4
+         (set_local $0
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -550,56 +547,56 @@
       )
       (br $break|2)
      )
-     (set_local $5
+     (set_local $4
       (i32.load
-       (get_local $3)
+       (get_local $1)
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
-       (tee_local $1
-        (get_local $4)
+       (tee_local $3
+        (get_local $0)
        )
        (i32.const 1)
       )
      )
      (i32.store8
-      (get_local $1)
+      (get_local $3)
       (block (result i32)
-       (set_local $3
+       (set_local $1
         (i32.add
-         (tee_local $1
-          (get_local $3)
+         (tee_local $3
+          (get_local $1)
          )
          (i32.const 1)
         )
        )
        (i32.load8_u
-        (get_local $1)
+        (get_local $3)
        )
       )
      )
@@ -617,17 +614,17 @@
        )
        (block
         (i32.store
-         (get_local $4)
+         (get_local $0)
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $1
+           (tee_local $3
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 2)
              )
             )
@@ -638,19 +635,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 4)
          )
          (i32.or
           (i32.shr_u
-           (get_local $1)
+           (get_local $3)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 6)
              )
             )
@@ -661,19 +658,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 8)
          )
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $1
+           (tee_local $3
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 10)
              )
             )
@@ -684,19 +681,19 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 12)
          )
          (i32.or
           (i32.shr_u
-           (get_local $1)
+           (get_local $3)
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
-              (get_local $3)
+              (get_local $1)
               (i32.const 14)
              )
             )
@@ -705,15 +702,15 @@
           )
          )
         )
-        (set_local $3
+        (set_local $1
          (i32.add
-          (get_local $3)
+          (get_local $1)
           (i32.const 16)
          )
         )
-        (set_local $4
+        (set_local $0
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 16)
          )
         )
@@ -729,32 +726,32 @@
      )
      (br $break|2)
     )
-    (set_local $5
+    (set_local $4
      (i32.load
-      (get_local $3)
+      (get_local $1)
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -772,17 +769,17 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $1
+          (tee_local $3
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 3)
             )
            )
@@ -793,19 +790,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.or
          (i32.shr_u
-          (get_local $1)
+          (get_local $3)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 7)
             )
            )
@@ -816,19 +813,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $1
+          (tee_local $3
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 11)
             )
            )
@@ -839,19 +836,19 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.or
          (i32.shr_u
-          (get_local $1)
+          (get_local $3)
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
-             (get_local $3)
+             (get_local $1)
              (i32.const 15)
             )
            )
@@ -860,15 +857,15 @@
          )
         )
        )
-       (set_local $3
+       (set_local $1
         (i32.add
-         (get_local $3)
+         (get_local $1)
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -890,387 +887,387 @@
     (i32.const 16)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1282,195 +1279,195 @@
     (i32.const 8)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1482,99 +1479,99 @@
     (i32.const 4)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1586,51 +1583,51 @@
     (i32.const 2)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
-      (tee_local $1
-       (get_local $4)
+      (tee_local $3
+       (get_local $0)
       )
       (i32.const 1)
      )
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $3
+      (set_local $1
        (i32.add
-        (tee_local $1
-         (get_local $3)
+        (tee_local $3
+         (get_local $1)
         )
         (i32.const 1)
        )
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
@@ -1642,23 +1639,23 @@
     (i32.const 1)
    )
    (block
-    (set_local $1
-     (get_local $4)
+    (set_local $3
+     (get_local $0)
     )
     (i32.store8
-     (get_local $1)
+     (get_local $3)
      (block (result i32)
-      (set_local $1
-       (get_local $3)
+      (set_local $3
+       (get_local $1)
       )
       (i32.load8_u
-       (get_local $1)
+       (get_local $3)
       )
      )
     )
    )
   )
-  (get_local $0)
+  (get_local $5)
  )
  (func $fmod/fmod (; 6 ;) (type $FFF) (param $0 f64) (param $1 f64) (result f64)
   (local $2 i64)
diff --git a/tests/compiler/showcase.wast b/tests/compiler/showcase.wast
index 4885d267..6b2b6940 100644
--- a/tests/compiler/showcase.wast
+++ b/tests/compiler/showcase.wast
@@ -121,14 +121,10 @@
   (local $4 i32)
   (local $5 i32)
   (local $6 i32)
-  (local $7 i32)
   (block
    (set_local $3
     (get_local $0)
    )
-   (set_local $4
-    (get_local $1)
-   )
   )
   (nop)
   (block $break|0
@@ -140,7 +136,7 @@
        (i32.const 0)
       )
       (i32.rem_u
-       (get_local $4)
+       (get_local $1)
        (i32.const 4)
       )
       (get_local $2)
@@ -149,29 +145,29 @@
       (block
        (i32.store8
         (block (result i32)
-         (set_local $7
-          (get_local $3)
+         (set_local $6
+          (get_local $0)
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $7)
+           (get_local $6)
            (i32.const 1)
           )
          )
-         (get_local $7)
+         (get_local $6)
         )
         (i32.load8_u
          (block (result i32)
-          (set_local $7
-           (get_local $4)
+          (set_local $6
+           (get_local $1)
           )
-          (set_local $4
+          (set_local $1
            (i32.add
-            (get_local $7)
+            (get_local $6)
             (i32.const 1)
            )
           )
-          (get_local $7)
+          (get_local $6)
          )
         )
        )
@@ -190,7 +186,7 @@
   (if
    (i32.eq
     (i32.rem_u
-     (get_local $3)
+     (get_local $0)
      (i32.const 4)
     )
     (i32.const 0)
@@ -206,56 +202,56 @@
        (block
         (block
          (i32.store
-          (get_local $3)
+          (get_local $0)
           (i32.load
-           (get_local $4)
+           (get_local $1)
           )
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.load
            (i32.add
-            (get_local $4)
+            (get_local $1)
             (i32.const 4)
            )
           )
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.load
            (i32.add
-            (get_local $4)
+            (get_local $1)
             (i32.const 8)
            )
           )
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.load
            (i32.add
-            (get_local $4)
+            (get_local $1)
             (i32.const 12)
            )
           )
          )
-         (set_local $4
+         (set_local $1
           (i32.add
-           (get_local $4)
+           (get_local $1)
            (i32.const 16)
           )
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -278,32 +274,32 @@
      )
      (block
       (i32.store
-       (get_local $3)
+       (get_local $0)
        (i32.load
-        (get_local $4)
+        (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
         (i32.add
-         (get_local $4)
+         (get_local $1)
          (i32.const 4)
         )
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 8)
        )
       )
-      (set_local $4
+      (set_local $1
        (i32.add
-        (get_local $4)
+        (get_local $1)
         (i32.const 8)
        )
       )
@@ -316,20 +312,20 @@
      )
      (block
       (i32.store
-       (get_local $3)
+       (get_local $0)
        (i32.load
-        (get_local $4)
+        (get_local $1)
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 4)
        )
       )
-      (set_local $4
+      (set_local $1
        (i32.add
-        (get_local $4)
+        (get_local $1)
         (i32.const 4)
        )
       )
@@ -342,20 +338,20 @@
      )
      (block
       (i32.store16
-       (get_local $3)
+       (get_local $0)
        (i32.load16_u
-        (get_local $4)
+        (get_local $1)
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 2)
        )
       )
-      (set_local $4
+      (set_local $1
        (i32.add
-        (get_local $4)
+        (get_local $1)
         (i32.const 2)
        )
       )
@@ -368,35 +364,35 @@
      )
      (i32.store8
       (block (result i32)
-       (set_local $7
-        (get_local $3)
+       (set_local $6
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $7
-         (get_local $4)
+        (set_local $6
+         (get_local $1)
         )
-        (set_local $4
+        (set_local $1
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
       )
      )
     )
     (return
-     (get_local $0)
+     (get_local $3)
     )
    )
   )
@@ -409,118 +405,118 @@
     (block $case2|2
      (block $case1|2
       (block $case0|2
-       (set_local $7
+       (set_local $6
         (i32.rem_u
-         (get_local $3)
+         (get_local $0)
          (i32.const 4)
         )
        )
        (br_if $case0|2
         (i32.eq
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
        (br_if $case1|2
         (i32.eq
-         (get_local $7)
+         (get_local $6)
          (i32.const 2)
         )
        )
        (br_if $case2|2
         (i32.eq
-         (get_local $7)
+         (get_local $6)
          (i32.const 3)
         )
        )
        (br $break|2)
       )
-      (set_local $5
+      (set_local $4
        (i32.load
-        (get_local $4)
+        (get_local $1)
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $7
-         (get_local $3)
+        (set_local $6
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $7
-          (get_local $4)
+         (set_local $6
+          (get_local $1)
          )
-         (set_local $4
+         (set_local $1
           (i32.add
-           (get_local $7)
+           (get_local $6)
            (i32.const 1)
           )
          )
-         (get_local $7)
+         (get_local $6)
         )
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $7
-         (get_local $3)
+        (set_local $6
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $7
-          (get_local $4)
+         (set_local $6
+          (get_local $1)
          )
-         (set_local $4
+         (set_local $1
           (i32.add
-           (get_local $7)
+           (get_local $6)
            (i32.const 1)
           )
          )
-         (get_local $7)
+         (get_local $6)
         )
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $7
-         (get_local $3)
+        (set_local $6
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $7
-          (get_local $4)
+         (set_local $6
+          (get_local $1)
          )
-         (set_local $4
+         (set_local $1
           (i32.add
-           (get_local $7)
+           (get_local $6)
            (i32.const 1)
           )
          )
-         (get_local $7)
+         (get_local $6)
         )
        )
       )
@@ -539,91 +535,19 @@
          )
          (block
           (block
-           (set_local $6
+           (set_local $5
             (i32.load
              (i32.add
-              (get_local $4)
+              (get_local $1)
               (i32.const 1)
              )
             )
            )
            (i32.store
-            (get_local $3)
+            (get_local $0)
             (i32.or
              (i32.shr_u
-              (get_local $5)
-              (i32.const 24)
-             )
-             (i32.shl
-              (get_local $6)
-              (i32.const 8)
-             )
-            )
-           )
-           (set_local $5
-            (i32.load
-             (i32.add
               (get_local $4)
-              (i32.const 5)
-             )
-            )
-           )
-           (i32.store
-            (i32.add
-             (get_local $3)
-             (i32.const 4)
-            )
-            (i32.or
-             (i32.shr_u
-              (get_local $6)
-              (i32.const 24)
-             )
-             (i32.shl
-              (get_local $5)
-              (i32.const 8)
-             )
-            )
-           )
-           (set_local $6
-            (i32.load
-             (i32.add
-              (get_local $4)
-              (i32.const 9)
-             )
-            )
-           )
-           (i32.store
-            (i32.add
-             (get_local $3)
-             (i32.const 8)
-            )
-            (i32.or
-             (i32.shr_u
-              (get_local $5)
-              (i32.const 24)
-             )
-             (i32.shl
-              (get_local $6)
-              (i32.const 8)
-             )
-            )
-           )
-           (set_local $5
-            (i32.load
-             (i32.add
-              (get_local $4)
-              (i32.const 13)
-             )
-            )
-           )
-           (i32.store
-            (i32.add
-             (get_local $3)
-             (i32.const 12)
-            )
-            (i32.or
-             (i32.shr_u
-              (get_local $6)
               (i32.const 24)
              )
              (i32.shl
@@ -633,14 +557,86 @@
             )
            )
            (set_local $4
+            (i32.load
+             (i32.add
+              (get_local $1)
+              (i32.const 5)
+             )
+            )
+           )
+           (i32.store
             (i32.add
-             (get_local $4)
+             (get_local $0)
+             (i32.const 4)
+            )
+            (i32.or
+             (i32.shr_u
+              (get_local $5)
+              (i32.const 24)
+             )
+             (i32.shl
+              (get_local $4)
+              (i32.const 8)
+             )
+            )
+           )
+           (set_local $5
+            (i32.load
+             (i32.add
+              (get_local $1)
+              (i32.const 9)
+             )
+            )
+           )
+           (i32.store
+            (i32.add
+             (get_local $0)
+             (i32.const 8)
+            )
+            (i32.or
+             (i32.shr_u
+              (get_local $4)
+              (i32.const 24)
+             )
+             (i32.shl
+              (get_local $5)
+              (i32.const 8)
+             )
+            )
+           )
+           (set_local $4
+            (i32.load
+             (i32.add
+              (get_local $1)
+              (i32.const 13)
+             )
+            )
+           )
+           (i32.store
+            (i32.add
+             (get_local $0)
+             (i32.const 12)
+            )
+            (i32.or
+             (i32.shr_u
+              (get_local $5)
+              (i32.const 24)
+             )
+             (i32.shl
+              (get_local $4)
+              (i32.const 8)
+             )
+            )
+           )
+           (set_local $1
+            (i32.add
+             (get_local $1)
              (i32.const 16)
             )
            )
-           (set_local $3
+           (set_local $0
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 16)
             )
            )
@@ -658,64 +654,64 @@
       )
       (br $break|2)
      )
-     (set_local $5
+     (set_local $4
       (i32.load
-       (get_local $4)
+       (get_local $1)
       )
      )
      (i32.store8
       (block (result i32)
-       (set_local $7
-        (get_local $3)
+       (set_local $6
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $7
-         (get_local $4)
+        (set_local $6
+         (get_local $1)
         )
-        (set_local $4
+        (set_local $1
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
       )
      )
      (i32.store8
       (block (result i32)
-       (set_local $7
-        (get_local $3)
+       (set_local $6
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $7
-         (get_local $4)
+        (set_local $6
+         (get_local $1)
         )
-        (set_local $4
+        (set_local $1
          (i32.add
-          (get_local $7)
+          (get_local $6)
           (i32.const 1)
          )
         )
-        (get_local $7)
+        (get_local $6)
        )
       )
      )
@@ -734,91 +730,19 @@
         )
         (block
          (block
-          (set_local $6
+          (set_local $5
            (i32.load
             (i32.add
-             (get_local $4)
+             (get_local $1)
              (i32.const 2)
             )
            )
           )
           (i32.store
-           (get_local $3)
+           (get_local $0)
            (i32.or
             (i32.shr_u
-             (get_local $5)
-             (i32.const 16)
-            )
-            (i32.shl
-             (get_local $6)
-             (i32.const 16)
-            )
-           )
-          )
-          (set_local $5
-           (i32.load
-            (i32.add
              (get_local $4)
-             (i32.const 6)
-            )
-           )
-          )
-          (i32.store
-           (i32.add
-            (get_local $3)
-            (i32.const 4)
-           )
-           (i32.or
-            (i32.shr_u
-             (get_local $6)
-             (i32.const 16)
-            )
-            (i32.shl
-             (get_local $5)
-             (i32.const 16)
-            )
-           )
-          )
-          (set_local $6
-           (i32.load
-            (i32.add
-             (get_local $4)
-             (i32.const 10)
-            )
-           )
-          )
-          (i32.store
-           (i32.add
-            (get_local $3)
-            (i32.const 8)
-           )
-           (i32.or
-            (i32.shr_u
-             (get_local $5)
-             (i32.const 16)
-            )
-            (i32.shl
-             (get_local $6)
-             (i32.const 16)
-            )
-           )
-          )
-          (set_local $5
-           (i32.load
-            (i32.add
-             (get_local $4)
-             (i32.const 14)
-            )
-           )
-          )
-          (i32.store
-           (i32.add
-            (get_local $3)
-            (i32.const 12)
-           )
-           (i32.or
-            (i32.shr_u
-             (get_local $6)
              (i32.const 16)
             )
             (i32.shl
@@ -828,14 +752,86 @@
            )
           )
           (set_local $4
+           (i32.load
+            (i32.add
+             (get_local $1)
+             (i32.const 6)
+            )
+           )
+          )
+          (i32.store
            (i32.add
-            (get_local $4)
+            (get_local $0)
+            (i32.const 4)
+           )
+           (i32.or
+            (i32.shr_u
+             (get_local $5)
+             (i32.const 16)
+            )
+            (i32.shl
+             (get_local $4)
+             (i32.const 16)
+            )
+           )
+          )
+          (set_local $5
+           (i32.load
+            (i32.add
+             (get_local $1)
+             (i32.const 10)
+            )
+           )
+          )
+          (i32.store
+           (i32.add
+            (get_local $0)
+            (i32.const 8)
+           )
+           (i32.or
+            (i32.shr_u
+             (get_local $4)
+             (i32.const 16)
+            )
+            (i32.shl
+             (get_local $5)
+             (i32.const 16)
+            )
+           )
+          )
+          (set_local $4
+           (i32.load
+            (i32.add
+             (get_local $1)
+             (i32.const 14)
+            )
+           )
+          )
+          (i32.store
+           (i32.add
+            (get_local $0)
+            (i32.const 12)
+           )
+           (i32.or
+            (i32.shr_u
+             (get_local $5)
+             (i32.const 16)
+            )
+            (i32.shl
+             (get_local $4)
+             (i32.const 16)
+            )
+           )
+          )
+          (set_local $1
+           (i32.add
+            (get_local $1)
             (i32.const 16)
            )
           )
-          (set_local $3
+          (set_local $0
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 16)
            )
           )
@@ -853,36 +849,36 @@
      )
      (br $break|2)
     )
-    (set_local $5
+    (set_local $4
      (i32.load
-      (get_local $4)
+      (get_local $1)
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -901,91 +897,19 @@
        )
        (block
         (block
-         (set_local $6
+         (set_local $5
           (i32.load
            (i32.add
-            (get_local $4)
+            (get_local $1)
             (i32.const 3)
            )
           )
          )
          (i32.store
-          (get_local $3)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $5)
-            (i32.const 8)
-           )
-           (i32.shl
-            (get_local $6)
-            (i32.const 24)
-           )
-          )
-         )
-         (set_local $5
-          (i32.load
-           (i32.add
             (get_local $4)
-            (i32.const 7)
-           )
-          )
-         )
-         (i32.store
-          (i32.add
-           (get_local $3)
-           (i32.const 4)
-          )
-          (i32.or
-           (i32.shr_u
-            (get_local $6)
-            (i32.const 8)
-           )
-           (i32.shl
-            (get_local $5)
-            (i32.const 24)
-           )
-          )
-         )
-         (set_local $6
-          (i32.load
-           (i32.add
-            (get_local $4)
-            (i32.const 11)
-           )
-          )
-         )
-         (i32.store
-          (i32.add
-           (get_local $3)
-           (i32.const 8)
-          )
-          (i32.or
-           (i32.shr_u
-            (get_local $5)
-            (i32.const 8)
-           )
-           (i32.shl
-            (get_local $6)
-            (i32.const 24)
-           )
-          )
-         )
-         (set_local $5
-          (i32.load
-           (i32.add
-            (get_local $4)
-            (i32.const 15)
-           )
-          )
-         )
-         (i32.store
-          (i32.add
-           (get_local $3)
-           (i32.const 12)
-          )
-          (i32.or
-           (i32.shr_u
-            (get_local $6)
             (i32.const 8)
            )
            (i32.shl
@@ -995,14 +919,86 @@
           )
          )
          (set_local $4
+          (i32.load
+           (i32.add
+            (get_local $1)
+            (i32.const 7)
+           )
+          )
+         )
+         (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
+           (i32.const 4)
+          )
+          (i32.or
+           (i32.shr_u
+            (get_local $5)
+            (i32.const 8)
+           )
+           (i32.shl
+            (get_local $4)
+            (i32.const 24)
+           )
+          )
+         )
+         (set_local $5
+          (i32.load
+           (i32.add
+            (get_local $1)
+            (i32.const 11)
+           )
+          )
+         )
+         (i32.store
+          (i32.add
+           (get_local $0)
+           (i32.const 8)
+          )
+          (i32.or
+           (i32.shr_u
+            (get_local $4)
+            (i32.const 8)
+           )
+           (i32.shl
+            (get_local $5)
+            (i32.const 24)
+           )
+          )
+         )
+         (set_local $4
+          (i32.load
+           (i32.add
+            (get_local $1)
+            (i32.const 15)
+           )
+          )
+         )
+         (i32.store
+          (i32.add
+           (get_local $0)
+           (i32.const 12)
+          )
+          (i32.or
+           (i32.shr_u
+            (get_local $5)
+            (i32.const 8)
+           )
+           (i32.shl
+            (get_local $4)
+            (i32.const 24)
+           )
+          )
+         )
+         (set_local $1
+          (i32.add
+           (get_local $1)
            (i32.const 16)
           )
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -1029,449 +1025,449 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -1485,225 +1481,225 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -1717,113 +1713,113 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -1837,57 +1833,57 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $7
-       (get_local $3)
+      (set_local $6
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $7
-        (get_local $4)
+       (set_local $6
+        (get_local $1)
        )
-       (set_local $4
+       (set_local $1
         (i32.add
-         (get_local $7)
+         (get_local $6)
          (i32.const 1)
         )
        )
-       (get_local $7)
+       (get_local $6)
       )
      )
     )
@@ -1900,35 +1896,35 @@
    )
    (i32.store8
     (block (result i32)
-     (set_local $7
-      (get_local $3)
+     (set_local $6
+      (get_local $0)
      )
-     (set_local $3
+     (set_local $0
       (i32.add
-       (get_local $7)
+       (get_local $6)
        (i32.const 1)
       )
      )
-     (get_local $7)
+     (get_local $6)
     )
     (i32.load8_u
      (block (result i32)
-      (set_local $7
-       (get_local $4)
+      (set_local $6
+       (get_local $1)
       )
-      (set_local $4
+      (set_local $1
        (i32.add
-        (get_local $7)
+        (get_local $6)
         (i32.const 1)
        )
       )
-      (get_local $7)
+      (get_local $6)
      )
     )
    )
   )
   (return
-   (get_local $0)
+   (get_local $3)
   )
  )
  (func $fmod/fmod (; 6 ;) (type $FFF) (param $0 f64) (param $1 f64) (result f64)
diff --git a/tests/compiler/std/array.optimized-inlined.wast b/tests/compiler/std/array.optimized-inlined.wast
index 4b0efda5..dbfbab36 100644
--- a/tests/compiler/std/array.optimized-inlined.wast
+++ b/tests/compiler/std/array.optimized-inlined.wast
@@ -1,6 +1,6 @@
 (module
  (type $ii (func (param i32) (result i32)))
- (type $iiii (func (param i32 i32 i32) (result i32)))
+ (type $iiiv (func (param i32 i32 i32)))
  (type $iv (func (param i32)))
  (type $iiv (func (param i32 i32)))
  (type $iii (func (param i32 i32) (result i32)))
@@ -12,7 +12,7 @@
  (memory $0 1)
  (export "memory" (memory $0))
  (start $start)
- (func $std:heap/Heap.allocate (; 0 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:heap/allocate_memory (; 0 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
   (if
@@ -101,20 +101,9 @@
   )
   (get_local $1)
  )
- (func $std:heap/Heap.copy (; 1 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/copy_memory (; 1 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
-  (if
-   (i32.lt_u
-    (get_local $0)
-    (get_global $HEAP_BASE)
-   )
-   (unreachable)
-  )
-  (set_local $4
-   (get_local $0)
-  )
   (loop $continue|0
    (if
     (if (result i32)
@@ -126,10 +115,10 @@
      (get_local $2)
     )
     (block
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -163,7 +152,7 @@
   (if
    (i32.eqz
     (i32.rem_u
-     (get_local $4)
+     (get_local $0)
      (i32.const 4)
     )
    )
@@ -176,14 +165,14 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.load
          (get_local $1)
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.load
@@ -195,7 +184,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.load
@@ -207,7 +196,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.load
@@ -223,9 +212,9 @@
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -246,14 +235,14 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
@@ -263,9 +252,9 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 8)
        )
       )
@@ -284,14 +273,14 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
       )
@@ -310,14 +299,14 @@
      )
      (block
       (i32.store16
-       (get_local $4)
+       (get_local $0)
        (i32.load16_u
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 2)
        )
       )
@@ -336,7 +325,7 @@
      )
      (block
       (set_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.store8
        (get_local $3)
@@ -351,9 +340,7 @@
       )
      )
     )
-    (return
-     (get_local $0)
-    )
+    (return)
    )
   )
   (if
@@ -369,7 +356,7 @@
         (br_table $case0|2 $case1|2 $case2|2 $tablify|0
          (i32.sub
           (i32.rem_u
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.const 1)
@@ -378,15 +365,15 @@
        )
        (br $break|2)
       )
-      (set_local $5
+      (set_local $4
        (i32.load
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -407,10 +394,10 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -431,10 +418,10 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -469,10 +456,10 @@
         )
         (block
          (i32.store
-          (get_local $4)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
@@ -490,7 +477,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.or
@@ -499,7 +486,7 @@
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
                (get_local $1)
@@ -513,12 +500,12 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
@@ -536,7 +523,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.or
@@ -545,7 +532,7 @@
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
                (get_local $1)
@@ -563,9 +550,9 @@
            (i32.const 16)
           )
          )
-         (set_local $4
+         (set_local $0
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -581,15 +568,15 @@
       )
       (br $break|2)
      )
-     (set_local $5
+     (set_local $4
       (i32.load
        (get_local $1)
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -610,10 +597,10 @@
        )
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -648,10 +635,10 @@
        )
        (block
         (i32.store
-         (get_local $4)
+         (get_local $0)
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
@@ -669,7 +656,7 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 4)
          )
          (i32.or
@@ -678,7 +665,7 @@
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -692,12 +679,12 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 8)
          )
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
@@ -715,7 +702,7 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 12)
          )
          (i32.or
@@ -724,7 +711,7 @@
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -742,9 +729,9 @@
           (i32.const 16)
          )
         )
-        (set_local $4
+        (set_local $0
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 16)
          )
         )
@@ -760,15 +747,15 @@
      )
      (br $break|2)
     )
-    (set_local $5
+    (set_local $4
      (i32.load
       (get_local $1)
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -803,10 +790,10 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
@@ -824,7 +811,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.or
@@ -833,7 +820,7 @@
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -847,12 +834,12 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
@@ -870,7 +857,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.or
@@ -879,7 +866,7 @@
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -897,9 +884,9 @@
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -921,10 +908,10 @@
     (i32.const 16)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -945,10 +932,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -969,10 +956,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -993,10 +980,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1017,10 +1004,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1041,10 +1028,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1065,10 +1052,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1089,10 +1076,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1113,10 +1100,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1137,10 +1124,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1161,10 +1148,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1185,10 +1172,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1209,10 +1196,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1233,10 +1220,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1257,10 +1244,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1281,10 +1268,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1313,10 +1300,10 @@
     (i32.const 8)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1337,10 +1324,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1361,10 +1348,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1385,10 +1372,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1409,10 +1396,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1433,10 +1420,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1457,10 +1444,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1481,10 +1468,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1513,10 +1500,10 @@
     (i32.const 4)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1537,10 +1524,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1561,10 +1548,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1585,10 +1572,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1617,10 +1604,10 @@
     (i32.const 2)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1641,10 +1628,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1674,7 +1661,7 @@
    )
    (block
     (set_local $3
-     (get_local $4)
+     (get_local $0)
     )
     (i32.store8
      (get_local $3)
@@ -1689,9 +1676,298 @@
     )
    )
   )
-  (get_local $0)
  )
- (func $std:array/Array#__grow (; 2 ;) (type $iiv) (param $0 i32) (param $1 i32)
+ (func $std:heap/move_memory (; 2 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (local $3 i32)
+  (if
+   (i32.eq
+    (get_local $0)
+    (get_local $1)
+   )
+   (return)
+  )
+  (if
+   (i32.and
+    (if (result i32)
+     (tee_local $3
+      (i32.le_u
+       (i32.add
+        (get_local $1)
+        (get_local $2)
+       )
+       (get_local $0)
+      )
+     )
+     (get_local $3)
+     (i32.le_u
+      (i32.add
+       (get_local $0)
+       (get_local $2)
+      )
+      (get_local $1)
+     )
+    )
+    (i32.const 1)
+   )
+   (block
+    (call $std:heap/copy_memory
+     (get_local $0)
+     (get_local $1)
+     (get_local $2)
+    )
+    (return)
+   )
+  )
+  (if
+   (i32.lt_u
+    (get_local $0)
+    (get_local $1)
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (loop $continue|0
+       (if
+        (i32.rem_u
+         (get_local $0)
+         (i32.const 8)
+        )
+        (block
+         (if
+          (i32.eqz
+           (get_local $2)
+          )
+          (return)
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+         (set_local $0
+          (i32.add
+           (tee_local $3
+            (get_local $0)
+           )
+           (i32.const 1)
+          )
+         )
+         (i32.store8
+          (get_local $3)
+          (block (result i32)
+           (set_local $1
+            (i32.add
+             (tee_local $3
+              (get_local $1)
+             )
+             (i32.const 1)
+            )
+           )
+           (i32.load8_u
+            (get_local $3)
+           )
+          )
+         )
+         (br $continue|0)
+        )
+       )
+      )
+      (loop $continue|1
+       (if
+        (i32.ge_u
+         (get_local $2)
+         (i32.const 8)
+        )
+        (block
+         (i64.store
+          (get_local $0)
+          (i64.load
+           (get_local $1)
+          )
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 8)
+          )
+         )
+         (set_local $0
+          (i32.add
+           (get_local $0)
+           (i32.const 8)
+          )
+         )
+         (set_local $1
+          (i32.add
+           (get_local $1)
+           (i32.const 8)
+          )
+         )
+         (br $continue|1)
+        )
+       )
+      )
+     )
+    )
+    (loop $continue|2
+     (if
+      (get_local $2)
+      (block
+       (set_local $0
+        (i32.add
+         (tee_local $3
+          (get_local $0)
+         )
+         (i32.const 1)
+        )
+       )
+       (i32.store8
+        (get_local $3)
+        (block (result i32)
+         (set_local $1
+          (i32.add
+           (tee_local $3
+            (get_local $1)
+           )
+           (i32.const 1)
+          )
+         )
+         (i32.load8_u
+          (get_local $3)
+         )
+        )
+       )
+       (set_local $2
+        (i32.sub
+         (get_local $2)
+         (i32.const 1)
+        )
+       )
+       (br $continue|2)
+      )
+     )
+    )
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (loop $continue|3
+       (if
+        (i32.rem_u
+         (i32.add
+          (get_local $0)
+          (get_local $2)
+         )
+         (i32.const 8)
+        )
+        (block
+         (if
+          (i32.eqz
+           (get_local $2)
+          )
+          (return)
+         )
+         (i32.store8
+          (i32.add
+           (get_local $0)
+           (tee_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 1)
+            )
+           )
+          )
+          (i32.load8_u
+           (i32.add
+            (get_local $1)
+            (get_local $2)
+           )
+          )
+         )
+         (br $continue|3)
+        )
+       )
+      )
+      (loop $continue|4
+       (if
+        (i32.ge_u
+         (get_local $2)
+         (i32.const 8)
+        )
+        (block
+         (i64.store
+          (i32.add
+           (get_local $0)
+           (tee_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+          )
+          (i64.load
+           (i32.add
+            (get_local $1)
+            (get_local $2)
+           )
+          )
+         )
+         (br $continue|4)
+        )
+       )
+      )
+     )
+    )
+    (loop $continue|5
+     (if
+      (get_local $2)
+      (block
+       (i32.store8
+        (i32.add
+         (get_local $0)
+         (tee_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+        )
+        (i32.load8_u
+         (i32.add
+          (get_local $1)
+          (get_local $2)
+         )
+        )
+       )
+       (br $continue|5)
+      )
+     )
+    )
+   )
+  )
+ )
+ (func $std:array/Array#__grow (; 3 ;) (type $iiv) (param $0 i32) (param $1 i32)
   (local $2 i32)
   (local $3 i32)
   (if
@@ -1704,7 +1980,7 @@
    (unreachable)
   )
   (set_local $2
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.mul
      (get_local $1)
      (i32.const 4)
@@ -1715,8 +1991,8 @@
    (i32.load
     (get_local $0)
    )
-   (drop
-    (call $std:heap/Heap.copy
+   (block
+    (call $std:heap/move_memory
      (get_local $2)
      (i32.load
       (get_local $0)
@@ -1728,16 +2004,16 @@
       (i32.const 4)
      )
     )
-   )
-  )
-  (block
-   (block $__inlined_func$std:heap/Heap.dispose
-    (set_local $3
-     (i32.load
-      (get_local $0)
+    (block
+     (block $__inlined_func$std:heap/free_memory
+      (set_local $3
+       (i32.load
+        (get_local $0)
+       )
+      )
+      (nop)
      )
     )
-    (nop)
    )
   )
   (i32.store
@@ -1749,7 +2025,7 @@
    (get_local $1)
   )
  )
- (func $std:array/Array#push (; 3 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#push (; 4 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
   (if
@@ -1773,11 +2049,11 @@
       )
      )
      (tee_local $3
-      (i32.mul
+      (i32.shl
        (i32.load offset=4
         (get_local $0)
        )
-       (i32.const 2)
+       (i32.const 1)
       )
      )
      (i32.gt_s
@@ -1814,7 +2090,7 @@
   )
   (get_local $2)
  )
- (func $std:array/Array#__get (; 4 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#__get (; 5 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (if
    (i32.ge_u
     (get_local $1)
@@ -1836,7 +2112,7 @@
    )
   )
  )
- (func $std:array/Array#pop (; 5 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:array/Array#pop (; 6 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (if
    (i32.and
@@ -1886,7 +2162,7 @@
    )
   )
  )
- (func $std:array/Array#unshift (; 6 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#unshift (; 7 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
   (local $4 i32)
@@ -1989,35 +2265,23 @@
   )
   (get_local $3)
  )
- (func $std:heap/Heap.fill (; 7 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
-  (local $3 i32)
-  (local $4 i64)
-  (local $5 i32)
-  (if
-   (i32.lt_u
-    (get_local $0)
-    (get_global $HEAP_BASE)
-   )
-   (unreachable)
-  )
+ (func $std:heap/set_memory (; 8 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (local $3 i64)
+  (local $4 i32)
   (if
    (i32.eqz
     (get_local $2)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
-   (tee_local $3
-    (get_local $0)
-   )
+   (get_local $0)
    (get_local $1)
   )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 1)
@@ -2029,38 +2293,36 @@
     (get_local $2)
     (i32.const 2)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 1)
    )
    (get_local $1)
   )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 2)
    )
    (get_local $1)
   )
-  (i32.store8
-   (i32.add
-    (get_local $3)
-    (i32.const 2)
-   )
-   (get_local $1)
-  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 3)
@@ -2072,13 +2334,11 @@
     (get_local $2)
     (i32.const 6)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 3)
    )
    (get_local $1)
@@ -2086,7 +2346,7 @@
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 4)
@@ -2098,19 +2358,17 @@
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
-   (tee_local $3
+   (tee_local $0
     (i32.add
-     (get_local $3)
-     (tee_local $5
+     (get_local $0)
+     (tee_local $4
       (i32.and
        (i32.sub
         (i32.const 0)
-        (get_local $3)
+        (get_local $0)
        )
        (i32.const 3)
       )
@@ -2127,12 +2385,12 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (tee_local $2
       (i32.and
        (i32.sub
         (get_local $2)
-        (get_local $5)
+        (get_local $4)
        )
        (i32.const -4)
       )
@@ -2147,20 +2405,18 @@
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 4)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 8)
    )
    (get_local $1)
@@ -2168,7 +2424,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 12)
@@ -2178,7 +2434,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 8)
@@ -2190,34 +2446,32 @@
     (get_local $2)
     (i32.const 24)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 12)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 16)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 20)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 24)
    )
    (get_local $1)
@@ -2225,7 +2479,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 28)
@@ -2235,7 +2489,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 24)
@@ -2245,7 +2499,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 20)
@@ -2255,20 +2509,20 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 16)
    )
    (get_local $1)
   )
-  (set_local $3
+  (set_local $0
    (i32.add
-    (get_local $3)
-    (tee_local $5
+    (get_local $0)
+    (tee_local $4
      (i32.add
       (i32.and
-       (get_local $3)
+       (get_local $0)
        (i32.const 4)
       )
       (i32.const 24)
@@ -2279,10 +2533,10 @@
   (set_local $2
    (i32.sub
     (get_local $2)
-    (get_local $5)
+    (get_local $4)
    )
   )
-  (set_local $4
+  (set_local $3
    (i64.or
     (i64.extend_u/i32
      (get_local $1)
@@ -2303,29 +2557,29 @@
     )
     (block
      (i64.store
+      (get_local $0)
       (get_local $3)
-      (get_local $4)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 8)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 16)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 24)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (set_local $2
       (i32.sub
@@ -2333,9 +2587,9 @@
        (i32.const 32)
       )
      )
-     (set_local $3
+     (set_local $0
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 32)
       )
      )
@@ -2343,9 +2597,8 @@
     )
    )
   )
-  (get_local $0)
  )
- (func $std:array/Array#shift (; 8 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:array/Array#shift (; 9 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (if
    (i32.and
@@ -2379,16 +2632,30 @@
     )
    )
   )
-  (drop
-   (call $std:heap/Heap.copy
+  (call $std:heap/move_memory
+   (i32.load
+    (get_local $0)
+   )
+   (i32.add
     (i32.load
      (get_local $0)
     )
-    (i32.add
-     (i32.load
+    (i32.const 4)
+   )
+   (i32.mul
+    (i32.sub
+     (i32.load offset=4
       (get_local $0)
      )
-     (i32.const 4)
+     (i32.const 1)
+    )
+    (i32.const 4)
+   )
+  )
+  (call $std:heap/set_memory
+   (i32.add
+    (i32.load
+     (get_local $0)
     )
     (i32.mul
      (i32.sub
@@ -2400,26 +2667,8 @@
      (i32.const 4)
     )
    )
-  )
-  (drop
-   (call $std:heap/Heap.fill
-    (i32.add
-     (i32.load
-      (get_local $0)
-     )
-     (i32.mul
-      (i32.sub
-       (i32.load offset=4
-        (get_local $0)
-       )
-       (i32.const 1)
-      )
-      (i32.const 4)
-     )
-    )
-    (i32.const 0)
-    (i32.const 4)
-   )
+   (i32.const 0)
+   (i32.const 4)
   )
   (i32.store offset=8
    (get_local $0)
@@ -2432,12 +2681,12 @@
   )
   (get_local $1)
  )
- (func $start (; 9 ;) (type $v)
+ (func $start (; 10 ;) (type $v)
   (set_global $std:heap/HEAP_OFFSET
    (get_global $HEAP_BASE)
   )
   (set_global $std/array/arr
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.const 12)
    )
   )
diff --git a/tests/compiler/std/array.optimized.wast b/tests/compiler/std/array.optimized.wast
index 0c9e3bc4..1a005641 100644
--- a/tests/compiler/std/array.optimized.wast
+++ b/tests/compiler/std/array.optimized.wast
@@ -1,6 +1,6 @@
 (module
  (type $ii (func (param i32) (result i32)))
- (type $iiii (func (param i32 i32 i32) (result i32)))
+ (type $iiiv (func (param i32 i32 i32)))
  (type $iv (func (param i32)))
  (type $iiv (func (param i32 i32)))
  (type $iii (func (param i32 i32) (result i32)))
@@ -12,7 +12,7 @@
  (memory $0 1)
  (export "memory" (memory $0))
  (start $start)
- (func $std:heap/Heap.allocate (; 0 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:heap/allocate_memory (; 0 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
   (if
@@ -101,20 +101,9 @@
   )
   (get_local $1)
  )
- (func $std:heap/Heap.copy (; 1 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/copy_memory (; 1 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
-  (if
-   (i32.lt_u
-    (get_local $0)
-    (get_global $HEAP_BASE)
-   )
-   (unreachable)
-  )
-  (set_local $4
-   (get_local $0)
-  )
   (loop $continue|0
    (if
     (if (result i32)
@@ -126,10 +115,10 @@
      (get_local $2)
     )
     (block
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -163,7 +152,7 @@
   (if
    (i32.eqz
     (i32.rem_u
-     (get_local $4)
+     (get_local $0)
      (i32.const 4)
     )
    )
@@ -176,14 +165,14 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.load
          (get_local $1)
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.load
@@ -195,7 +184,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.load
@@ -207,7 +196,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.load
@@ -223,9 +212,9 @@
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -246,14 +235,14 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
@@ -263,9 +252,9 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 8)
        )
       )
@@ -284,14 +273,14 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
       )
@@ -310,14 +299,14 @@
      )
      (block
       (i32.store16
-       (get_local $4)
+       (get_local $0)
        (i32.load16_u
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 2)
        )
       )
@@ -336,7 +325,7 @@
      )
      (block
       (set_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.store8
        (get_local $3)
@@ -351,9 +340,7 @@
       )
      )
     )
-    (return
-     (get_local $0)
-    )
+    (return)
    )
   )
   (if
@@ -369,7 +356,7 @@
         (br_table $case0|2 $case1|2 $case2|2 $tablify|0
          (i32.sub
           (i32.rem_u
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.const 1)
@@ -378,15 +365,15 @@
        )
        (br $break|2)
       )
-      (set_local $5
+      (set_local $4
        (i32.load
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -407,10 +394,10 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -431,10 +418,10 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -469,10 +456,10 @@
         )
         (block
          (i32.store
-          (get_local $4)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
@@ -490,7 +477,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.or
@@ -499,7 +486,7 @@
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
                (get_local $1)
@@ -513,12 +500,12 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
@@ -536,7 +523,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.or
@@ -545,7 +532,7 @@
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
                (get_local $1)
@@ -563,9 +550,9 @@
            (i32.const 16)
           )
          )
-         (set_local $4
+         (set_local $0
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -581,15 +568,15 @@
       )
       (br $break|2)
      )
-     (set_local $5
+     (set_local $4
       (i32.load
        (get_local $1)
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -610,10 +597,10 @@
        )
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -648,10 +635,10 @@
        )
        (block
         (i32.store
-         (get_local $4)
+         (get_local $0)
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
@@ -669,7 +656,7 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 4)
          )
          (i32.or
@@ -678,7 +665,7 @@
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -692,12 +679,12 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 8)
          )
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
@@ -715,7 +702,7 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 12)
          )
          (i32.or
@@ -724,7 +711,7 @@
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -742,9 +729,9 @@
           (i32.const 16)
          )
         )
-        (set_local $4
+        (set_local $0
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 16)
          )
         )
@@ -760,15 +747,15 @@
      )
      (br $break|2)
     )
-    (set_local $5
+    (set_local $4
      (i32.load
       (get_local $1)
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -803,10 +790,10 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
@@ -824,7 +811,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.or
@@ -833,7 +820,7 @@
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -847,12 +834,12 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
@@ -870,7 +857,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.or
@@ -879,7 +866,7 @@
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -897,9 +884,9 @@
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -921,10 +908,10 @@
     (i32.const 16)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -945,10 +932,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -969,10 +956,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -993,10 +980,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1017,10 +1004,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1041,10 +1028,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1065,10 +1052,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1089,10 +1076,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1113,10 +1100,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1137,10 +1124,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1161,10 +1148,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1185,10 +1172,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1209,10 +1196,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1233,10 +1220,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1257,10 +1244,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1281,10 +1268,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1313,10 +1300,10 @@
     (i32.const 8)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1337,10 +1324,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1361,10 +1348,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1385,10 +1372,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1409,10 +1396,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1433,10 +1420,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1457,10 +1444,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1481,10 +1468,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1513,10 +1500,10 @@
     (i32.const 4)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1537,10 +1524,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1561,10 +1548,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1585,10 +1572,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1617,10 +1604,10 @@
     (i32.const 2)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1641,10 +1628,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1674,7 +1661,7 @@
    )
    (block
     (set_local $3
-     (get_local $4)
+     (get_local $0)
     )
     (i32.store8
      (get_local $3)
@@ -1689,12 +1676,301 @@
     )
    )
   )
-  (get_local $0)
  )
- (func $std:heap/Heap.dispose (; 2 ;) (type $iv) (param $0 i32)
+ (func $std:heap/move_memory (; 2 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (local $3 i32)
+  (if
+   (i32.eq
+    (get_local $0)
+    (get_local $1)
+   )
+   (return)
+  )
+  (if
+   (i32.and
+    (if (result i32)
+     (tee_local $3
+      (i32.le_u
+       (i32.add
+        (get_local $1)
+        (get_local $2)
+       )
+       (get_local $0)
+      )
+     )
+     (get_local $3)
+     (i32.le_u
+      (i32.add
+       (get_local $0)
+       (get_local $2)
+      )
+      (get_local $1)
+     )
+    )
+    (i32.const 1)
+   )
+   (block
+    (call $std:heap/copy_memory
+     (get_local $0)
+     (get_local $1)
+     (get_local $2)
+    )
+    (return)
+   )
+  )
+  (if
+   (i32.lt_u
+    (get_local $0)
+    (get_local $1)
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (loop $continue|0
+       (if
+        (i32.rem_u
+         (get_local $0)
+         (i32.const 8)
+        )
+        (block
+         (if
+          (i32.eqz
+           (get_local $2)
+          )
+          (return)
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+         (set_local $0
+          (i32.add
+           (tee_local $3
+            (get_local $0)
+           )
+           (i32.const 1)
+          )
+         )
+         (i32.store8
+          (get_local $3)
+          (block (result i32)
+           (set_local $1
+            (i32.add
+             (tee_local $3
+              (get_local $1)
+             )
+             (i32.const 1)
+            )
+           )
+           (i32.load8_u
+            (get_local $3)
+           )
+          )
+         )
+         (br $continue|0)
+        )
+       )
+      )
+      (loop $continue|1
+       (if
+        (i32.ge_u
+         (get_local $2)
+         (i32.const 8)
+        )
+        (block
+         (i64.store
+          (get_local $0)
+          (i64.load
+           (get_local $1)
+          )
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 8)
+          )
+         )
+         (set_local $0
+          (i32.add
+           (get_local $0)
+           (i32.const 8)
+          )
+         )
+         (set_local $1
+          (i32.add
+           (get_local $1)
+           (i32.const 8)
+          )
+         )
+         (br $continue|1)
+        )
+       )
+      )
+     )
+    )
+    (loop $continue|2
+     (if
+      (get_local $2)
+      (block
+       (set_local $0
+        (i32.add
+         (tee_local $3
+          (get_local $0)
+         )
+         (i32.const 1)
+        )
+       )
+       (i32.store8
+        (get_local $3)
+        (block (result i32)
+         (set_local $1
+          (i32.add
+           (tee_local $3
+            (get_local $1)
+           )
+           (i32.const 1)
+          )
+         )
+         (i32.load8_u
+          (get_local $3)
+         )
+        )
+       )
+       (set_local $2
+        (i32.sub
+         (get_local $2)
+         (i32.const 1)
+        )
+       )
+       (br $continue|2)
+      )
+     )
+    )
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (loop $continue|3
+       (if
+        (i32.rem_u
+         (i32.add
+          (get_local $0)
+          (get_local $2)
+         )
+         (i32.const 8)
+        )
+        (block
+         (if
+          (i32.eqz
+           (get_local $2)
+          )
+          (return)
+         )
+         (i32.store8
+          (i32.add
+           (get_local $0)
+           (tee_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 1)
+            )
+           )
+          )
+          (i32.load8_u
+           (i32.add
+            (get_local $1)
+            (get_local $2)
+           )
+          )
+         )
+         (br $continue|3)
+        )
+       )
+      )
+      (loop $continue|4
+       (if
+        (i32.ge_u
+         (get_local $2)
+         (i32.const 8)
+        )
+        (block
+         (i64.store
+          (i32.add
+           (get_local $0)
+           (tee_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+          )
+          (i64.load
+           (i32.add
+            (get_local $1)
+            (get_local $2)
+           )
+          )
+         )
+         (br $continue|4)
+        )
+       )
+      )
+     )
+    )
+    (loop $continue|5
+     (if
+      (get_local $2)
+      (block
+       (i32.store8
+        (i32.add
+         (get_local $0)
+         (tee_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+        )
+        (i32.load8_u
+         (i32.add
+          (get_local $1)
+          (get_local $2)
+         )
+        )
+       )
+       (br $continue|5)
+      )
+     )
+    )
+   )
+  )
+ )
+ (func $std:heap/free_memory (; 3 ;) (type $iv) (param $0 i32)
   (nop)
  )
- (func $std:array/Array#__grow (; 3 ;) (type $iiv) (param $0 i32) (param $1 i32)
+ (func $std:array/Array#__grow (; 4 ;) (type $iiv) (param $0 i32) (param $1 i32)
   (local $2 i32)
   (if
    (i32.le_s
@@ -1706,7 +1982,7 @@
    (unreachable)
   )
   (set_local $2
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.mul
      (get_local $1)
      (i32.const 4)
@@ -1717,8 +1993,8 @@
    (i32.load
     (get_local $0)
    )
-   (drop
-    (call $std:heap/Heap.copy
+   (block
+    (call $std:heap/move_memory
      (get_local $2)
      (i32.load
       (get_local $0)
@@ -1730,11 +2006,11 @@
       (i32.const 4)
      )
     )
-   )
-  )
-  (call $std:heap/Heap.dispose
-   (i32.load
-    (get_local $0)
+    (call $std:heap/free_memory
+     (i32.load
+      (get_local $0)
+     )
+    )
    )
   )
   (i32.store
@@ -1746,7 +2022,7 @@
    (get_local $1)
   )
  )
- (func $std:array/Array#push (; 4 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#push (; 5 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
   (if
@@ -1770,11 +2046,11 @@
       )
      )
      (tee_local $3
-      (i32.mul
+      (i32.shl
        (i32.load offset=4
         (get_local $0)
        )
-       (i32.const 2)
+       (i32.const 1)
       )
      )
      (i32.gt_s
@@ -1811,7 +2087,7 @@
   )
   (get_local $2)
  )
- (func $std:array/Array#__get (; 5 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#__get (; 6 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (if
    (i32.ge_u
     (get_local $1)
@@ -1833,7 +2109,7 @@
    )
   )
  )
- (func $std:array/Array#pop (; 6 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:array/Array#pop (; 7 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (if
    (i32.and
@@ -1883,7 +2159,7 @@
    )
   )
  )
- (func $std:array/Array#unshift (; 7 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#unshift (; 8 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
   (local $4 i32)
@@ -1986,35 +2262,23 @@
   )
   (get_local $3)
  )
- (func $std:heap/Heap.fill (; 8 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
-  (local $3 i32)
-  (local $4 i64)
-  (local $5 i32)
-  (if
-   (i32.lt_u
-    (get_local $0)
-    (get_global $HEAP_BASE)
-   )
-   (unreachable)
-  )
+ (func $std:heap/set_memory (; 9 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (local $3 i64)
+  (local $4 i32)
   (if
    (i32.eqz
     (get_local $2)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
-   (tee_local $3
-    (get_local $0)
-   )
+   (get_local $0)
    (get_local $1)
   )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 1)
@@ -2026,38 +2290,36 @@
     (get_local $2)
     (i32.const 2)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 1)
    )
    (get_local $1)
   )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 2)
    )
    (get_local $1)
   )
-  (i32.store8
-   (i32.add
-    (get_local $3)
-    (i32.const 2)
-   )
-   (get_local $1)
-  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 3)
@@ -2069,13 +2331,11 @@
     (get_local $2)
     (i32.const 6)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 3)
    )
    (get_local $1)
@@ -2083,7 +2343,7 @@
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 4)
@@ -2095,19 +2355,17 @@
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
-   (tee_local $3
+   (tee_local $0
     (i32.add
-     (get_local $3)
-     (tee_local $5
+     (get_local $0)
+     (tee_local $4
       (i32.and
        (i32.sub
         (i32.const 0)
-        (get_local $3)
+        (get_local $0)
        )
        (i32.const 3)
       )
@@ -2124,12 +2382,12 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (tee_local $2
       (i32.and
        (i32.sub
         (get_local $2)
-        (get_local $5)
+        (get_local $4)
        )
        (i32.const -4)
       )
@@ -2144,20 +2402,18 @@
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 4)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 8)
    )
    (get_local $1)
@@ -2165,7 +2421,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 12)
@@ -2175,7 +2431,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 8)
@@ -2187,34 +2443,32 @@
     (get_local $2)
     (i32.const 24)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 12)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 16)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 20)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 24)
    )
    (get_local $1)
@@ -2222,7 +2476,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 28)
@@ -2232,7 +2486,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 24)
@@ -2242,7 +2496,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 20)
@@ -2252,20 +2506,20 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 16)
    )
    (get_local $1)
   )
-  (set_local $3
+  (set_local $0
    (i32.add
-    (get_local $3)
-    (tee_local $5
+    (get_local $0)
+    (tee_local $4
      (i32.add
       (i32.and
-       (get_local $3)
+       (get_local $0)
        (i32.const 4)
       )
       (i32.const 24)
@@ -2276,10 +2530,10 @@
   (set_local $2
    (i32.sub
     (get_local $2)
-    (get_local $5)
+    (get_local $4)
    )
   )
-  (set_local $4
+  (set_local $3
    (i64.or
     (i64.extend_u/i32
      (get_local $1)
@@ -2300,29 +2554,29 @@
     )
     (block
      (i64.store
+      (get_local $0)
       (get_local $3)
-      (get_local $4)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 8)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 16)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 24)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (set_local $2
       (i32.sub
@@ -2330,9 +2584,9 @@
        (i32.const 32)
       )
      )
-     (set_local $3
+     (set_local $0
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 32)
       )
      )
@@ -2340,9 +2594,8 @@
     )
    )
   )
-  (get_local $0)
  )
- (func $std:array/Array#shift (; 9 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:array/Array#shift (; 10 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (if
    (i32.and
@@ -2376,16 +2629,30 @@
     )
    )
   )
-  (drop
-   (call $std:heap/Heap.copy
+  (call $std:heap/move_memory
+   (i32.load
+    (get_local $0)
+   )
+   (i32.add
     (i32.load
      (get_local $0)
     )
-    (i32.add
-     (i32.load
+    (i32.const 4)
+   )
+   (i32.mul
+    (i32.sub
+     (i32.load offset=4
       (get_local $0)
      )
-     (i32.const 4)
+     (i32.const 1)
+    )
+    (i32.const 4)
+   )
+  )
+  (call $std:heap/set_memory
+   (i32.add
+    (i32.load
+     (get_local $0)
     )
     (i32.mul
      (i32.sub
@@ -2397,26 +2664,8 @@
      (i32.const 4)
     )
    )
-  )
-  (drop
-   (call $std:heap/Heap.fill
-    (i32.add
-     (i32.load
-      (get_local $0)
-     )
-     (i32.mul
-      (i32.sub
-       (i32.load offset=4
-        (get_local $0)
-       )
-       (i32.const 1)
-      )
-      (i32.const 4)
-     )
-    )
-    (i32.const 0)
-    (i32.const 4)
-   )
+   (i32.const 0)
+   (i32.const 4)
   )
   (i32.store offset=8
    (get_local $0)
@@ -2429,12 +2678,12 @@
   )
   (get_local $1)
  )
- (func $start (; 10 ;) (type $v)
+ (func $start (; 11 ;) (type $v)
   (set_global $std:heap/HEAP_OFFSET
    (get_global $HEAP_BASE)
   )
   (set_global $std/array/arr
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.const 12)
    )
   )
diff --git a/tests/compiler/std/array.ts b/tests/compiler/std/array.ts
index 31b5c0e4..3e27683b 100644
--- a/tests/compiler/std/array.ts
+++ b/tests/compiler/std/array.ts
@@ -1,4 +1,4 @@
-var arr = changetype<i32[]>(Heap.allocate(sizeof<usize>() + 2 * sizeof<i32>()));
+var arr = changetype<i32[]>(allocate_memory(sizeof<usize>() + 2 * sizeof<i32>()));
 
 assert(arr.length == 0);
 assert(arr.__capacity == 0);
@@ -36,7 +36,7 @@ assert(arr[0] == 43);
 assert(arr[1] == 44);
 assert(arr[2] == 45);
 
-arr.unshift(42); // see FIXME in std:array
+arr.unshift(42);
 
 assert(arr.length == 4);
 assert(arr.__capacity == 4);
diff --git a/tests/compiler/std/array.wast b/tests/compiler/std/array.wast
index 265e87d3..7d3bd821 100644
--- a/tests/compiler/std/array.wast
+++ b/tests/compiler/std/array.wast
@@ -1,7 +1,7 @@
 (module
  (type $i (func (result i32)))
  (type $ii (func (param i32) (result i32)))
- (type $iiii (func (param i32 i32 i32) (result i32)))
+ (type $iiiv (func (param i32 i32 i32)))
  (type $iv (func (param i32)))
  (type $iiv (func (param i32 i32)))
  (type $iii (func (param i32 i32) (result i32)))
@@ -16,7 +16,7 @@
  (memory $0 1)
  (export "memory" (memory $0))
  (start $start)
- (func $std:heap/Heap.allocate (; 0 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:heap/allocate_memory (; 0 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
   (local $3 i32)
@@ -113,25 +113,10 @@
    (get_local $4)
   )
  )
- (func $std:heap/Heap.copy (; 1 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/copy_memory (; 1 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  (local $6 i32)
-  (if
-   (i32.eqz
-    (i32.ge_u
-     (get_local $0)
-     (get_global $HEAP_BASE)
-    )
-   )
-   (unreachable)
-  )
-  (block
-   (set_local $3
-    (get_local $0)
-   )
-  )
   (nop)
   (block $break|0
    (loop $continue|0
@@ -151,29 +136,29 @@
       (block
        (i32.store8
         (block (result i32)
-         (set_local $6
-          (get_local $3)
+         (set_local $5
+          (get_local $0)
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $6)
+           (get_local $5)
            (i32.const 1)
           )
          )
-         (get_local $6)
+         (get_local $5)
         )
         (i32.load8_u
          (block (result i32)
-          (set_local $6
+          (set_local $5
            (get_local $1)
           )
           (set_local $1
            (i32.add
-            (get_local $6)
+            (get_local $5)
             (i32.const 1)
            )
           )
-          (get_local $6)
+          (get_local $5)
          )
         )
        )
@@ -192,7 +177,7 @@
   (if
    (i32.eq
     (i32.rem_u
-     (get_local $3)
+     (get_local $0)
      (i32.const 4)
     )
     (i32.const 0)
@@ -208,14 +193,14 @@
        (block
         (block
          (i32.store
-          (get_local $3)
+          (get_local $0)
           (i32.load
            (get_local $1)
           )
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.load
@@ -227,7 +212,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.load
@@ -239,7 +224,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.load
@@ -255,9 +240,9 @@
            (i32.const 16)
           )
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -280,14 +265,14 @@
      )
      (block
       (i32.store
-       (get_local $3)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
@@ -297,9 +282,9 @@
         )
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 8)
        )
       )
@@ -318,14 +303,14 @@
      )
      (block
       (i32.store
-       (get_local $3)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 4)
        )
       )
@@ -344,14 +329,14 @@
      )
      (block
       (i32.store16
-       (get_local $3)
+       (get_local $0)
        (i32.load16_u
         (get_local $1)
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 2)
        )
       )
@@ -370,36 +355,34 @@
      )
      (i32.store8
       (block (result i32)
-       (set_local $6
-        (get_local $3)
+       (set_local $5
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $6
+        (set_local $5
          (get_local $1)
         )
         (set_local $1
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
       )
      )
     )
-    (return
-     (get_local $0)
-    )
+    (return)
    )
   )
   (if
@@ -411,118 +394,118 @@
     (block $case2|2
      (block $case1|2
       (block $case0|2
-       (set_local $6
+       (set_local $5
         (i32.rem_u
-         (get_local $3)
+         (get_local $0)
          (i32.const 4)
         )
        )
        (br_if $case0|2
         (i32.eq
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
        (br_if $case1|2
         (i32.eq
-         (get_local $6)
+         (get_local $5)
          (i32.const 2)
         )
        )
        (br_if $case2|2
         (i32.eq
-         (get_local $6)
+         (get_local $5)
          (i32.const 3)
         )
        )
        (br $break|2)
       )
-      (set_local $4
+      (set_local $3
        (i32.load
         (get_local $1)
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $6
-         (get_local $3)
+        (set_local $5
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $6
+         (set_local $5
           (get_local $1)
          )
          (set_local $1
           (i32.add
-           (get_local $6)
+           (get_local $5)
            (i32.const 1)
           )
          )
-         (get_local $6)
+         (get_local $5)
         )
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $6
-         (get_local $3)
+        (set_local $5
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $6
+         (set_local $5
           (get_local $1)
          )
          (set_local $1
           (i32.add
-           (get_local $6)
+           (get_local $5)
            (i32.const 1)
           )
          )
-         (get_local $6)
+         (get_local $5)
         )
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $6
-         (get_local $3)
+        (set_local $5
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $6
+         (set_local $5
           (get_local $1)
          )
          (set_local $1
           (i32.add
-           (get_local $6)
+           (get_local $5)
            (i32.const 1)
           )
          )
-         (get_local $6)
+         (get_local $5)
         )
        )
       )
@@ -541,7 +524,7 @@
          )
          (block
           (block
-           (set_local $5
+           (set_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -550,19 +533,19 @@
             )
            )
            (i32.store
-            (get_local $3)
+            (get_local $0)
             (i32.or
              (i32.shr_u
-              (get_local $4)
+              (get_local $3)
               (i32.const 24)
              )
              (i32.shl
-              (get_local $5)
+              (get_local $4)
               (i32.const 8)
              )
             )
            )
-           (set_local $4
+           (set_local $3
             (i32.load
              (i32.add
               (get_local $1)
@@ -572,21 +555,21 @@
            )
            (i32.store
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 4)
             )
             (i32.or
              (i32.shr_u
-              (get_local $5)
+              (get_local $4)
               (i32.const 24)
              )
              (i32.shl
-              (get_local $4)
+              (get_local $3)
               (i32.const 8)
              )
             )
            )
-           (set_local $5
+           (set_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -596,21 +579,21 @@
            )
            (i32.store
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 8)
             )
             (i32.or
              (i32.shr_u
-              (get_local $4)
+              (get_local $3)
               (i32.const 24)
              )
              (i32.shl
-              (get_local $5)
+              (get_local $4)
               (i32.const 8)
              )
             )
            )
-           (set_local $4
+           (set_local $3
             (i32.load
              (i32.add
               (get_local $1)
@@ -620,16 +603,16 @@
            )
            (i32.store
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 12)
             )
             (i32.or
              (i32.shr_u
-              (get_local $5)
+              (get_local $4)
               (i32.const 24)
              )
              (i32.shl
-              (get_local $4)
+              (get_local $3)
               (i32.const 8)
              )
             )
@@ -640,9 +623,9 @@
              (i32.const 16)
             )
            )
-           (set_local $3
+           (set_local $0
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 16)
             )
            )
@@ -660,64 +643,64 @@
       )
       (br $break|2)
      )
-     (set_local $4
+     (set_local $3
       (i32.load
        (get_local $1)
       )
      )
      (i32.store8
       (block (result i32)
-       (set_local $6
-        (get_local $3)
+       (set_local $5
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $6
+        (set_local $5
          (get_local $1)
         )
         (set_local $1
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
       )
      )
      (i32.store8
       (block (result i32)
-       (set_local $6
-        (get_local $3)
+       (set_local $5
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $6
+        (set_local $5
          (get_local $1)
         )
         (set_local $1
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
       )
      )
@@ -736,7 +719,7 @@
         )
         (block
          (block
-          (set_local $5
+          (set_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -745,19 +728,19 @@
            )
           )
           (i32.store
-           (get_local $3)
+           (get_local $0)
            (i32.or
             (i32.shr_u
-             (get_local $4)
+             (get_local $3)
              (i32.const 16)
             )
             (i32.shl
-             (get_local $5)
+             (get_local $4)
              (i32.const 16)
             )
            )
           )
-          (set_local $4
+          (set_local $3
            (i32.load
             (i32.add
              (get_local $1)
@@ -767,21 +750,21 @@
           )
           (i32.store
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 4)
            )
            (i32.or
             (i32.shr_u
-             (get_local $5)
+             (get_local $4)
              (i32.const 16)
             )
             (i32.shl
-             (get_local $4)
+             (get_local $3)
              (i32.const 16)
             )
            )
           )
-          (set_local $5
+          (set_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -791,21 +774,21 @@
           )
           (i32.store
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 8)
            )
            (i32.or
             (i32.shr_u
-             (get_local $4)
+             (get_local $3)
              (i32.const 16)
             )
             (i32.shl
-             (get_local $5)
+             (get_local $4)
              (i32.const 16)
             )
            )
           )
-          (set_local $4
+          (set_local $3
            (i32.load
             (i32.add
              (get_local $1)
@@ -815,16 +798,16 @@
           )
           (i32.store
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 12)
            )
            (i32.or
             (i32.shr_u
-             (get_local $5)
+             (get_local $4)
              (i32.const 16)
             )
             (i32.shl
-             (get_local $4)
+             (get_local $3)
              (i32.const 16)
             )
            )
@@ -835,9 +818,9 @@
             (i32.const 16)
            )
           )
-          (set_local $3
+          (set_local $0
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 16)
            )
           )
@@ -855,36 +838,36 @@
      )
      (br $break|2)
     )
-    (set_local $4
+    (set_local $3
      (i32.load
       (get_local $1)
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -903,7 +886,7 @@
        )
        (block
         (block
-         (set_local $5
+         (set_local $4
           (i32.load
            (i32.add
             (get_local $1)
@@ -912,19 +895,19 @@
           )
          )
          (i32.store
-          (get_local $3)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $4)
+            (get_local $3)
             (i32.const 8)
            )
            (i32.shl
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
           )
          )
-         (set_local $4
+         (set_local $3
           (i32.load
            (i32.add
             (get_local $1)
@@ -934,21 +917,21 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 8)
            )
            (i32.shl
-            (get_local $4)
+            (get_local $3)
             (i32.const 24)
            )
           )
          )
-         (set_local $5
+         (set_local $4
           (i32.load
            (i32.add
             (get_local $1)
@@ -958,21 +941,21 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.or
            (i32.shr_u
-            (get_local $4)
+            (get_local $3)
             (i32.const 8)
            )
            (i32.shl
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
           )
          )
-         (set_local $4
+         (set_local $3
           (i32.load
            (i32.add
             (get_local $1)
@@ -982,16 +965,16 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 8)
            )
            (i32.shl
-            (get_local $4)
+            (get_local $3)
             (i32.const 24)
            )
           )
@@ -1002,9 +985,9 @@
            (i32.const 16)
           )
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -1031,449 +1014,449 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -1487,225 +1470,225 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -1719,113 +1702,113 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -1839,57 +1822,57 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -1902,40 +1885,361 @@
    )
    (i32.store8
     (block (result i32)
-     (set_local $6
-      (get_local $3)
+     (set_local $5
+      (get_local $0)
      )
-     (set_local $3
+     (set_local $0
       (i32.add
-       (get_local $6)
+       (get_local $5)
        (i32.const 1)
       )
      )
-     (get_local $6)
+     (get_local $5)
     )
     (i32.load8_u
      (block (result i32)
-      (set_local $6
+      (set_local $5
        (get_local $1)
       )
       (set_local $1
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
     )
    )
   )
-  (return
-   (get_local $0)
+ )
+ (func $std:heap/move_memory (; 2 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (local $3 i32)
+  (if
+   (i32.eq
+    (get_local $0)
+    (get_local $1)
+   )
+   (return)
+  )
+  (if
+   (i32.and
+    (if (result i32)
+     (i32.ne
+      (tee_local $3
+       (i32.le_u
+        (i32.add
+         (get_local $1)
+         (get_local $2)
+        )
+        (get_local $0)
+       )
+      )
+      (i32.const 0)
+     )
+     (get_local $3)
+     (i32.le_u
+      (i32.add
+       (get_local $0)
+       (get_local $2)
+      )
+      (get_local $1)
+     )
+    )
+    (i32.const 1)
+   )
+   (block
+    (call $std:heap/copy_memory
+     (get_local $0)
+     (get_local $1)
+     (get_local $2)
+    )
+    (return)
+   )
+  )
+  (if
+   (i32.lt_u
+    (get_local $0)
+    (get_local $1)
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (block $break|0
+       (loop $continue|0
+        (if
+         (i32.rem_u
+          (get_local $0)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (if
+            (i32.eqz
+             (get_local $2)
+            )
+            (return)
+           )
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 1)
+            )
+           )
+           (i32.store8
+            (block (result i32)
+             (set_local $3
+              (get_local $0)
+             )
+             (set_local $0
+              (i32.add
+               (get_local $3)
+               (i32.const 1)
+              )
+             )
+             (get_local $3)
+            )
+            (i32.load8_u
+             (block (result i32)
+              (set_local $3
+               (get_local $1)
+              )
+              (set_local $1
+               (i32.add
+                (get_local $3)
+                (i32.const 1)
+               )
+              )
+              (get_local $3)
+             )
+            )
+           )
+          )
+          (br $continue|0)
+         )
+        )
+       )
+      )
+      (block $break|1
+       (loop $continue|1
+        (if
+         (i32.ge_u
+          (get_local $2)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (i64.store
+            (get_local $0)
+            (i64.load
+             (get_local $1)
+            )
+           )
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+           (set_local $0
+            (i32.add
+             (get_local $0)
+             (i32.const 8)
+            )
+           )
+           (set_local $1
+            (i32.add
+             (get_local $1)
+             (i32.const 8)
+            )
+           )
+          )
+          (br $continue|1)
+         )
+        )
+       )
+      )
+     )
+    )
+    (block $break|2
+     (loop $continue|2
+      (if
+       (get_local $2)
+       (block
+        (block
+         (i32.store8
+          (block (result i32)
+           (set_local $3
+            (get_local $0)
+           )
+           (set_local $0
+            (i32.add
+             (get_local $3)
+             (i32.const 1)
+            )
+           )
+           (get_local $3)
+          )
+          (i32.load8_u
+           (block (result i32)
+            (set_local $3
+             (get_local $1)
+            )
+            (set_local $1
+             (i32.add
+              (get_local $3)
+              (i32.const 1)
+             )
+            )
+            (get_local $3)
+           )
+          )
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+        )
+        (br $continue|2)
+       )
+      )
+     )
+    )
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (block $break|3
+       (loop $continue|3
+        (if
+         (i32.rem_u
+          (i32.add
+           (get_local $0)
+           (get_local $2)
+          )
+          (i32.const 8)
+         )
+         (block
+          (block
+           (if
+            (i32.eqz
+             (get_local $2)
+            )
+            (return)
+           )
+           (i32.store8
+            (i32.add
+             (get_local $0)
+             (tee_local $2
+              (i32.sub
+               (get_local $2)
+               (i32.const 1)
+              )
+             )
+            )
+            (i32.load8_u
+             (i32.add
+              (get_local $1)
+              (get_local $2)
+             )
+            )
+           )
+          )
+          (br $continue|3)
+         )
+        )
+       )
+      )
+      (block $break|4
+       (loop $continue|4
+        (if
+         (i32.ge_u
+          (get_local $2)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+           (i64.store
+            (i32.add
+             (get_local $0)
+             (get_local $2)
+            )
+            (i64.load
+             (i32.add
+              (get_local $1)
+              (get_local $2)
+             )
+            )
+           )
+          )
+          (br $continue|4)
+         )
+        )
+       )
+      )
+     )
+    )
+    (block $break|5
+     (loop $continue|5
+      (if
+       (get_local $2)
+       (block
+        (i32.store8
+         (i32.add
+          (get_local $0)
+          (tee_local $2
+           (i32.sub
+            (get_local $2)
+            (i32.const 1)
+           )
+          )
+         )
+         (i32.load8_u
+          (i32.add
+           (get_local $1)
+           (get_local $2)
+          )
+         )
+        )
+        (br $continue|5)
+       )
+      )
+     )
+    )
+   )
   )
  )
- (func $std:heap/Heap.dispose (; 2 ;) (type $iv) (param $0 i32)
+ (func $std:heap/free_memory (; 3 ;) (type $iv) (param $0 i32)
  )
- (func $std:array/Array#__grow (; 3 ;) (type $iiv) (param $0 i32) (param $1 i32)
+ (func $std:array/Array#__grow (; 4 ;) (type $iiv) (param $0 i32) (param $1 i32)
   (local $2 i32)
   (if
    (i32.eqz
@@ -1950,7 +2254,7 @@
   )
   (block
    (set_local $2
-    (call $std:heap/Heap.allocate
+    (call $std:heap/allocate_memory
      (i32.mul
       (get_local $1)
       (i32.const 4)
@@ -1962,8 +2266,8 @@
    (i32.load
     (get_local $0)
    )
-   (drop
-    (call $std:heap/Heap.copy
+   (block
+    (call $std:heap/move_memory
      (get_local $2)
      (i32.load
       (get_local $0)
@@ -1975,11 +2279,11 @@
       (i32.const 4)
      )
     )
-   )
-  )
-  (call $std:heap/Heap.dispose
-   (i32.load
-    (get_local $0)
+    (call $std:heap/free_memory
+     (i32.load
+      (get_local $0)
+     )
+    )
    )
   )
   (i32.store
@@ -1991,7 +2295,7 @@
    (get_local $1)
   )
  )
- (func $std:array/Array#push (; 4 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#push (; 5 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
   (if
@@ -2015,11 +2319,11 @@
       )
      )
      (tee_local $3
-      (i32.mul
+      (i32.shl
        (i32.load offset=4
         (get_local $0)
        )
-       (i32.const 2)
+       (i32.const 1)
       )
      )
      (i32.gt_s
@@ -2061,7 +2365,7 @@
    )
   )
  )
- (func $std:array/Array#__get (; 5 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#__get (; 6 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (if
    (i32.ge_u
     (get_local $1)
@@ -2085,7 +2389,7 @@
    )
   )
  )
- (func $std:array/Array#pop (; 6 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:array/Array#pop (; 7 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (if
    (i32.and
@@ -2140,7 +2444,7 @@
    )
   )
  )
- (func $std:array/Array#unshift (; 7 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+ (func $std:array/Array#unshift (; 8 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
   (local $4 i32)
@@ -2259,41 +2563,24 @@
    )
   )
  )
- (func $std:heap/Heap.fill (; 8 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/set_memory (; 9 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
-  (local $6 i64)
-  (if
-   (i32.eqz
-    (i32.ge_u
-     (get_local $0)
-     (get_global $HEAP_BASE)
-    )
-   )
-   (unreachable)
-  )
+  (local $5 i64)
   (if
    (i32.eqz
     (get_local $2)
    )
-   (return
-    (get_local $0)
-   )
-  )
-  (block
-   (set_local $3
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
-   (get_local $3)
+   (get_local $0)
    (get_local $1)
   )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 1)
@@ -2305,38 +2592,36 @@
     (get_local $2)
     (i32.const 2)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 1)
    )
    (get_local $1)
   )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 2)
    )
    (get_local $1)
   )
-  (i32.store8
-   (i32.add
-    (get_local $3)
-    (i32.const 2)
-   )
-   (get_local $1)
-  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 3)
@@ -2348,13 +2633,11 @@
     (get_local $2)
     (i32.const 6)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 3)
    )
    (get_local $1)
@@ -2362,7 +2645,7 @@
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 4)
@@ -2374,31 +2657,29 @@
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (block
-   (set_local $4
+   (set_local $3
     (i32.and
      (i32.sub
       (i32.const 0)
-      (get_local $3)
+      (get_local $0)
      )
      (i32.const 3)
     )
    )
   )
-  (set_local $3
+  (set_local $0
    (i32.add
+    (get_local $0)
     (get_local $3)
-    (get_local $4)
    )
   )
   (set_local $2
    (i32.sub
     (get_local $2)
-    (get_local $4)
+    (get_local $3)
    )
   )
   (set_local $2
@@ -2411,7 +2692,7 @@
    )
   )
   (block
-   (set_local $5
+   (set_local $4
     (i32.mul
      (i32.div_u
       (i32.sub
@@ -2425,169 +2706,165 @@
    )
   )
   (i32.store
-   (get_local $3)
-   (get_local $5)
+   (get_local $0)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 4)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (if
    (i32.le_u
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 4)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 8)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 12)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 8)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (if
    (i32.le_u
     (get_local $2)
     (i32.const 24)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 12)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 16)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 20)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 24)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 28)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 24)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 20)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 16)
    )
-   (get_local $5)
+   (get_local $4)
   )
-  (set_local $4
+  (set_local $3
    (i32.add
     (i32.const 24)
     (i32.and
-     (get_local $3)
+     (get_local $0)
      (i32.const 4)
     )
    )
   )
-  (set_local $3
+  (set_local $0
    (i32.add
+    (get_local $0)
     (get_local $3)
-    (get_local $4)
    )
   )
   (set_local $2
    (i32.sub
     (get_local $2)
-    (get_local $4)
+    (get_local $3)
    )
   )
   (block
-   (set_local $6
+   (set_local $5
     (i64.or
      (i64.extend_u/i32
-      (get_local $5)
+      (get_local $4)
      )
      (i64.shl
       (i64.extend_u/i32
-       (get_local $5)
+       (get_local $4)
       )
       (i64.const 32)
      )
@@ -2604,29 +2881,29 @@
      (block
       (block
        (i64.store
-        (get_local $3)
-        (get_local $6)
+        (get_local $0)
+        (get_local $5)
        )
        (i64.store
         (i32.add
-         (get_local $3)
+         (get_local $0)
          (i32.const 8)
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i64.store
         (i32.add
-         (get_local $3)
+         (get_local $0)
          (i32.const 16)
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i64.store
         (i32.add
-         (get_local $3)
+         (get_local $0)
          (i32.const 24)
         )
-        (get_local $6)
+        (get_local $5)
        )
        (set_local $2
         (i32.sub
@@ -2634,9 +2911,9 @@
          (i32.const 32)
         )
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $3)
+         (get_local $0)
          (i32.const 32)
         )
        )
@@ -2646,11 +2923,8 @@
     )
    )
   )
-  (return
-   (get_local $0)
-  )
  )
- (func $std:array/Array#shift (; 9 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:array/Array#shift (; 10 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
   (if
@@ -2690,16 +2964,30 @@
     )
    )
   )
-  (drop
-   (call $std:heap/Heap.copy
+  (call $std:heap/move_memory
+   (i32.load
+    (get_local $0)
+   )
+   (i32.add
     (i32.load
      (get_local $0)
     )
-    (i32.add
-     (i32.load
+    (i32.const 4)
+   )
+   (i32.mul
+    (i32.sub
+     (i32.load offset=4
       (get_local $0)
      )
-     (i32.const 4)
+     (i32.const 1)
+    )
+    (i32.const 4)
+   )
+  )
+  (call $std:heap/set_memory
+   (i32.add
+    (i32.load
+     (get_local $0)
     )
     (i32.mul
      (i32.sub
@@ -2711,26 +2999,8 @@
      (i32.const 4)
     )
    )
-  )
-  (drop
-   (call $std:heap/Heap.fill
-    (i32.add
-     (i32.load
-      (get_local $0)
-     )
-     (i32.mul
-      (i32.sub
-       (i32.load offset=4
-        (get_local $0)
-       )
-       (i32.const 1)
-      )
-      (i32.const 4)
-     )
-    )
-    (i32.const 0)
-    (i32.const 4)
-   )
+   (i32.const 0)
+   (i32.const 4)
   )
   (i32.store offset=8
    (get_local $0)
@@ -2745,12 +3015,12 @@
    (get_local $2)
   )
  )
- (func $start (; 10 ;) (type $v)
+ (func $start (; 11 ;) (type $v)
   (set_global $std:heap/HEAP_OFFSET
    (get_global $HEAP_BASE)
   )
   (set_global $std/array/arr
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.add
      (i32.const 4)
      (i32.mul
@@ -3392,16 +3662,17 @@
   GLOBAL: std:heap/ALIGN_SIZE
   GLOBAL: std:heap/ALIGN_MASK
   GLOBAL: std:heap/HEAP_OFFSET
-  CLASS_PROTOTYPE: std:heap/Heap
-  CLASS_PROTOTYPE: Heap
-  PROPERTY: std:heap/Heap.used
-  PROPERTY: std:heap/Heap.free
-  PROPERTY: std:heap/Heap.size
-  FUNCTION_PROTOTYPE: std:heap/Heap.allocate
-  FUNCTION_PROTOTYPE: std:heap/Heap.dispose
-  FUNCTION_PROTOTYPE: std:heap/Heap.copy
-  FUNCTION_PROTOTYPE: std:heap/Heap.fill
-  FUNCTION_PROTOTYPE: std:heap/Heap.compare
+  FUNCTION_PROTOTYPE: std:heap/allocate_memory
+  FUNCTION_PROTOTYPE: allocate_memory
+  FUNCTION_PROTOTYPE: std:heap/free_memory
+  FUNCTION_PROTOTYPE: free_memory
+  FUNCTION_PROTOTYPE: std:heap/copy_memory
+  FUNCTION_PROTOTYPE: std:heap/move_memory
+  FUNCTION_PROTOTYPE: move_memory
+  FUNCTION_PROTOTYPE: std:heap/set_memory
+  FUNCTION_PROTOTYPE: set_memory
+  FUNCTION_PROTOTYPE: std:heap/compare_memory
+  FUNCTION_PROTOTYPE: compare_memory
   CLASS_PROTOTYPE: std:map/Map
   CLASS_PROTOTYPE: Map
   CLASS_PROTOTYPE: std:regexp/RegExp
@@ -3423,7 +3694,11 @@
   CLASS_PROTOTYPE: std:array/CArray
   CLASS_PROTOTYPE: std:error/Error
   CLASS_PROTOTYPE: std:error/RangeError
-  CLASS_PROTOTYPE: std:heap/Heap
+  FUNCTION_PROTOTYPE: std:heap/allocate_memory
+  FUNCTION_PROTOTYPE: std:heap/free_memory
+  FUNCTION_PROTOTYPE: std:heap/move_memory
+  FUNCTION_PROTOTYPE: std:heap/set_memory
+  FUNCTION_PROTOTYPE: std:heap/compare_memory
   CLASS_PROTOTYPE: std:map/Map
   CLASS_PROTOTYPE: std:regexp/RegExp
   CLASS_PROTOTYPE: std:set/Set
diff --git a/tests/compiler/std/carray.optimized.wast b/tests/compiler/std/carray.optimized.wast
index 4247e770..49d3f567 100644
--- a/tests/compiler/std/carray.optimized.wast
+++ b/tests/compiler/std/carray.optimized.wast
@@ -8,6 +8,13 @@
  (export "memory" (memory $0))
  (start $start)
  (func $std:array/CArray#__get (; 0 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+  (if
+   (i32.lt_s
+    (get_local $1)
+    (i32.const 0)
+   )
+   (unreachable)
+  )
   (i32.load
    (i32.add
     (get_local $0)
@@ -19,6 +26,13 @@
   )
  )
  (func $std:array/CArray#__set (; 1 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (if
+   (i32.lt_s
+    (get_local $1)
+    (i32.const 0)
+   )
+   (unreachable)
+  )
   (i32.store
    (i32.add
     (get_local $0)
diff --git a/tests/compiler/std/carray.wast b/tests/compiler/std/carray.wast
index 651ea64b..73092756 100644
--- a/tests/compiler/std/carray.wast
+++ b/tests/compiler/std/carray.wast
@@ -8,6 +8,13 @@
  (export "memory" (memory $0))
  (start $start)
  (func $std:array/CArray#__get (; 0 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
+  (if
+   (i32.lt_s
+    (get_local $1)
+    (i32.const 0)
+   )
+   (unreachable)
+  )
   (return
    (i32.load
     (i32.add
@@ -21,6 +28,13 @@
   )
  )
  (func $std:array/CArray#__set (; 1 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (if
+   (i32.lt_s
+    (get_local $1)
+    (i32.const 0)
+   )
+   (unreachable)
+  )
   (i32.store
    (i32.add
     (get_local $0)
@@ -247,16 +261,17 @@
   GLOBAL: std:heap/ALIGN_SIZE
   GLOBAL: std:heap/ALIGN_MASK
   GLOBAL: std:heap/HEAP_OFFSET
-  CLASS_PROTOTYPE: std:heap/Heap
-  CLASS_PROTOTYPE: Heap
-  PROPERTY: std:heap/Heap.used
-  PROPERTY: std:heap/Heap.free
-  PROPERTY: std:heap/Heap.size
-  FUNCTION_PROTOTYPE: std:heap/Heap.allocate
-  FUNCTION_PROTOTYPE: std:heap/Heap.dispose
-  FUNCTION_PROTOTYPE: std:heap/Heap.copy
-  FUNCTION_PROTOTYPE: std:heap/Heap.fill
-  FUNCTION_PROTOTYPE: std:heap/Heap.compare
+  FUNCTION_PROTOTYPE: std:heap/allocate_memory
+  FUNCTION_PROTOTYPE: allocate_memory
+  FUNCTION_PROTOTYPE: std:heap/free_memory
+  FUNCTION_PROTOTYPE: free_memory
+  FUNCTION_PROTOTYPE: std:heap/copy_memory
+  FUNCTION_PROTOTYPE: std:heap/move_memory
+  FUNCTION_PROTOTYPE: move_memory
+  FUNCTION_PROTOTYPE: std:heap/set_memory
+  FUNCTION_PROTOTYPE: set_memory
+  FUNCTION_PROTOTYPE: std:heap/compare_memory
+  FUNCTION_PROTOTYPE: compare_memory
   CLASS_PROTOTYPE: std:map/Map
   CLASS_PROTOTYPE: Map
   CLASS_PROTOTYPE: std:regexp/RegExp
@@ -277,7 +292,11 @@
   CLASS_PROTOTYPE: std:array/CArray
   CLASS_PROTOTYPE: std:error/Error
   CLASS_PROTOTYPE: std:error/RangeError
-  CLASS_PROTOTYPE: std:heap/Heap
+  FUNCTION_PROTOTYPE: std:heap/allocate_memory
+  FUNCTION_PROTOTYPE: std:heap/free_memory
+  FUNCTION_PROTOTYPE: std:heap/move_memory
+  FUNCTION_PROTOTYPE: std:heap/set_memory
+  FUNCTION_PROTOTYPE: std:heap/compare_memory
   CLASS_PROTOTYPE: std:map/Map
   CLASS_PROTOTYPE: std:regexp/RegExp
   CLASS_PROTOTYPE: std:set/Set
diff --git a/tests/compiler/std/heap.optimized.wast b/tests/compiler/std/heap.optimized.wast
index d9de4fb8..0d8ea488 100644
--- a/tests/compiler/std/heap.optimized.wast
+++ b/tests/compiler/std/heap.optimized.wast
@@ -1,5 +1,6 @@
 (module
  (type $ii (func (param i32) (result i32)))
+ (type $iiiv (func (param i32 i32 i32)))
  (type $iiii (func (param i32 i32 i32) (result i32)))
  (type $iv (func (param i32)))
  (type $v (func))
@@ -11,7 +12,7 @@
  (memory $0 1)
  (export "memory" (memory $0))
  (start $start)
- (func $std:heap/Heap.allocate (; 0 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:heap/allocate_memory (; 0 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
   (if
@@ -100,35 +101,23 @@
   )
   (get_local $1)
  )
- (func $std:heap/Heap.fill (; 1 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
-  (local $3 i32)
-  (local $4 i64)
-  (local $5 i32)
-  (if
-   (i32.lt_u
-    (get_local $0)
-    (get_global $HEAP_BASE)
-   )
-   (unreachable)
-  )
+ (func $std:heap/set_memory (; 1 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (local $3 i64)
+  (local $4 i32)
   (if
    (i32.eqz
     (get_local $2)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
-   (tee_local $3
-    (get_local $0)
-   )
+   (get_local $0)
    (get_local $1)
   )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 1)
@@ -140,38 +129,36 @@
     (get_local $2)
     (i32.const 2)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 1)
    )
    (get_local $1)
   )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 2)
    )
    (get_local $1)
   )
-  (i32.store8
-   (i32.add
-    (get_local $3)
-    (i32.const 2)
-   )
-   (get_local $1)
-  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 3)
@@ -183,13 +170,11 @@
     (get_local $2)
     (i32.const 6)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 3)
    )
    (get_local $1)
@@ -197,7 +182,7 @@
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 4)
@@ -209,19 +194,17 @@
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
-   (tee_local $3
+   (tee_local $0
     (i32.add
-     (get_local $3)
-     (tee_local $5
+     (get_local $0)
+     (tee_local $4
       (i32.and
        (i32.sub
         (i32.const 0)
-        (get_local $3)
+        (get_local $0)
        )
        (i32.const 3)
       )
@@ -238,12 +221,12 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (tee_local $2
       (i32.and
        (i32.sub
         (get_local $2)
-        (get_local $5)
+        (get_local $4)
        )
        (i32.const -4)
       )
@@ -258,20 +241,18 @@
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 4)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 8)
    )
    (get_local $1)
@@ -279,7 +260,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 12)
@@ -289,7 +270,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 8)
@@ -301,34 +282,32 @@
     (get_local $2)
     (i32.const 24)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 12)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 16)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 20)
    )
    (get_local $1)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 24)
    )
    (get_local $1)
@@ -336,7 +315,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 28)
@@ -346,7 +325,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 24)
@@ -356,7 +335,7 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 20)
@@ -366,20 +345,20 @@
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 16)
    )
    (get_local $1)
   )
-  (set_local $3
+  (set_local $0
    (i32.add
-    (get_local $3)
-    (tee_local $5
+    (get_local $0)
+    (tee_local $4
      (i32.add
       (i32.and
-       (get_local $3)
+       (get_local $0)
        (i32.const 4)
       )
       (i32.const 24)
@@ -390,10 +369,10 @@
   (set_local $2
    (i32.sub
     (get_local $2)
-    (get_local $5)
+    (get_local $4)
    )
   )
-  (set_local $4
+  (set_local $3
    (i64.or
     (i64.extend_u/i32
      (get_local $1)
@@ -414,29 +393,29 @@
     )
     (block
      (i64.store
+      (get_local $0)
       (get_local $3)
-      (get_local $4)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 8)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 16)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (i64.store
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 24)
       )
-      (get_local $4)
+      (get_local $3)
      )
      (set_local $2
       (i32.sub
@@ -444,9 +423,9 @@
        (i32.const 32)
       )
      )
-     (set_local $3
+     (set_local $0
       (i32.add
-       (get_local $3)
+       (get_local $0)
        (i32.const 32)
       )
      )
@@ -454,22 +433,10 @@
     )
    )
   )
-  (get_local $0)
  )
- (func $std:heap/Heap.copy (; 2 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/copy_memory (; 2 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
-  (if
-   (i32.lt_u
-    (get_local $0)
-    (get_global $HEAP_BASE)
-   )
-   (unreachable)
-  )
-  (set_local $4
-   (get_local $0)
-  )
   (loop $continue|0
    (if
     (if (result i32)
@@ -481,10 +448,10 @@
      (get_local $2)
     )
     (block
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -518,7 +485,7 @@
   (if
    (i32.eqz
     (i32.rem_u
-     (get_local $4)
+     (get_local $0)
      (i32.const 4)
     )
    )
@@ -531,14 +498,14 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.load
          (get_local $1)
         )
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.load
@@ -550,7 +517,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.load
@@ -562,7 +529,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.load
@@ -578,9 +545,9 @@
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -601,14 +568,14 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
@@ -618,9 +585,9 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 8)
        )
       )
@@ -639,14 +606,14 @@
      )
      (block
       (i32.store
-       (get_local $4)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 4)
        )
       )
@@ -665,14 +632,14 @@
      )
      (block
       (i32.store16
-       (get_local $4)
+       (get_local $0)
        (i32.load16_u
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
-        (get_local $4)
+        (get_local $0)
         (i32.const 2)
        )
       )
@@ -691,7 +658,7 @@
      )
      (block
       (set_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.store8
        (get_local $3)
@@ -706,9 +673,7 @@
       )
      )
     )
-    (return
-     (get_local $0)
-    )
+    (return)
    )
   )
   (if
@@ -724,7 +689,7 @@
         (br_table $case0|2 $case1|2 $case2|2 $tablify|0
          (i32.sub
           (i32.rem_u
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.const 1)
@@ -733,15 +698,15 @@
        )
        (br $break|2)
       )
-      (set_local $5
+      (set_local $4
        (i32.load
         (get_local $1)
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -762,10 +727,10 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -786,10 +751,10 @@
         )
        )
       )
-      (set_local $4
+      (set_local $0
        (i32.add
         (tee_local $3
-         (get_local $4)
+         (get_local $0)
         )
         (i32.const 1)
        )
@@ -824,10 +789,10 @@
         )
         (block
          (i32.store
-          (get_local $4)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
@@ -845,7 +810,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.or
@@ -854,7 +819,7 @@
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
                (get_local $1)
@@ -868,12 +833,12 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
            (i32.shl
@@ -891,7 +856,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.or
@@ -900,7 +865,7 @@
             (i32.const 24)
            )
            (i32.shl
-            (tee_local $5
+            (tee_local $4
              (i32.load
               (i32.add
                (get_local $1)
@@ -918,9 +883,9 @@
            (i32.const 16)
           )
          )
-         (set_local $4
+         (set_local $0
           (i32.add
-           (get_local $4)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -936,15 +901,15 @@
       )
       (br $break|2)
      )
-     (set_local $5
+     (set_local $4
       (i32.load
        (get_local $1)
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -965,10 +930,10 @@
        )
       )
      )
-     (set_local $4
+     (set_local $0
       (i32.add
        (tee_local $3
-        (get_local $4)
+        (get_local $0)
        )
        (i32.const 1)
       )
@@ -1003,10 +968,10 @@
        )
        (block
         (i32.store
-         (get_local $4)
+         (get_local $0)
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
@@ -1024,7 +989,7 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 4)
          )
          (i32.or
@@ -1033,7 +998,7 @@
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -1047,12 +1012,12 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 8)
          )
          (i32.or
           (i32.shr_u
-           (get_local $5)
+           (get_local $4)
            (i32.const 16)
           )
           (i32.shl
@@ -1070,7 +1035,7 @@
         )
         (i32.store
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 12)
          )
          (i32.or
@@ -1079,7 +1044,7 @@
            (i32.const 16)
           )
           (i32.shl
-           (tee_local $5
+           (tee_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -1097,9 +1062,9 @@
           (i32.const 16)
          )
         )
-        (set_local $4
+        (set_local $0
          (i32.add
-          (get_local $4)
+          (get_local $0)
           (i32.const 16)
          )
         )
@@ -1115,15 +1080,15 @@
      )
      (br $break|2)
     )
-    (set_local $5
+    (set_local $4
      (i32.load
       (get_local $1)
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1158,10 +1123,10 @@
       )
       (block
        (i32.store
-        (get_local $4)
+        (get_local $0)
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
@@ -1179,7 +1144,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 4)
         )
         (i32.or
@@ -1188,7 +1153,7 @@
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -1202,12 +1167,12 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 8)
         )
         (i32.or
          (i32.shr_u
-          (get_local $5)
+          (get_local $4)
           (i32.const 8)
          )
          (i32.shl
@@ -1225,7 +1190,7 @@
        )
        (i32.store
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 12)
         )
         (i32.or
@@ -1234,7 +1199,7 @@
           (i32.const 8)
          )
          (i32.shl
-          (tee_local $5
+          (tee_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -1252,9 +1217,9 @@
          (i32.const 16)
         )
        )
-       (set_local $4
+       (set_local $0
         (i32.add
-         (get_local $4)
+         (get_local $0)
          (i32.const 16)
         )
        )
@@ -1276,10 +1241,10 @@
     (i32.const 16)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1300,10 +1265,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1324,10 +1289,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1348,10 +1313,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1372,10 +1337,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1396,10 +1361,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1420,10 +1385,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1444,10 +1409,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1468,10 +1433,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1492,10 +1457,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1516,10 +1481,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1540,10 +1505,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1564,10 +1529,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1588,10 +1553,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1612,10 +1577,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1636,10 +1601,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1668,10 +1633,10 @@
     (i32.const 8)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1692,10 +1657,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1716,10 +1681,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1740,10 +1705,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1764,10 +1729,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1788,10 +1753,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1812,10 +1777,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1836,10 +1801,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1868,10 +1833,10 @@
     (i32.const 4)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1892,10 +1857,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1916,10 +1881,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1940,10 +1905,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1972,10 +1937,10 @@
     (i32.const 2)
    )
    (block
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -1996,10 +1961,10 @@
       )
      )
     )
-    (set_local $4
+    (set_local $0
      (i32.add
       (tee_local $3
-       (get_local $4)
+       (get_local $0)
       )
       (i32.const 1)
      )
@@ -2029,7 +1994,7 @@
    )
    (block
     (set_local $3
-     (get_local $4)
+     (get_local $0)
     )
     (i32.store8
      (get_local $3)
@@ -2044,9 +2009,298 @@
     )
    )
   )
-  (get_local $0)
  )
- (func $std:heap/Heap.compare (; 3 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/move_memory (; 3 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (local $3 i32)
+  (if
+   (i32.eq
+    (get_local $0)
+    (get_local $1)
+   )
+   (return)
+  )
+  (if
+   (i32.and
+    (if (result i32)
+     (tee_local $3
+      (i32.le_u
+       (i32.add
+        (get_local $1)
+        (get_local $2)
+       )
+       (get_local $0)
+      )
+     )
+     (get_local $3)
+     (i32.le_u
+      (i32.add
+       (get_local $0)
+       (get_local $2)
+      )
+      (get_local $1)
+     )
+    )
+    (i32.const 1)
+   )
+   (block
+    (call $std:heap/copy_memory
+     (get_local $0)
+     (get_local $1)
+     (get_local $2)
+    )
+    (return)
+   )
+  )
+  (if
+   (i32.lt_u
+    (get_local $0)
+    (get_local $1)
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (loop $continue|0
+       (if
+        (i32.rem_u
+         (get_local $0)
+         (i32.const 8)
+        )
+        (block
+         (if
+          (i32.eqz
+           (get_local $2)
+          )
+          (return)
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+         (set_local $0
+          (i32.add
+           (tee_local $3
+            (get_local $0)
+           )
+           (i32.const 1)
+          )
+         )
+         (i32.store8
+          (get_local $3)
+          (block (result i32)
+           (set_local $1
+            (i32.add
+             (tee_local $3
+              (get_local $1)
+             )
+             (i32.const 1)
+            )
+           )
+           (i32.load8_u
+            (get_local $3)
+           )
+          )
+         )
+         (br $continue|0)
+        )
+       )
+      )
+      (loop $continue|1
+       (if
+        (i32.ge_u
+         (get_local $2)
+         (i32.const 8)
+        )
+        (block
+         (i64.store
+          (get_local $0)
+          (i64.load
+           (get_local $1)
+          )
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 8)
+          )
+         )
+         (set_local $0
+          (i32.add
+           (get_local $0)
+           (i32.const 8)
+          )
+         )
+         (set_local $1
+          (i32.add
+           (get_local $1)
+           (i32.const 8)
+          )
+         )
+         (br $continue|1)
+        )
+       )
+      )
+     )
+    )
+    (loop $continue|2
+     (if
+      (get_local $2)
+      (block
+       (set_local $0
+        (i32.add
+         (tee_local $3
+          (get_local $0)
+         )
+         (i32.const 1)
+        )
+       )
+       (i32.store8
+        (get_local $3)
+        (block (result i32)
+         (set_local $1
+          (i32.add
+           (tee_local $3
+            (get_local $1)
+           )
+           (i32.const 1)
+          )
+         )
+         (i32.load8_u
+          (get_local $3)
+         )
+        )
+       )
+       (set_local $2
+        (i32.sub
+         (get_local $2)
+         (i32.const 1)
+        )
+       )
+       (br $continue|2)
+      )
+     )
+    )
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (loop $continue|3
+       (if
+        (i32.rem_u
+         (i32.add
+          (get_local $0)
+          (get_local $2)
+         )
+         (i32.const 8)
+        )
+        (block
+         (if
+          (i32.eqz
+           (get_local $2)
+          )
+          (return)
+         )
+         (i32.store8
+          (i32.add
+           (get_local $0)
+           (tee_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 1)
+            )
+           )
+          )
+          (i32.load8_u
+           (i32.add
+            (get_local $1)
+            (get_local $2)
+           )
+          )
+         )
+         (br $continue|3)
+        )
+       )
+      )
+      (loop $continue|4
+       (if
+        (i32.ge_u
+         (get_local $2)
+         (i32.const 8)
+        )
+        (block
+         (i64.store
+          (i32.add
+           (get_local $0)
+           (tee_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+          )
+          (i64.load
+           (i32.add
+            (get_local $1)
+            (get_local $2)
+           )
+          )
+         )
+         (br $continue|4)
+        )
+       )
+      )
+     )
+    )
+    (loop $continue|5
+     (if
+      (get_local $2)
+      (block
+       (i32.store8
+        (i32.add
+         (get_local $0)
+         (tee_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+        )
+        (i32.load8_u
+         (i32.add
+          (get_local $1)
+          (get_local $2)
+         )
+        )
+       )
+       (br $continue|5)
+      )
+     )
+    )
+   )
+  )
+ )
+ (func $std:heap/compare_memory (; 4 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
   (if
    (i32.eq
     (get_local $0)
@@ -2106,20 +2360,20 @@
    (i32.const 0)
   )
  )
- (func $std:heap/Heap.dispose (; 4 ;) (type $iv) (param $0 i32)
+ (func $std:heap/free_memory (; 5 ;) (type $iv) (param $0 i32)
   (nop)
  )
- (func $start (; 5 ;) (type $v)
+ (func $start (; 6 ;) (type $v)
   (set_global $std:heap/HEAP_OFFSET
    (get_global $HEAP_BASE)
   )
   (set_global $std/heap/ptr1
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.const 42)
    )
   )
   (set_global $std/heap/ptr2
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.const 42)
    )
   )
@@ -2130,12 +2384,10 @@
    )
    (unreachable)
   )
-  (drop
-   (call $std:heap/Heap.fill
-    (get_global $std/heap/ptr1)
-    (i32.const 18)
-    (i32.const 42)
-   )
+  (call $std:heap/set_memory
+   (get_global $std/heap/ptr1)
+   (i32.const 18)
+   (i32.const 42)
   )
   (set_global $std/heap/i
    (i32.const 0)
@@ -2169,12 +2421,10 @@
     )
    )
   )
-  (drop
-   (call $std:heap/Heap.copy
-    (get_global $std/heap/ptr2)
-    (get_global $std/heap/ptr1)
-    (i32.const 42)
-   )
+  (call $std:heap/move_memory
+   (get_global $std/heap/ptr2)
+   (get_global $std/heap/ptr1)
+   (i32.const 42)
   )
   (set_global $std/heap/i
    (i32.const 0)
@@ -2209,17 +2459,17 @@
    )
   )
   (if
-   (call $std:heap/Heap.compare
+   (call $std:heap/compare_memory
     (get_global $std/heap/ptr1)
     (get_global $std/heap/ptr2)
     (i32.const 42)
    )
    (unreachable)
   )
-  (call $std:heap/Heap.dispose
+  (call $std:heap/free_memory
    (get_global $std/heap/ptr1)
   )
-  (call $std:heap/Heap.dispose
+  (call $std:heap/free_memory
    (get_global $std/heap/ptr2)
   )
  )
diff --git a/tests/compiler/std/heap.ts b/tests/compiler/std/heap.ts
index 8f45bd08..05a8c7a1 100644
--- a/tests/compiler/std/heap.ts
+++ b/tests/compiler/std/heap.ts
@@ -1,21 +1,21 @@
 const size: usize = 42;
-let ptr1: usize = Heap.allocate(size);
-let ptr2: usize = Heap.allocate(size);
+let ptr1: usize = allocate_memory(size);
+let ptr2: usize = allocate_memory(size);
 
 assert(ptr1 != ptr2);
 
-Heap.fill(ptr1, 0x12, size);
+set_memory(ptr1, 0x12, size);
 
 let i: usize;
 for (i = 0; i < size; ++i)
   assert(load<u8>(ptr1 + i) == 0x12);
 
-Heap.copy(ptr2, ptr1, size);
+move_memory(ptr2, ptr1, size);
 
 for (i = 0; i < size; ++i)
   assert(load<u8>(ptr2 + i) == 0x12);
 
-assert(Heap.compare(ptr1, ptr2, size) == 0);
+assert(compare_memory(ptr1, ptr2, size) == 0);
 
-Heap.dispose(ptr1);
-Heap.dispose(ptr2);
+free_memory(ptr1);
+free_memory(ptr2);
diff --git a/tests/compiler/std/heap.wast b/tests/compiler/std/heap.wast
index e12a33ea..dd6957a6 100644
--- a/tests/compiler/std/heap.wast
+++ b/tests/compiler/std/heap.wast
@@ -1,6 +1,7 @@
 (module
  (type $i (func (result i32)))
  (type $ii (func (param i32) (result i32)))
+ (type $iiiv (func (param i32 i32 i32)))
  (type $iiii (func (param i32 i32 i32) (result i32)))
  (type $iv (func (param i32)))
  (type $v (func))
@@ -16,7 +17,7 @@
  (memory $0 1)
  (export "memory" (memory $0))
  (start $start)
- (func $std:heap/Heap.allocate (; 0 ;) (type $ii) (param $0 i32) (result i32)
+ (func $std:heap/allocate_memory (; 0 ;) (type $ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
   (local $3 i32)
@@ -113,41 +114,24 @@
    (get_local $4)
   )
  )
- (func $std:heap/Heap.fill (; 1 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/set_memory (; 1 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
   (local $4 i32)
-  (local $5 i32)
-  (local $6 i64)
-  (if
-   (i32.eqz
-    (i32.ge_u
-     (get_local $0)
-     (get_global $HEAP_BASE)
-    )
-   )
-   (unreachable)
-  )
+  (local $5 i64)
   (if
    (i32.eqz
     (get_local $2)
    )
-   (return
-    (get_local $0)
-   )
-  )
-  (block
-   (set_local $3
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
-   (get_local $3)
+   (get_local $0)
    (get_local $1)
   )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 1)
@@ -159,38 +143,36 @@
     (get_local $2)
     (i32.const 2)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 1)
    )
    (get_local $1)
   )
+  (i32.store8
+   (i32.add
+    (get_local $0)
+    (i32.const 2)
+   )
+   (get_local $1)
+  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 2)
    )
    (get_local $1)
   )
-  (i32.store8
-   (i32.add
-    (get_local $3)
-    (i32.const 2)
-   )
-   (get_local $1)
-  )
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 3)
@@ -202,13 +184,11 @@
     (get_local $2)
     (i32.const 6)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store8
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 3)
    )
    (get_local $1)
@@ -216,7 +196,7 @@
   (i32.store8
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 4)
@@ -228,31 +208,29 @@
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (block
-   (set_local $4
+   (set_local $3
     (i32.and
      (i32.sub
       (i32.const 0)
-      (get_local $3)
+      (get_local $0)
      )
      (i32.const 3)
     )
    )
   )
-  (set_local $3
+  (set_local $0
    (i32.add
+    (get_local $0)
     (get_local $3)
-    (get_local $4)
    )
   )
   (set_local $2
    (i32.sub
     (get_local $2)
-    (get_local $4)
+    (get_local $3)
    )
   )
   (set_local $2
@@ -265,7 +243,7 @@
    )
   )
   (block
-   (set_local $5
+   (set_local $4
     (i32.mul
      (i32.div_u
       (i32.sub
@@ -279,169 +257,165 @@
    )
   )
   (i32.store
-   (get_local $3)
-   (get_local $5)
+   (get_local $0)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 4)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (if
    (i32.le_u
     (get_local $2)
     (i32.const 8)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 4)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 8)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 12)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 8)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (if
    (i32.le_u
     (get_local $2)
     (i32.const 24)
    )
-   (return
-    (get_local $0)
-   )
+   (return)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 12)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 16)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 20)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.add
-    (get_local $3)
+    (get_local $0)
     (i32.const 24)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 28)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 24)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 20)
    )
-   (get_local $5)
+   (get_local $4)
   )
   (i32.store
    (i32.sub
     (i32.add
-     (get_local $3)
+     (get_local $0)
      (get_local $2)
     )
     (i32.const 16)
    )
-   (get_local $5)
+   (get_local $4)
   )
-  (set_local $4
+  (set_local $3
    (i32.add
     (i32.const 24)
     (i32.and
-     (get_local $3)
+     (get_local $0)
      (i32.const 4)
     )
    )
   )
-  (set_local $3
+  (set_local $0
    (i32.add
+    (get_local $0)
     (get_local $3)
-    (get_local $4)
    )
   )
   (set_local $2
    (i32.sub
     (get_local $2)
-    (get_local $4)
+    (get_local $3)
    )
   )
   (block
-   (set_local $6
+   (set_local $5
     (i64.or
      (i64.extend_u/i32
-      (get_local $5)
+      (get_local $4)
      )
      (i64.shl
       (i64.extend_u/i32
-       (get_local $5)
+       (get_local $4)
       )
       (i64.const 32)
      )
@@ -458,29 +432,29 @@
      (block
       (block
        (i64.store
-        (get_local $3)
-        (get_local $6)
+        (get_local $0)
+        (get_local $5)
        )
        (i64.store
         (i32.add
-         (get_local $3)
+         (get_local $0)
          (i32.const 8)
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i64.store
         (i32.add
-         (get_local $3)
+         (get_local $0)
          (i32.const 16)
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i64.store
         (i32.add
-         (get_local $3)
+         (get_local $0)
          (i32.const 24)
         )
-        (get_local $6)
+        (get_local $5)
        )
        (set_local $2
         (i32.sub
@@ -488,9 +462,9 @@
          (i32.const 32)
         )
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $3)
+         (get_local $0)
          (i32.const 32)
         )
        )
@@ -500,29 +474,11 @@
     )
    )
   )
-  (return
-   (get_local $0)
-  )
  )
- (func $std:heap/Heap.copy (; 2 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/copy_memory (; 2 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  (local $6 i32)
-  (if
-   (i32.eqz
-    (i32.ge_u
-     (get_local $0)
-     (get_global $HEAP_BASE)
-    )
-   )
-   (unreachable)
-  )
-  (block
-   (set_local $3
-    (get_local $0)
-   )
-  )
   (nop)
   (block $break|0
    (loop $continue|0
@@ -542,29 +498,29 @@
       (block
        (i32.store8
         (block (result i32)
-         (set_local $6
-          (get_local $3)
+         (set_local $5
+          (get_local $0)
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $6)
+           (get_local $5)
            (i32.const 1)
           )
          )
-         (get_local $6)
+         (get_local $5)
         )
         (i32.load8_u
          (block (result i32)
-          (set_local $6
+          (set_local $5
            (get_local $1)
           )
           (set_local $1
            (i32.add
-            (get_local $6)
+            (get_local $5)
             (i32.const 1)
            )
           )
-          (get_local $6)
+          (get_local $5)
          )
         )
        )
@@ -583,7 +539,7 @@
   (if
    (i32.eq
     (i32.rem_u
-     (get_local $3)
+     (get_local $0)
      (i32.const 4)
     )
     (i32.const 0)
@@ -599,14 +555,14 @@
        (block
         (block
          (i32.store
-          (get_local $3)
+          (get_local $0)
           (i32.load
            (get_local $1)
           )
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.load
@@ -618,7 +574,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.load
@@ -630,7 +586,7 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.load
@@ -646,9 +602,9 @@
            (i32.const 16)
           )
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -671,14 +627,14 @@
      )
      (block
       (i32.store
-       (get_local $3)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
       (i32.store
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 4)
        )
        (i32.load
@@ -688,9 +644,9 @@
         )
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 8)
        )
       )
@@ -709,14 +665,14 @@
      )
      (block
       (i32.store
-       (get_local $3)
+       (get_local $0)
        (i32.load
         (get_local $1)
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 4)
        )
       )
@@ -735,14 +691,14 @@
      )
      (block
       (i32.store16
-       (get_local $3)
+       (get_local $0)
        (i32.load16_u
         (get_local $1)
        )
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $3)
+        (get_local $0)
         (i32.const 2)
        )
       )
@@ -761,36 +717,34 @@
      )
      (i32.store8
       (block (result i32)
-       (set_local $6
-        (get_local $3)
+       (set_local $5
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $6
+        (set_local $5
          (get_local $1)
         )
         (set_local $1
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
       )
      )
     )
-    (return
-     (get_local $0)
-    )
+    (return)
    )
   )
   (if
@@ -802,118 +756,118 @@
     (block $case2|2
      (block $case1|2
       (block $case0|2
-       (set_local $6
+       (set_local $5
         (i32.rem_u
-         (get_local $3)
+         (get_local $0)
          (i32.const 4)
         )
        )
        (br_if $case0|2
         (i32.eq
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
        (br_if $case1|2
         (i32.eq
-         (get_local $6)
+         (get_local $5)
          (i32.const 2)
         )
        )
        (br_if $case2|2
         (i32.eq
-         (get_local $6)
+         (get_local $5)
          (i32.const 3)
         )
        )
        (br $break|2)
       )
-      (set_local $4
+      (set_local $3
        (i32.load
         (get_local $1)
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $6
-         (get_local $3)
+        (set_local $5
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $6
+         (set_local $5
           (get_local $1)
          )
          (set_local $1
           (i32.add
-           (get_local $6)
+           (get_local $5)
            (i32.const 1)
           )
          )
-         (get_local $6)
+         (get_local $5)
         )
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $6
-         (get_local $3)
+        (set_local $5
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $6
+         (set_local $5
           (get_local $1)
          )
          (set_local $1
           (i32.add
-           (get_local $6)
+           (get_local $5)
            (i32.const 1)
           )
          )
-         (get_local $6)
+         (get_local $5)
         )
        )
       )
       (i32.store8
        (block (result i32)
-        (set_local $6
-         (get_local $3)
+        (set_local $5
+         (get_local $0)
         )
-        (set_local $3
+        (set_local $0
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
        (i32.load8_u
         (block (result i32)
-         (set_local $6
+         (set_local $5
           (get_local $1)
          )
          (set_local $1
           (i32.add
-           (get_local $6)
+           (get_local $5)
            (i32.const 1)
           )
          )
-         (get_local $6)
+         (get_local $5)
         )
        )
       )
@@ -932,7 +886,7 @@
          )
          (block
           (block
-           (set_local $5
+           (set_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -941,19 +895,19 @@
             )
            )
            (i32.store
-            (get_local $3)
+            (get_local $0)
             (i32.or
              (i32.shr_u
-              (get_local $4)
+              (get_local $3)
               (i32.const 24)
              )
              (i32.shl
-              (get_local $5)
+              (get_local $4)
               (i32.const 8)
              )
             )
            )
-           (set_local $4
+           (set_local $3
             (i32.load
              (i32.add
               (get_local $1)
@@ -963,21 +917,21 @@
            )
            (i32.store
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 4)
             )
             (i32.or
              (i32.shr_u
-              (get_local $5)
+              (get_local $4)
               (i32.const 24)
              )
              (i32.shl
-              (get_local $4)
+              (get_local $3)
               (i32.const 8)
              )
             )
            )
-           (set_local $5
+           (set_local $4
             (i32.load
              (i32.add
               (get_local $1)
@@ -987,21 +941,21 @@
            )
            (i32.store
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 8)
             )
             (i32.or
              (i32.shr_u
-              (get_local $4)
+              (get_local $3)
               (i32.const 24)
              )
              (i32.shl
-              (get_local $5)
+              (get_local $4)
               (i32.const 8)
              )
             )
            )
-           (set_local $4
+           (set_local $3
             (i32.load
              (i32.add
               (get_local $1)
@@ -1011,16 +965,16 @@
            )
            (i32.store
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 12)
             )
             (i32.or
              (i32.shr_u
-              (get_local $5)
+              (get_local $4)
               (i32.const 24)
              )
              (i32.shl
-              (get_local $4)
+              (get_local $3)
               (i32.const 8)
              )
             )
@@ -1031,9 +985,9 @@
              (i32.const 16)
             )
            )
-           (set_local $3
+           (set_local $0
             (i32.add
-             (get_local $3)
+             (get_local $0)
              (i32.const 16)
             )
            )
@@ -1051,64 +1005,64 @@
       )
       (br $break|2)
      )
-     (set_local $4
+     (set_local $3
       (i32.load
        (get_local $1)
       )
      )
      (i32.store8
       (block (result i32)
-       (set_local $6
-        (get_local $3)
+       (set_local $5
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $6
+        (set_local $5
          (get_local $1)
         )
         (set_local $1
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
       )
      )
      (i32.store8
       (block (result i32)
-       (set_local $6
-        (get_local $3)
+       (set_local $5
+        (get_local $0)
        )
-       (set_local $3
+       (set_local $0
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
       (i32.load8_u
        (block (result i32)
-        (set_local $6
+        (set_local $5
          (get_local $1)
         )
         (set_local $1
          (i32.add
-          (get_local $6)
+          (get_local $5)
           (i32.const 1)
          )
         )
-        (get_local $6)
+        (get_local $5)
        )
       )
      )
@@ -1127,7 +1081,7 @@
         )
         (block
          (block
-          (set_local $5
+          (set_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -1136,19 +1090,19 @@
            )
           )
           (i32.store
-           (get_local $3)
+           (get_local $0)
            (i32.or
             (i32.shr_u
-             (get_local $4)
+             (get_local $3)
              (i32.const 16)
             )
             (i32.shl
-             (get_local $5)
+             (get_local $4)
              (i32.const 16)
             )
            )
           )
-          (set_local $4
+          (set_local $3
            (i32.load
             (i32.add
              (get_local $1)
@@ -1158,21 +1112,21 @@
           )
           (i32.store
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 4)
            )
            (i32.or
             (i32.shr_u
-             (get_local $5)
+             (get_local $4)
              (i32.const 16)
             )
             (i32.shl
-             (get_local $4)
+             (get_local $3)
              (i32.const 16)
             )
            )
           )
-          (set_local $5
+          (set_local $4
            (i32.load
             (i32.add
              (get_local $1)
@@ -1182,21 +1136,21 @@
           )
           (i32.store
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 8)
            )
            (i32.or
             (i32.shr_u
-             (get_local $4)
+             (get_local $3)
              (i32.const 16)
             )
             (i32.shl
-             (get_local $5)
+             (get_local $4)
              (i32.const 16)
             )
            )
           )
-          (set_local $4
+          (set_local $3
            (i32.load
             (i32.add
              (get_local $1)
@@ -1206,16 +1160,16 @@
           )
           (i32.store
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 12)
            )
            (i32.or
             (i32.shr_u
-             (get_local $5)
+             (get_local $4)
              (i32.const 16)
             )
             (i32.shl
-             (get_local $4)
+             (get_local $3)
              (i32.const 16)
             )
            )
@@ -1226,9 +1180,9 @@
             (i32.const 16)
            )
           )
-          (set_local $3
+          (set_local $0
            (i32.add
-            (get_local $3)
+            (get_local $0)
             (i32.const 16)
            )
           )
@@ -1246,36 +1200,36 @@
      )
      (br $break|2)
     )
-    (set_local $4
+    (set_local $3
      (i32.load
       (get_local $1)
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -1294,7 +1248,7 @@
        )
        (block
         (block
-         (set_local $5
+         (set_local $4
           (i32.load
            (i32.add
             (get_local $1)
@@ -1303,19 +1257,19 @@
           )
          )
          (i32.store
-          (get_local $3)
+          (get_local $0)
           (i32.or
            (i32.shr_u
-            (get_local $4)
+            (get_local $3)
             (i32.const 8)
            )
            (i32.shl
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
           )
          )
-         (set_local $4
+         (set_local $3
           (i32.load
            (i32.add
             (get_local $1)
@@ -1325,21 +1279,21 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 4)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 8)
            )
            (i32.shl
-            (get_local $4)
+            (get_local $3)
             (i32.const 24)
            )
           )
          )
-         (set_local $5
+         (set_local $4
           (i32.load
            (i32.add
             (get_local $1)
@@ -1349,21 +1303,21 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 8)
           )
           (i32.or
            (i32.shr_u
-            (get_local $4)
+            (get_local $3)
             (i32.const 8)
            )
            (i32.shl
-            (get_local $5)
+            (get_local $4)
             (i32.const 24)
            )
           )
          )
-         (set_local $4
+         (set_local $3
           (i32.load
            (i32.add
             (get_local $1)
@@ -1373,16 +1327,16 @@
          )
          (i32.store
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 12)
           )
           (i32.or
            (i32.shr_u
-            (get_local $5)
+            (get_local $4)
             (i32.const 8)
            )
            (i32.shl
-            (get_local $4)
+            (get_local $3)
             (i32.const 24)
            )
           )
@@ -1393,9 +1347,9 @@
            (i32.const 16)
           )
          )
-         (set_local $3
+         (set_local $0
           (i32.add
-           (get_local $3)
+           (get_local $0)
            (i32.const 16)
           )
          )
@@ -1422,449 +1376,449 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -1878,225 +1832,225 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -2110,113 +2064,113 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -2230,57 +2184,57 @@
    (block
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
     (i32.store8
      (block (result i32)
-      (set_local $6
-       (get_local $3)
+      (set_local $5
+       (get_local $0)
       )
-      (set_local $3
+      (set_local $0
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
      (i32.load8_u
       (block (result i32)
-       (set_local $6
+       (set_local $5
         (get_local $1)
        )
        (set_local $1
         (i32.add
-         (get_local $6)
+         (get_local $5)
          (i32.const 1)
         )
        )
-       (get_local $6)
+       (get_local $5)
       )
      )
     )
@@ -2293,38 +2247,359 @@
    )
    (i32.store8
     (block (result i32)
-     (set_local $6
-      (get_local $3)
+     (set_local $5
+      (get_local $0)
      )
-     (set_local $3
+     (set_local $0
       (i32.add
-       (get_local $6)
+       (get_local $5)
        (i32.const 1)
       )
      )
-     (get_local $6)
+     (get_local $5)
     )
     (i32.load8_u
      (block (result i32)
-      (set_local $6
+      (set_local $5
        (get_local $1)
       )
       (set_local $1
        (i32.add
-        (get_local $6)
+        (get_local $5)
         (i32.const 1)
        )
       )
-      (get_local $6)
+      (get_local $5)
      )
     )
    )
   )
-  (return
-   (get_local $0)
+ )
+ (func $std:heap/move_memory (; 3 ;) (type $iiiv) (param $0 i32) (param $1 i32) (param $2 i32)
+  (local $3 i32)
+  (if
+   (i32.eq
+    (get_local $0)
+    (get_local $1)
+   )
+   (return)
+  )
+  (if
+   (i32.and
+    (if (result i32)
+     (i32.ne
+      (tee_local $3
+       (i32.le_u
+        (i32.add
+         (get_local $1)
+         (get_local $2)
+        )
+        (get_local $0)
+       )
+      )
+      (i32.const 0)
+     )
+     (get_local $3)
+     (i32.le_u
+      (i32.add
+       (get_local $0)
+       (get_local $2)
+      )
+      (get_local $1)
+     )
+    )
+    (i32.const 1)
+   )
+   (block
+    (call $std:heap/copy_memory
+     (get_local $0)
+     (get_local $1)
+     (get_local $2)
+    )
+    (return)
+   )
+  )
+  (if
+   (i32.lt_u
+    (get_local $0)
+    (get_local $1)
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (block $break|0
+       (loop $continue|0
+        (if
+         (i32.rem_u
+          (get_local $0)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (if
+            (i32.eqz
+             (get_local $2)
+            )
+            (return)
+           )
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 1)
+            )
+           )
+           (i32.store8
+            (block (result i32)
+             (set_local $3
+              (get_local $0)
+             )
+             (set_local $0
+              (i32.add
+               (get_local $3)
+               (i32.const 1)
+              )
+             )
+             (get_local $3)
+            )
+            (i32.load8_u
+             (block (result i32)
+              (set_local $3
+               (get_local $1)
+              )
+              (set_local $1
+               (i32.add
+                (get_local $3)
+                (i32.const 1)
+               )
+              )
+              (get_local $3)
+             )
+            )
+           )
+          )
+          (br $continue|0)
+         )
+        )
+       )
+      )
+      (block $break|1
+       (loop $continue|1
+        (if
+         (i32.ge_u
+          (get_local $2)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (i64.store
+            (get_local $0)
+            (i64.load
+             (get_local $1)
+            )
+           )
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+           (set_local $0
+            (i32.add
+             (get_local $0)
+             (i32.const 8)
+            )
+           )
+           (set_local $1
+            (i32.add
+             (get_local $1)
+             (i32.const 8)
+            )
+           )
+          )
+          (br $continue|1)
+         )
+        )
+       )
+      )
+     )
+    )
+    (block $break|2
+     (loop $continue|2
+      (if
+       (get_local $2)
+       (block
+        (block
+         (i32.store8
+          (block (result i32)
+           (set_local $3
+            (get_local $0)
+           )
+           (set_local $0
+            (i32.add
+             (get_local $3)
+             (i32.const 1)
+            )
+           )
+           (get_local $3)
+          )
+          (i32.load8_u
+           (block (result i32)
+            (set_local $3
+             (get_local $1)
+            )
+            (set_local $1
+             (i32.add
+              (get_local $3)
+              (i32.const 1)
+             )
+            )
+            (get_local $3)
+           )
+          )
+         )
+         (set_local $2
+          (i32.sub
+           (get_local $2)
+           (i32.const 1)
+          )
+         )
+        )
+        (br $continue|2)
+       )
+      )
+     )
+    )
+   )
+   (block
+    (if
+     (i32.eq
+      (i32.rem_u
+       (get_local $1)
+       (i32.const 8)
+      )
+      (i32.rem_u
+       (get_local $0)
+       (i32.const 8)
+      )
+     )
+     (block
+      (block $break|3
+       (loop $continue|3
+        (if
+         (i32.rem_u
+          (i32.add
+           (get_local $0)
+           (get_local $2)
+          )
+          (i32.const 8)
+         )
+         (block
+          (block
+           (if
+            (i32.eqz
+             (get_local $2)
+            )
+            (return)
+           )
+           (i32.store8
+            (i32.add
+             (get_local $0)
+             (tee_local $2
+              (i32.sub
+               (get_local $2)
+               (i32.const 1)
+              )
+             )
+            )
+            (i32.load8_u
+             (i32.add
+              (get_local $1)
+              (get_local $2)
+             )
+            )
+           )
+          )
+          (br $continue|3)
+         )
+        )
+       )
+      )
+      (block $break|4
+       (loop $continue|4
+        (if
+         (i32.ge_u
+          (get_local $2)
+          (i32.const 8)
+         )
+         (block
+          (block
+           (set_local $2
+            (i32.sub
+             (get_local $2)
+             (i32.const 8)
+            )
+           )
+           (i64.store
+            (i32.add
+             (get_local $0)
+             (get_local $2)
+            )
+            (i64.load
+             (i32.add
+              (get_local $1)
+              (get_local $2)
+             )
+            )
+           )
+          )
+          (br $continue|4)
+         )
+        )
+       )
+      )
+     )
+    )
+    (block $break|5
+     (loop $continue|5
+      (if
+       (get_local $2)
+       (block
+        (i32.store8
+         (i32.add
+          (get_local $0)
+          (tee_local $2
+           (i32.sub
+            (get_local $2)
+            (i32.const 1)
+           )
+          )
+         )
+         (i32.load8_u
+          (i32.add
+           (get_local $1)
+           (get_local $2)
+          )
+         )
+        )
+        (br $continue|5)
+       )
+      )
+     )
+    )
+   )
   )
  )
- (func $std:heap/Heap.compare (; 3 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
+ (func $std:heap/compare_memory (; 4 ;) (type $iiii) (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
   (if
    (i32.eq
     (get_local $0)
@@ -2393,19 +2668,19 @@
    )
   )
  )
- (func $std:heap/Heap.dispose (; 4 ;) (type $iv) (param $0 i32)
+ (func $std:heap/free_memory (; 5 ;) (type $iv) (param $0 i32)
  )
- (func $start (; 5 ;) (type $v)
+ (func $start (; 6 ;) (type $v)
   (set_global $std:heap/HEAP_OFFSET
    (get_global $HEAP_BASE)
   )
   (set_global $std/heap/ptr1
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.const 42)
    )
   )
   (set_global $std/heap/ptr2
-   (call $std:heap/Heap.allocate
+   (call $std:heap/allocate_memory
     (i32.const 42)
    )
   )
@@ -2418,12 +2693,10 @@
    )
    (unreachable)
   )
-  (drop
-   (call $std:heap/Heap.fill
-    (get_global $std/heap/ptr1)
-    (i32.const 18)
-    (i32.const 42)
-   )
+  (call $std:heap/set_memory
+   (get_global $std/heap/ptr1)
+   (i32.const 18)
+   (i32.const 42)
   )
   (block $break|0
    (set_global $std/heap/i
@@ -2461,12 +2734,10 @@
     )
    )
   )
-  (drop
-   (call $std:heap/Heap.copy
-    (get_global $std/heap/ptr2)
-    (get_global $std/heap/ptr1)
-    (i32.const 42)
-   )
+  (call $std:heap/move_memory
+   (get_global $std/heap/ptr2)
+   (get_global $std/heap/ptr1)
+   (i32.const 42)
   )
   (block $break|1
    (set_global $std/heap/i
@@ -2507,7 +2778,7 @@
   (if
    (i32.eqz
     (i32.eq
-     (call $std:heap/Heap.compare
+     (call $std:heap/compare_memory
       (get_global $std/heap/ptr1)
       (get_global $std/heap/ptr2)
       (i32.const 42)
@@ -2517,10 +2788,10 @@
    )
    (unreachable)
   )
-  (call $std:heap/Heap.dispose
+  (call $std:heap/free_memory
    (get_global $std/heap/ptr1)
   )
-  (call $std:heap/Heap.dispose
+  (call $std:heap/free_memory
    (get_global $std/heap/ptr2)
   )
  )
@@ -2581,16 +2852,17 @@
   GLOBAL: std:heap/ALIGN_SIZE
   GLOBAL: std:heap/ALIGN_MASK
   GLOBAL: std:heap/HEAP_OFFSET
-  CLASS_PROTOTYPE: std:heap/Heap
-  CLASS_PROTOTYPE: Heap
-  PROPERTY: std:heap/Heap.used
-  PROPERTY: std:heap/Heap.free
-  PROPERTY: std:heap/Heap.size
-  FUNCTION_PROTOTYPE: std:heap/Heap.allocate
-  FUNCTION_PROTOTYPE: std:heap/Heap.dispose
-  FUNCTION_PROTOTYPE: std:heap/Heap.copy
-  FUNCTION_PROTOTYPE: std:heap/Heap.fill
-  FUNCTION_PROTOTYPE: std:heap/Heap.compare
+  FUNCTION_PROTOTYPE: std:heap/allocate_memory
+  FUNCTION_PROTOTYPE: allocate_memory
+  FUNCTION_PROTOTYPE: std:heap/free_memory
+  FUNCTION_PROTOTYPE: free_memory
+  FUNCTION_PROTOTYPE: std:heap/copy_memory
+  FUNCTION_PROTOTYPE: std:heap/move_memory
+  FUNCTION_PROTOTYPE: move_memory
+  FUNCTION_PROTOTYPE: std:heap/set_memory
+  FUNCTION_PROTOTYPE: set_memory
+  FUNCTION_PROTOTYPE: std:heap/compare_memory
+  FUNCTION_PROTOTYPE: compare_memory
   CLASS_PROTOTYPE: std:map/Map
   CLASS_PROTOTYPE: Map
   CLASS_PROTOTYPE: std:regexp/RegExp
@@ -2614,7 +2886,11 @@
   CLASS_PROTOTYPE: std:array/CArray
   CLASS_PROTOTYPE: std:error/Error
   CLASS_PROTOTYPE: std:error/RangeError
-  CLASS_PROTOTYPE: std:heap/Heap
+  FUNCTION_PROTOTYPE: std:heap/allocate_memory
+  FUNCTION_PROTOTYPE: std:heap/free_memory
+  FUNCTION_PROTOTYPE: std:heap/move_memory
+  FUNCTION_PROTOTYPE: std:heap/set_memory
+  FUNCTION_PROTOTYPE: std:heap/compare_memory
   CLASS_PROTOTYPE: std:map/Map
   CLASS_PROTOTYPE: std:regexp/RegExp
   CLASS_PROTOTYPE: std:set/Set