From 00766b42fb6627f79a15c72ff2bae7138c1e90d6 Mon Sep 17 00:00:00 2001
From: dcode <dcode@dcode.io>
Date: Wed, 28 Aug 2019 23:01:03 +0200
Subject: [PATCH] Improve stub RT reallocation

---
 std/assembly/rt/stub.ts                       |  38 +-
 tests/compiler/call-super.optimized.wat       |  34 +-
 tests/compiler/call-super.untouched.wat       |  31 +-
 tests/compiler/constructor.optimized.wat      |  34 +-
 tests/compiler/constructor.untouched.wat      |  31 +-
 tests/compiler/exports.optimized.wat          |  23 +-
 tests/compiler/exports.untouched.wat          |  31 +-
 tests/compiler/getter-call.optimized.wat      |  23 +-
 tests/compiler/getter-call.untouched.wat      |  31 +-
 tests/compiler/inlining.optimized.wat         |  34 +-
 tests/compiler/inlining.untouched.wat         |  31 +-
 tests/compiler/number.optimized.wat           | 135 ++-
 tests/compiler/number.untouched.wat           | 117 ++-
 .../optional-typeparameters.optimized.wat     |  23 +-
 .../optional-typeparameters.untouched.wat     |  31 +-
 tests/compiler/resolve-access.optimized.wat   |  34 +-
 tests/compiler/resolve-access.untouched.wat   |  31 +-
 tests/compiler/retain-release.optimized.wat   |  34 +-
 tests/compiler/retain-release.untouched.wat   |  31 +-
 tests/compiler/rt/instanceof.optimized.wat    |  23 +-
 tests/compiler/rt/instanceof.untouched.wat    |  31 +-
 tests/compiler/rt/stub-realloc.optimized.wat  | 180 +++-
 tests/compiler/rt/stub-realloc.ts             |   3 +
 tests/compiler/rt/stub-realloc.untouched.wat  | 188 +++-
 tests/compiler/runtime-stub.optimized.wat     |  34 +-
 tests/compiler/runtime-stub.untouched.wat     |  31 +-
 tests/compiler/simd.untouched.wat             | 850 +-----------------
 tests/compiler/std/date.optimized.wat         |  23 +-
 tests/compiler/std/date.untouched.wat         |  31 +-
 tests/compiler/std/new.optimized.wat          |  23 +-
 tests/compiler/std/new.untouched.wat          |  31 +-
 .../compiler/std/object-literal.optimized.wat |  34 +-
 .../compiler/std/object-literal.untouched.wat |  31 +-
 .../std/operator-overloading.optimized.wat    |  23 +-
 .../std/operator-overloading.untouched.wat    |  31 +-
 tests/compiler/std/static-array.optimized.wat | 117 ++-
 tests/compiler/std/static-array.untouched.wat | 116 ++-
 tests/compiler/std/symbol.optimized.wat       |  34 +-
 tests/compiler/std/symbol.untouched.wat       |  31 +-
 39 files changed, 1215 insertions(+), 1427 deletions(-)

diff --git a/std/assembly/rt/stub.ts b/std/assembly/rt/stub.ts
index c715c962bf..6efff7088d 100644
--- a/std/assembly/rt/stub.ts
+++ b/std/assembly/rt/stub.ts
@@ -1,4 +1,4 @@
-import { AL_MASK, BLOCK, BLOCK_OVERHEAD, BLOCK_MAXSIZE } from "rt/common";
+import { AL_MASK, BLOCK, BLOCK_OVERHEAD, BLOCK_MAXSIZE, AL_SIZE, DEBUG } from "rt/common";
 
 // @ts-ignore: decorator
 @lazy
@@ -9,7 +9,7 @@ var startOffset: usize = (__heap_base + AL_MASK) & ~AL_MASK;
 var offset: usize = startOffset;
 
 function maybeGrowMemory(newOffset: usize): void {
-  newOffset = (newOffset + AL_MASK) & ~AL_MASK;
+  // assumes newOffset is aligned
   var pagesBefore = memory.size();
   var maxOffset = <usize>pagesBefore << 16;
   if (newOffset > maxOffset) {
@@ -27,8 +27,11 @@ function maybeGrowMemory(newOffset: usize): void {
 export function __alloc(size: usize, id: u32): usize {
   if (size > BLOCK_MAXSIZE) unreachable();
   var ptr = offset + BLOCK_OVERHEAD;
-  maybeGrowMemory(ptr + max<usize>(size, 1));
+  var actualSize = max<usize>((size + AL_MASK) & ~AL_MASK, AL_SIZE);
+  maybeGrowMemory(ptr + actualSize);
   var block = changetype<BLOCK>(ptr - BLOCK_OVERHEAD);
+  block.mmInfo = actualSize;
+  if (DEBUG) block.gcInfo = -1;
   block.rtId = id;
   block.rtSize = size;
   return ptr;
@@ -39,19 +42,24 @@ export function __alloc(size: usize, id: u32): usize {
 export function __realloc(ptr: usize, size: usize): usize {
   assert(ptr != 0 && !(ptr & AL_MASK)); // must exist and be aligned
   var block = changetype<BLOCK>(ptr - BLOCK_OVERHEAD);
-  var actualSize = (<usize>block.rtSize + AL_MASK) & ~AL_MASK;
-  var isLast = ptr + actualSize == offset;
+  var actualSize = block.mmInfo;
+  if (DEBUG) assert(block.gcInfo == -1);
   if (size > actualSize) {
-    if (isLast) { // maybe grow
+    if (ptr + actualSize == offset) { // last block: grow
       if (size > BLOCK_MAXSIZE) unreachable();
-      maybeGrowMemory(ptr + size);
+      actualSize = (size + AL_MASK) & ~AL_MASK;
+      maybeGrowMemory(ptr + actualSize);
+      block.mmInfo = actualSize;
     } else { // copy to new block at least double the size
-      let newPtr = __alloc(max<usize>(size, actualSize << 1), block.rtId);
-      memory.copy(newPtr, ptr, actualSize);
+      actualSize = max<usize>((size + AL_MASK) & ~AL_MASK, actualSize << 1);
+      let newPtr = __alloc(actualSize, block.rtId);
+      memory.copy(newPtr, ptr, block.rtSize);
       block = changetype<BLOCK>((ptr = newPtr) - BLOCK_OVERHEAD);
     }
-  } else if (isLast) { // shrink
-    offset = (ptr + size + AL_MASK) & ~AL_MASK;
+  } else if (ptr + actualSize == offset) { // last block: shrink
+    actualSize = (size + AL_MASK) & ~AL_MASK;
+    offset = ptr + actualSize;
+    block.mmInfo = actualSize;
   }
   block.rtSize = size;
   return ptr;
@@ -59,7 +67,13 @@ export function __realloc(ptr: usize, size: usize): usize {
 
 // @ts-ignore: decorator
 @unsafe @global
-export function __free(ref: usize): void {
+export function __free(ptr: usize): void {
+  assert(ptr != 0 && !(ptr & AL_MASK)); // must exist and be aligned
+  var block = changetype<BLOCK>(ptr - BLOCK_OVERHEAD);
+  if (DEBUG) assert(block.gcInfo == -1);
+  if (ptr + block.mmInfo == offset) { // last block: discard
+    offset = changetype<usize>(block);
+  }
 }
 
 // @ts-ignore: decorator
diff --git a/tests/compiler/call-super.optimized.wat b/tests/compiler/call-super.optimized.wat
index afe7f7d5cc..4d182dd23b 100644
--- a/tests/compiler/call-super.optimized.wat
+++ b/tests/compiler/call-super.optimized.wat
@@ -16,11 +16,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -62,6 +57,7 @@
  (func $~lib/rt/stub/__alloc (; 2 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -71,25 +67,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $call-super/A#constructor (; 3 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
   local.get $0
diff --git a/tests/compiler/call-super.untouched.wat b/tests/compiler/call-super.untouched.wat
index 3576b7b267..bb24c05c58 100644
--- a/tests/compiler/call-super.untouched.wat
+++ b/tests/compiler/call-super.untouched.wat
@@ -20,14 +20,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -81,6 +73,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -91,25 +84,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/constructor.optimized.wat b/tests/compiler/constructor.optimized.wat
index 04c91348ea..fbd2b98ef4 100644
--- a/tests/compiler/constructor.optimized.wat
+++ b/tests/compiler/constructor.optimized.wat
@@ -22,11 +22,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -68,6 +63,7 @@
  (func $~lib/rt/stub/__alloc (; 1 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -77,25 +73,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $start:constructor (; 2 ;) (type $FUNCSIG$v)
   (local $0 i32)
diff --git a/tests/compiler/constructor.untouched.wat b/tests/compiler/constructor.untouched.wat
index 61b3283e95..55e22eb4a9 100644
--- a/tests/compiler/constructor.untouched.wat
+++ b/tests/compiler/constructor.untouched.wat
@@ -28,14 +28,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -89,6 +81,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -99,25 +92,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/exports.optimized.wat b/tests/compiler/exports.optimized.wat
index 7de1f70d26..07e64a70cb 100644
--- a/tests/compiler/exports.optimized.wat
+++ b/tests/compiler/exports.optimized.wat
@@ -64,11 +64,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -113,20 +108,26 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $1
-  i32.const 4
+  local.tee $2
+  i32.const 16
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $1
+  local.get $2
   i32.const 16
   i32.sub
-  local.tee $2
+  local.tee $1
+  i32.const 16
+  i32.store
+  local.get $1
+  i32.const -1
+  i32.store offset=4
+  local.get $1
   local.get $0
   i32.store offset=8
-  local.get $2
+  local.get $1
   i32.const 4
   i32.store offset=12
-  local.get $1
+  local.get $2
  )
  (func $exports/Car#get:numDoors (; 5 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
   local.get $0
diff --git a/tests/compiler/exports.untouched.wat b/tests/compiler/exports.untouched.wat
index 83d21f38ff..5ad9605083 100644
--- a/tests/compiler/exports.untouched.wat
+++ b/tests/compiler/exports.untouched.wat
@@ -74,14 +74,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -135,6 +127,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -145,25 +138,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/getter-call.optimized.wat b/tests/compiler/getter-call.optimized.wat
index 8af8457747..5edabea077 100644
--- a/tests/compiler/getter-call.optimized.wat
+++ b/tests/compiler/getter-call.optimized.wat
@@ -15,11 +15,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -64,20 +59,26 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $0
-  i32.const 1
+  local.tee $1
+  i32.const 16
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $0
+  local.get $1
   i32.const 16
   i32.sub
-  local.tee $1
+  local.tee $0
+  i32.const 16
+  i32.store
+  local.get $0
+  i32.const -1
+  i32.store offset=4
+  local.get $0
   i32.const 3
   i32.store offset=8
-  local.get $1
+  local.get $0
   i32.const 0
   i32.store offset=12
-  local.get $0
+  local.get $1
  )
  (func $getter-call/C#get:x~anonymous|0 (; 2 ;) (type $FUNCSIG$i) (result i32)
   i32.const 42
diff --git a/tests/compiler/getter-call.untouched.wat b/tests/compiler/getter-call.untouched.wat
index 6e23747da5..9bd7e0e965 100644
--- a/tests/compiler/getter-call.untouched.wat
+++ b/tests/compiler/getter-call.untouched.wat
@@ -20,14 +20,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -81,6 +73,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -91,25 +84,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/inlining.optimized.wat b/tests/compiler/inlining.optimized.wat
index 738b0b807a..df12a651e7 100644
--- a/tests/compiler/inlining.optimized.wat
+++ b/tests/compiler/inlining.optimized.wat
@@ -42,11 +42,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -88,6 +83,7 @@
  (func $~lib/rt/stub/__alloc (; 5 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -97,25 +93,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $inlining/test_ctor (; 6 ;) (type $FUNCSIG$v)
   (local $0 i32)
diff --git a/tests/compiler/inlining.untouched.wat b/tests/compiler/inlining.untouched.wat
index 9ea72bd92f..8edcb26a51 100644
--- a/tests/compiler/inlining.untouched.wat
+++ b/tests/compiler/inlining.untouched.wat
@@ -281,14 +281,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -342,6 +334,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -352,25 +345,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/number.optimized.wat b/tests/compiler/number.optimized.wat
index ae8b8dde48..fa626de37d 100644
--- a/tests/compiler/number.optimized.wat
+++ b/tests/compiler/number.optimized.wat
@@ -26,13 +26,14 @@
  (data (i32.const 1184) "(\00\00\00\01\00\00\00\00\00\00\00(\00\00\00\01\00\00\00\n\00\00\00d\00\00\00\e8\03\00\00\10\'\00\00\a0\86\01\00@B\0f\00\80\96\98\00\00\e1\f5\05\00\ca\9a;")
  (data (i32.const 1240) "\10\00\00\00\01\00\00\00\05\00\00\00\10\00\00\00\b0\04\00\00\b0\04\00\00(\00\00\00\n")
  (data (i32.const 1276) "\01\00\00\00\01")
- (data (i32.const 1288) "\06\00\00\00\01\00\00\00\01\00\00\00\06\00\00\002\00.\000")
- (data (i32.const 1312) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\003")
- (data (i32.const 1336) "\04\00\00\00\01\00\00\00\01\00\00\00\04\00\00\00-\005")
- (data (i32.const 1360) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\004")
- (data (i32.const 1384) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\002")
- (data (i32.const 1408) "\08\00\00\00\01\00\00\00\01\00\00\00\08\00\00\00t\00r\00u\00e")
- (data (i32.const 1432) "\n\00\00\00\01\00\00\00\01\00\00\00\n\00\00\00f\00a\00l\00s\00e")
+ (data (i32.const 1288) "\1e\00\00\00\01\00\00\00\01\00\00\00\1e\00\00\00~\00l\00i\00b\00/\00r\00t\00/\00s\00t\00u\00b\00.\00t\00s")
+ (data (i32.const 1336) "\06\00\00\00\01\00\00\00\01\00\00\00\06\00\00\002\00.\000")
+ (data (i32.const 1360) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\003")
+ (data (i32.const 1384) "\04\00\00\00\01\00\00\00\01\00\00\00\04\00\00\00-\005")
+ (data (i32.const 1408) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\004")
+ (data (i32.const 1432) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\002")
+ (data (i32.const 1456) "\08\00\00\00\01\00\00\00\01\00\00\00\08\00\00\00t\00r\00u\00e")
+ (data (i32.const 1480) "\n\00\00\00\01\00\00\00\01\00\00\00\n\00\00\00f\00a\00l\00s\00e")
  (global $number/a (mut i32) (i32.const 1))
  (global $~lib/rt/stub/startOffset (mut i32) (i32.const 0))
  (global $~lib/rt/stub/offset (mut i32) (i32.const 0))
@@ -96,11 +97,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -142,6 +138,7 @@
  (func $~lib/rt/stub/__alloc (; 3 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
   (local $1 i32)
   (local $2 i32)
+  (local $3 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -151,25 +148,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $1
-  local.get $0
-  i32.const 1
+  local.tee $2
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $1
+  i32.const 16
+  local.get $1
+  i32.const 16
   i32.gt_u
   select
+  local.tee $3
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $1
+  local.get $2
   i32.const 16
   i32.sub
-  local.tee $2
+  local.tee $1
+  local.get $3
+  i32.store
+  local.get $1
+  i32.const -1
+  i32.store offset=4
+  local.get $1
   i32.const 1
   i32.store offset=8
-  local.get $2
+  local.get $1
   local.get $0
   i32.store offset=12
-  local.get $1
+  local.get $2
  )
  (func $~lib/util/number/utoa_simple<u32> (; 4 ;) (type $FUNCSIG$viii) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
@@ -1395,7 +1404,51 @@
   call $~lib/memory/memory.copy
   local.get $1
  )
- (func $~lib/util/number/dtoa (; 16 ;) (type $FUNCSIG$i) (result i32)
+ (func $~lib/rt/stub/__free (; 16 ;) (type $FUNCSIG$vi) (param $0 i32)
+  (local $1 i32)
+  local.get $0
+  i32.const 15
+  i32.and
+  i32.eqz
+  i32.const 0
+  local.get $0
+  select
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 1304
+   i32.const 71
+   i32.const 2
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  i32.const 16
+  i32.sub
+  local.tee $1
+  i32.load offset=4
+  i32.const -1
+  i32.ne
+  if
+   i32.const 0
+   i32.const 1304
+   i32.const 73
+   i32.const 13
+   call $~lib/builtins/abort
+   unreachable
+  end
+  global.get $~lib/rt/stub/offset
+  local.get $1
+  i32.load
+  local.get $0
+  i32.add
+  i32.eq
+  if
+   local.get $1
+   global.set $~lib/rt/stub/offset
+  end
+ )
+ (func $~lib/util/number/dtoa (; 17 ;) (type $FUNCSIG$i) (result i32)
   (local $0 i32)
   (local $1 i32)
   f64.const 2
@@ -1425,14 +1478,16 @@
   local.get $0
   local.get $1
   call $~lib/string/String#substring
+  local.get $0
+  call $~lib/rt/stub/__free
  )
- (func $~lib/number/Bool#toString (; 17 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
-  i32.const 1424
-  i32.const 1448
+ (func $~lib/number/Bool#toString (; 18 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
+  i32.const 1472
+  i32.const 1496
   local.get $0
   select
  )
- (func $~lib/number/F32.isSafeInteger (; 18 ;) (type $FUNCSIG$if) (param $0 f32) (result i32)
+ (func $~lib/number/F32.isSafeInteger (; 19 ;) (type $FUNCSIG$if) (param $0 f32) (result i32)
   local.get $0
   f32.trunc
   local.get $0
@@ -1444,7 +1499,7 @@
   f32.le
   select
  )
- (func $~lib/number/F32.isInteger (; 19 ;) (type $FUNCSIG$if) (param $0 f32) (result i32)
+ (func $~lib/number/F32.isInteger (; 20 ;) (type $FUNCSIG$if) (param $0 f32) (result i32)
   local.get $0
   f32.trunc
   local.get $0
@@ -1457,7 +1512,7 @@
   f32.eq
   select
  )
- (func $~lib/number/F64.isSafeInteger (; 20 ;) (type $FUNCSIG$id) (param $0 f64) (result i32)
+ (func $~lib/number/F64.isSafeInteger (; 21 ;) (type $FUNCSIG$id) (param $0 f64) (result i32)
   local.get $0
   f64.trunc
   local.get $0
@@ -1469,7 +1524,7 @@
   f64.le
   select
  )
- (func $~lib/number/F64.isInteger (; 21 ;) (type $FUNCSIG$id) (param $0 f64) (result i32)
+ (func $~lib/number/F64.isInteger (; 22 ;) (type $FUNCSIG$id) (param $0 f64) (result i32)
   local.get $0
   call $~lib/number/isFinite<f64>
   if (result i32)
@@ -1481,9 +1536,9 @@
    i32.const 0
   end
  )
- (func $start:number (; 22 ;) (type $FUNCSIG$v)
+ (func $start:number (; 23 ;) (type $FUNCSIG$v)
   (local $0 i32)
-  i32.const 1472
+  i32.const 1520
   global.set $~lib/rt/stub/startOffset
   global.get $~lib/rt/stub/startOffset
   global.set $~lib/rt/stub/offset
@@ -1501,7 +1556,7 @@
    unreachable
   end
   call $~lib/util/number/dtoa
-  i32.const 1304
+  i32.const 1352
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -1514,7 +1569,7 @@
   end
   i32.const 3
   call $~lib/util/number/itoa32
-  i32.const 1328
+  i32.const 1376
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -1527,7 +1582,7 @@
   end
   i32.const -5
   call $~lib/util/number/itoa32
-  i32.const 1352
+  i32.const 1400
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -1540,7 +1595,7 @@
   end
   i32.const 4
   call $~lib/util/number/itoa32
-  i32.const 1376
+  i32.const 1424
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -1557,7 +1612,7 @@
   global.set $number/a
   global.get $number/a
   call $~lib/util/number/itoa32
-  i32.const 1400
+  i32.const 1448
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -1587,7 +1642,7 @@
   end
   i32.const 1
   call $~lib/number/Bool#toString
-  i32.const 1424
+  i32.const 1472
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -1600,7 +1655,7 @@
   end
   i32.const 0
   call $~lib/number/Bool#toString
-  i32.const 1448
+  i32.const 1496
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -1636,7 +1691,7 @@
   global.set $number/a
   local.get $0
   call $~lib/util/number/itoa32
-  i32.const 1400
+  i32.const 1448
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -2099,10 +2154,10 @@
    unreachable
   end
  )
- (func $start (; 23 ;) (type $FUNCSIG$v)
+ (func $start (; 24 ;) (type $FUNCSIG$v)
   call $start:number
  )
- (func $null (; 24 ;) (type $FUNCSIG$v)
+ (func $null (; 25 ;) (type $FUNCSIG$v)
   nop
  )
 )
diff --git a/tests/compiler/number.untouched.wat b/tests/compiler/number.untouched.wat
index 3ae85d8c98..890afe11d7 100644
--- a/tests/compiler/number.untouched.wat
+++ b/tests/compiler/number.untouched.wat
@@ -31,13 +31,14 @@
  (data (i32.const 1632) "(\00\00\00\01\00\00\00\00\00\00\00(\00\00\00\01\00\00\00\n\00\00\00d\00\00\00\e8\03\00\00\10\'\00\00\a0\86\01\00@B\0f\00\80\96\98\00\00\e1\f5\05\00\ca\9a;")
  (data (i32.const 1688) "\10\00\00\00\01\00\00\00\03\00\00\00\10\00\00\00p\06\00\00p\06\00\00(\00\00\00\n\00\00\00")
  (data (i32.const 1720) "\00\00\00\00\01\00\00\00\01\00\00\00\00\00\00\00")
- (data (i32.const 1736) "\06\00\00\00\01\00\00\00\01\00\00\00\06\00\00\002\00.\000\00")
- (data (i32.const 1760) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\003\00")
- (data (i32.const 1784) "\04\00\00\00\01\00\00\00\01\00\00\00\04\00\00\00-\005\00")
- (data (i32.const 1808) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\004\00")
- (data (i32.const 1832) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\002\00")
- (data (i32.const 1856) "\08\00\00\00\01\00\00\00\01\00\00\00\08\00\00\00t\00r\00u\00e\00")
- (data (i32.const 1880) "\n\00\00\00\01\00\00\00\01\00\00\00\n\00\00\00f\00a\00l\00s\00e\00")
+ (data (i32.const 1736) "\1e\00\00\00\01\00\00\00\01\00\00\00\1e\00\00\00~\00l\00i\00b\00/\00r\00t\00/\00s\00t\00u\00b\00.\00t\00s\00")
+ (data (i32.const 1784) "\06\00\00\00\01\00\00\00\01\00\00\00\06\00\00\002\00.\000\00")
+ (data (i32.const 1808) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\003\00")
+ (data (i32.const 1832) "\04\00\00\00\01\00\00\00\01\00\00\00\04\00\00\00-\005\00")
+ (data (i32.const 1856) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\004\00")
+ (data (i32.const 1880) "\02\00\00\00\01\00\00\00\01\00\00\00\02\00\00\002\00")
+ (data (i32.const 1904) "\08\00\00\00\01\00\00\00\01\00\00\00\08\00\00\00t\00r\00u\00e\00")
+ (data (i32.const 1928) "\n\00\00\00\01\00\00\00\01\00\00\00\n\00\00\00f\00a\00l\00s\00e\00")
  (table $0 1 funcref)
  (elem (i32.const 0) $null)
  (global $number/a (mut i32) (i32.const 1))
@@ -58,7 +59,7 @@
  (global $~lib/builtins/f64.MIN_SAFE_INTEGER f64 (f64.const -9007199254740991))
  (global $~lib/builtins/f64.MAX_SAFE_INTEGER f64 (f64.const 9007199254740991))
  (global $~lib/builtins/f64.EPSILON f64 (f64.const 2.220446049250313e-16))
- (global $~lib/heap/__heap_base i32 (i32.const 1908))
+ (global $~lib/heap/__heap_base i32 (i32.const 1956))
  (export "memory" (memory $0))
  (start $start)
  (func $~lib/rt/stub/__retain (; 1 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
@@ -136,14 +137,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -197,6 +190,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -207,25 +201,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
@@ -3227,7 +3235,54 @@
   call $~lib/rt/stub/__retain
  )
  (func $~lib/rt/stub/__free (; 23 ;) (type $FUNCSIG$vi) (param $0 i32)
-  nop
+  (local $1 i32)
+  local.get $0
+  i32.const 0
+  i32.ne
+  if (result i32)
+   local.get $0
+   i32.const 15
+   i32.and
+   i32.eqz
+  else   
+   i32.const 0
+  end
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 1752
+   i32.const 71
+   i32.const 2
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  i32.const 16
+  i32.sub
+  local.set $1
+  local.get $1
+  i32.load offset=4
+  i32.const -1
+  i32.eq
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 1752
+   i32.const 73
+   i32.const 13
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  local.get $1
+  i32.load
+  i32.add
+  global.get $~lib/rt/stub/offset
+  i32.eq
+  if
+   local.get $1
+   global.set $~lib/rt/stub/offset
+  end
  )
  (func $~lib/util/number/dtoa (; 24 ;) (type $FUNCSIG$id) (param $0 f64) (result i32)
   (local $1 i32)
@@ -3297,11 +3352,11 @@
   (local $2 i32)
   local.get $0
   if (result i32)
-   i32.const 1872
+   i32.const 1920
    call $~lib/rt/stub/__retain
    local.tee $1
   else   
-   i32.const 1896
+   i32.const 1944
    call $~lib/rt/stub/__retain
    local.tee $2
   end
@@ -3411,7 +3466,7 @@
   i32.const 0
   call $~lib/number/F64#toString
   local.tee $1
-  i32.const 1752
+  i32.const 1800
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -3425,7 +3480,7 @@
   i32.const 3
   call $~lib/number/I32#toString
   local.tee $2
-  i32.const 1776
+  i32.const 1824
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -3439,7 +3494,7 @@
   i32.const -5
   call $~lib/number/I32#toString
   local.tee $3
-  i32.const 1800
+  i32.const 1848
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -3453,7 +3508,7 @@
   i32.const 4
   call $~lib/number/I32#toString
   local.tee $4
-  i32.const 1824
+  i32.const 1872
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -3471,7 +3526,7 @@
   global.get $number/a
   call $~lib/number/I32#toString
   local.tee $5
-  i32.const 1848
+  i32.const 1896
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -3503,7 +3558,7 @@
   i32.const 1
   call $~lib/number/Bool#toString
   local.tee $7
-  i32.const 1872
+  i32.const 1920
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -3517,7 +3572,7 @@
   i32.const 0
   call $~lib/number/Bool#toString
   local.tee $8
-  i32.const 1896
+  i32.const 1944
   call $~lib/string/String.__eq
   i32.eqz
   if
@@ -3555,7 +3610,7 @@
   local.get $10
   call $~lib/number/I32#toString
   local.tee $10
-  i32.const 1848
+  i32.const 1896
   call $~lib/string/String.__eq
   i32.eqz
   if
diff --git a/tests/compiler/optional-typeparameters.optimized.wat b/tests/compiler/optional-typeparameters.optimized.wat
index eff80db176..9169976248 100644
--- a/tests/compiler/optional-typeparameters.optimized.wat
+++ b/tests/compiler/optional-typeparameters.optimized.wat
@@ -13,11 +13,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -62,20 +57,26 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $1
-  i32.const 1
+  local.tee $2
+  i32.const 16
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $1
+  local.get $2
   i32.const 16
   i32.sub
-  local.tee $2
+  local.tee $1
+  i32.const 16
+  i32.store
+  local.get $1
+  i32.const -1
+  i32.store offset=4
+  local.get $1
   local.get $0
   i32.store offset=8
-  local.get $2
+  local.get $1
   i32.const 0
   i32.store offset=12
-  local.get $1
+  local.get $2
  )
  (func $start (; 2 ;) (type $FUNCSIG$v)
   i32.const 16
diff --git a/tests/compiler/optional-typeparameters.untouched.wat b/tests/compiler/optional-typeparameters.untouched.wat
index 73204cd63c..0eea36ce30 100644
--- a/tests/compiler/optional-typeparameters.untouched.wat
+++ b/tests/compiler/optional-typeparameters.untouched.wat
@@ -27,14 +27,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -88,6 +80,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -98,25 +91,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/resolve-access.optimized.wat b/tests/compiler/resolve-access.optimized.wat
index 67f6201554..57be3c3d7d 100644
--- a/tests/compiler/resolve-access.optimized.wat
+++ b/tests/compiler/resolve-access.optimized.wat
@@ -26,11 +26,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -72,6 +67,7 @@
  (func $~lib/rt/stub/__alloc (; 2 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -81,25 +77,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $~lib/memory/memory.copy (; 3 ;) (type $FUNCSIG$vi) (param $0 i32)
   (local $1 i32)
diff --git a/tests/compiler/resolve-access.untouched.wat b/tests/compiler/resolve-access.untouched.wat
index 1b40e0a894..ed9346b5dc 100644
--- a/tests/compiler/resolve-access.untouched.wat
+++ b/tests/compiler/resolve-access.untouched.wat
@@ -36,14 +36,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -97,6 +89,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -107,25 +100,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/retain-release.optimized.wat b/tests/compiler/retain-release.optimized.wat
index aa511b4705..2fafbf49b2 100644
--- a/tests/compiler/retain-release.optimized.wat
+++ b/tests/compiler/retain-release.optimized.wat
@@ -58,11 +58,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -104,6 +99,7 @@
  (func $~lib/rt/stub/__alloc (; 2 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -113,25 +109,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $retain-release/Ref#constructor (; 3 ;) (type $FUNCSIG$i) (result i32)
   i32.const 0
diff --git a/tests/compiler/retain-release.untouched.wat b/tests/compiler/retain-release.untouched.wat
index 2c7c9d0f32..30a6b5c8b3 100644
--- a/tests/compiler/retain-release.untouched.wat
+++ b/tests/compiler/retain-release.untouched.wat
@@ -61,14 +61,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -122,6 +114,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -132,25 +125,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/rt/instanceof.optimized.wat b/tests/compiler/rt/instanceof.optimized.wat
index bd0aad4a00..34e512c932 100644
--- a/tests/compiler/rt/instanceof.optimized.wat
+++ b/tests/compiler/rt/instanceof.optimized.wat
@@ -27,11 +27,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -76,20 +71,26 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $1
-  i32.const 1
+  local.tee $2
+  i32.const 16
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $1
+  local.get $2
   i32.const 16
   i32.sub
-  local.tee $2
+  local.tee $1
+  i32.const 16
+  i32.store
+  local.get $1
+  i32.const -1
+  i32.store offset=4
+  local.get $1
   local.get $0
   i32.store offset=8
-  local.get $2
+  local.get $1
   i32.const 0
   i32.store offset=12
-  local.get $1
+  local.get $2
  )
  (func $rt/instanceof/Animal#constructor (; 3 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
   local.get $0
diff --git a/tests/compiler/rt/instanceof.untouched.wat b/tests/compiler/rt/instanceof.untouched.wat
index ad3070de14..f7dc175185 100644
--- a/tests/compiler/rt/instanceof.untouched.wat
+++ b/tests/compiler/rt/instanceof.untouched.wat
@@ -32,14 +32,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -93,6 +85,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -103,25 +96,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/rt/stub-realloc.optimized.wat b/tests/compiler/rt/stub-realloc.optimized.wat
index f3a34c8e68..7052613dbd 100644
--- a/tests/compiler/rt/stub-realloc.optimized.wat
+++ b/tests/compiler/rt/stub-realloc.optimized.wat
@@ -28,11 +28,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -74,6 +69,7 @@
  (func $~lib/rt/stub/__alloc (; 2 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -83,25 +79,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $~lib/rt/stub/__retain (; 3 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
   local.get $0
@@ -302,31 +310,38 @@
   if
    i32.const 0
    i32.const 24
-   i32.const 40
+   i32.const 43
    i32.const 2
    call $~lib/builtins/abort
    unreachable
   end
-  global.get $~lib/rt/stub/offset
   local.get $0
   i32.const 16
   i32.sub
-  local.tee $2
-  i32.load offset=12
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $4
-  local.get $0
-  i32.add
-  i32.eq
-  local.set $3
+  local.tee $3
+  i32.load
+  local.set $2
+  local.get $3
+  i32.load offset=4
+  i32.const -1
+  i32.ne
+  if
+   i32.const 0
+   i32.const 24
+   i32.const 46
+   i32.const 13
+   call $~lib/builtins/abort
+   unreachable
+  end
   local.get $1
-  local.get $4
+  local.get $2
   i32.gt_u
   if
-   local.get $3
+   global.get $~lib/rt/stub/offset
+   local.get $0
+   local.get $2
+   i32.add
+   i32.eq
    if
     local.get $1
     i32.const 1073741808
@@ -334,52 +349,118 @@
     if
      unreachable
     end
-    local.get $0
     local.get $1
+    i32.const 15
+    i32.add
+    i32.const -16
+    i32.and
+    local.tee $2
+    local.get $0
     i32.add
     call $~lib/rt/stub/maybeGrowMemory
+    local.get $3
+    local.get $2
+    i32.store
    else    
     local.get $1
-    local.get $4
+    i32.const 15
+    i32.add
+    i32.const -16
+    i32.and
+    local.tee $4
+    local.get $2
     i32.const 1
     i32.shl
-    local.tee $3
-    local.get $1
-    local.get $3
+    local.tee $2
+    local.get $4
+    local.get $2
     i32.gt_u
     select
-    local.get $2
+    local.get $3
     i32.load offset=8
     call $~lib/rt/stub/__alloc
     local.tee $2
     local.get $0
-    local.get $4
+    local.get $3
+    i32.load offset=12
     call $~lib/memory/memory.copy
     local.get $2
     local.tee $0
     i32.const 16
     i32.sub
-    local.set $2
+    local.set $3
    end
   else   
-   local.get $3
+   global.get $~lib/rt/stub/offset
+   local.get $0
+   local.get $2
+   i32.add
+   i32.eq
    if
-    local.get $0
     local.get $1
-    i32.add
     i32.const 15
     i32.add
     i32.const -16
     i32.and
+    local.tee $2
+    local.get $0
+    i32.add
     global.set $~lib/rt/stub/offset
+    local.get $3
+    local.get $2
+    i32.store
    end
   end
-  local.get $2
+  local.get $3
   local.get $1
   i32.store offset=12
   local.get $0
  )
- (func $start:rt/stub-realloc (; 8 ;) (type $FUNCSIG$v)
+ (func $~lib/rt/stub/__free (; 8 ;) (type $FUNCSIG$vi) (param $0 i32)
+  (local $1 i32)
+  local.get $0
+  i32.const 15
+  i32.and
+  i32.eqz
+  i32.const 0
+  local.get $0
+  select
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 24
+   i32.const 71
+   i32.const 2
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  i32.const 16
+  i32.sub
+  local.tee $1
+  i32.load offset=4
+  i32.const -1
+  i32.ne
+  if
+   i32.const 0
+   i32.const 24
+   i32.const 73
+   i32.const 13
+   call $~lib/builtins/abort
+   unreachable
+  end
+  global.get $~lib/rt/stub/offset
+  local.get $1
+  i32.load
+  local.get $0
+  i32.add
+  i32.eq
+  if
+   local.get $1
+   global.set $~lib/rt/stub/offset
+  end
+ )
+ (func $start:rt/stub-realloc (; 9 ;) (type $FUNCSIG$v)
   i32.const 10
   i32.const 0
   call $~lib/rt/stub/__alloc
@@ -497,8 +578,23 @@
    call $~lib/builtins/abort
    unreachable
   end
+  global.get $rt/stub-realloc/ptr
+  call $~lib/rt/stub/__free
+  i32.const 1
+  i32.const 0
+  call $~lib/rt/stub/__alloc
+  global.get $rt/stub-realloc/ptr
+  i32.ne
+  if
+   i32.const 0
+   i32.const 72
+   i32.const 30
+   i32.const 0
+   call $~lib/builtins/abort
+   unreachable
+  end
  )
- (func $start (; 9 ;) (type $FUNCSIG$v)
+ (func $start (; 10 ;) (type $FUNCSIG$v)
   global.get $~lib/started
   if
    return
diff --git a/tests/compiler/rt/stub-realloc.ts b/tests/compiler/rt/stub-realloc.ts
index ccb6aa7e84..129011e08b 100644
--- a/tests/compiler/rt/stub-realloc.ts
+++ b/tests/compiler/rt/stub-realloc.ts
@@ -25,3 +25,6 @@ assert(ptr == originalPtr);
 
 ptr = __realloc(ptr, 33); // not last anymore: copies
 assert(ptr > originalPtr);
+
+__free(ptr); // discards last
+assert(__alloc(1, 0) == ptr);
diff --git a/tests/compiler/rt/stub-realloc.untouched.wat b/tests/compiler/rt/stub-realloc.untouched.wat
index 66e58360bc..b7fec9328b 100644
--- a/tests/compiler/rt/stub-realloc.untouched.wat
+++ b/tests/compiler/rt/stub-realloc.untouched.wat
@@ -34,14 +34,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -95,6 +87,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -105,25 +98,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
@@ -1395,7 +1402,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  (local $6 i32)
   local.get $0
   i32.const 0
   i32.ne
@@ -1411,7 +1417,7 @@
   if
    i32.const 0
    i32.const 24
-   i32.const 40
+   i32.const 43
    i32.const 2
    call $~lib/builtins/abort
    unreachable
@@ -1421,25 +1427,30 @@
   i32.sub
   local.set $2
   local.get $2
-  i32.load offset=12
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
+  i32.load
   local.set $3
-  local.get $0
-  local.get $3
-  i32.add
-  global.get $~lib/rt/stub/offset
+  local.get $2
+  i32.load offset=4
+  i32.const -1
   i32.eq
-  local.set $4
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 24
+   i32.const 46
+   i32.const 13
+   call $~lib/builtins/abort
+   unreachable
+  end
   local.get $1
   local.get $3
   i32.gt_u
   if
-   local.get $4
+   local.get $0
+   local.get $3
+   i32.add
+   global.get $~lib/rt/stub/offset
+   i32.eq
    if
     local.get $1
     i32.const 1073741808
@@ -1447,48 +1458,77 @@
     if
      unreachable
     end
-    local.get $0
     local.get $1
+    i32.const 15
+    i32.add
+    i32.const 15
+    i32.const -1
+    i32.xor
+    i32.and
+    local.set $3
+    local.get $0
+    local.get $3
     i32.add
     call $~lib/rt/stub/maybeGrowMemory
+    local.get $2
+    local.get $3
+    i32.store
    else    
     local.get $1
-    local.tee $5
+    i32.const 15
+    i32.add
+    i32.const 15
+    i32.const -1
+    i32.xor
+    i32.and
+    local.tee $4
     local.get $3
     i32.const 1
     i32.shl
-    local.tee $6
+    local.tee $5
+    local.get $4
     local.get $5
-    local.get $6
     i32.gt_u
     select
+    local.set $3
+    local.get $3
     local.get $2
     i32.load offset=8
     call $~lib/rt/stub/__alloc
-    local.set $5
-    local.get $5
+    local.set $4
+    local.get $4
     local.get $0
-    local.get $3
+    local.get $2
+    i32.load offset=12
     call $~lib/memory/memory.copy
-    local.get $5
+    local.get $4
     local.tee $0
     i32.const 16
     i32.sub
     local.set $2
    end
   else   
-   local.get $4
+   local.get $0
+   local.get $3
+   i32.add
+   global.get $~lib/rt/stub/offset
+   i32.eq
    if
-    local.get $0
     local.get $1
-    i32.add
     i32.const 15
     i32.add
     i32.const 15
     i32.const -1
     i32.xor
     i32.and
+    local.set $3
+    local.get $0
+    local.get $3
+    i32.add
     global.set $~lib/rt/stub/offset
+    local.get $2
+    local.get $3
+    i32.store
    end
   end
   local.get $2
@@ -1496,7 +1536,57 @@
   i32.store offset=12
   local.get $0
  )
- (func $start:rt/stub-realloc (; 9 ;) (type $FUNCSIG$v)
+ (func $~lib/rt/stub/__free (; 9 ;) (type $FUNCSIG$vi) (param $0 i32)
+  (local $1 i32)
+  local.get $0
+  i32.const 0
+  i32.ne
+  if (result i32)
+   local.get $0
+   i32.const 15
+   i32.and
+   i32.eqz
+  else   
+   i32.const 0
+  end
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 24
+   i32.const 71
+   i32.const 2
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  i32.const 16
+  i32.sub
+  local.set $1
+  local.get $1
+  i32.load offset=4
+  i32.const -1
+  i32.eq
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 24
+   i32.const 73
+   i32.const 13
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $0
+  local.get $1
+  i32.load
+  i32.add
+  global.get $~lib/rt/stub/offset
+  i32.eq
+  if
+   local.get $1
+   global.set $~lib/rt/stub/offset
+  end
+ )
+ (func $start:rt/stub-realloc (; 10 ;) (type $FUNCSIG$v)
   i32.const 10
   i32.const 0
   call $~lib/rt/stub/__alloc
@@ -1623,8 +1713,24 @@
    call $~lib/builtins/abort
    unreachable
   end
+  global.get $rt/stub-realloc/ptr
+  call $~lib/rt/stub/__free
+  i32.const 1
+  i32.const 0
+  call $~lib/rt/stub/__alloc
+  global.get $rt/stub-realloc/ptr
+  i32.eq
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 72
+   i32.const 30
+   i32.const 0
+   call $~lib/builtins/abort
+   unreachable
+  end
  )
- (func $start (; 10 ;) (type $FUNCSIG$v)
+ (func $start (; 11 ;) (type $FUNCSIG$v)
   global.get $~lib/started
   if
    return
@@ -1644,6 +1750,6 @@
   global.set $~lib/rt/stub/offset
   call $start:rt/stub-realloc
  )
- (func $null (; 11 ;) (type $FUNCSIG$v)
+ (func $null (; 12 ;) (type $FUNCSIG$v)
  )
 )
diff --git a/tests/compiler/runtime-stub.optimized.wat b/tests/compiler/runtime-stub.optimized.wat
index 5ac2180318..cf05ce677d 100644
--- a/tests/compiler/runtime-stub.optimized.wat
+++ b/tests/compiler/runtime-stub.optimized.wat
@@ -19,11 +19,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -65,6 +60,7 @@
  (func $~lib/rt/stub/__alloc (; 1 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -74,25 +70,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $~lib/rt/stub/__retain (; 2 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
   local.get $0
diff --git a/tests/compiler/runtime-stub.untouched.wat b/tests/compiler/runtime-stub.untouched.wat
index 15ab18ea48..e7cc0e83eb 100644
--- a/tests/compiler/runtime-stub.untouched.wat
+++ b/tests/compiler/runtime-stub.untouched.wat
@@ -24,14 +24,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -85,6 +77,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -95,25 +88,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/simd.untouched.wat b/tests/compiler/simd.untouched.wat
index 74e8c059ac..ed2aec3b8e 100644
--- a/tests/compiler/simd.untouched.wat
+++ b/tests/compiler/simd.untouched.wat
@@ -10,121 +10,7 @@
  (export "memory" (memory $0))
  (start $start)
  (func $simd/test_v128 (; 1 ;) (type $FUNCSIG$v)
-  v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d
-  v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 5
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d
-  v128.const i32x4 0x04030202 0x08070605 0x0c0b0a09 0x100f0e0d
-  i8x16.ne
-  i8x16.any_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 10
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d
-  v128.const i32x4 0x01010101 0x01010101 0x01010101 0x01010101
-  v128.and
-  v128.const i32x4 0x00010001 0x00010001 0x00010001 0x00010001
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 16
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d
-  v128.const i32x4 0x01010101 0x01010101 0x01010101 0x01010101
-  v128.or
-  v128.const i32x4 0x05030301 0x09070705 0x0d0b0b09 0x110f0f0d
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 23
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d
-  v128.const i32x4 0x01010101 0x01010101 0x01010101 0x01010101
-  v128.xor
-  v128.const i32x4 0x05020300 0x09060704 0x0d0a0b08 0x110e0f0c
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 30
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d
-  v128.not
-  v128.const i32x4 0xfbfcfdfe 0xf7f8f9fa 0xf3f4f5f6 0xeff0f1f2
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 37
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d
-  v128.const i32x4 0x0d0e0f10 0x090a0b0c 0x05060708 0x01020304
-  v128.const i32x4 0xff00ff00 0xff00ff00 0xff00ff00 0xff00ff00
-  v128.bitselect
-  v128.const i32x4 0x040e0210 0x080a060c 0x0c060a08 0x10020e04
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 43
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
+  nop
  )
  (func $simd/test_i8x16 (; 2 ;) (type $FUNCSIG$v)
   (local $0 v128)
@@ -321,170 +207,6 @@
    call $~lib/builtins/abort
    unreachable
   end
-  v128.const i32x4 0x7f7f7f7e 0x7f7f7f7f 0x7f7f7f7f 0x7f7f7f7f
-  i32.const 2
-  i8x16.splat
-  i8x16.add_saturate_s
-  i32.const 127
-  i8x16.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 85
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0xfffffffe 0xffffffff 0xffffffff 0xffffffff
-  i32.const 2
-  i8x16.splat
-  i8x16.add_saturate_u
-  i32.const -1
-  i8x16.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 91
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x80808081 0x80808080 0x80808080 0x80808080
-  i32.const 2
-  i8x16.splat
-  i8x16.sub_saturate_s
-  i32.const -128
-  i8x16.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 97
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x00000001 0x00000000 0x00000000 0x00000000
-  i32.const 2
-  i8x16.splat
-  i8x16.sub_saturate_u
-  i32.const 0
-  i8x16.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 103
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const 1
-  i8x16.splat
-  i32.const 1
-  i8x16.shl
-  i32.const 2
-  i8x16.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 109
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const -2
-  i8x16.splat
-  i32.const 1
-  i8x16.shr_s
-  i32.const -1
-  i8x16.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 110
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const -1
-  i8x16.splat
-  i32.const 1
-  i8x16.shr_u
-  i32.const 127
-  i8x16.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 111
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x00000001 0x00000000 0x00000000 0x00000000
-  i8x16.any_true
-  i32.const 0
-  i32.ne
-  i32.const 1
-  i32.eq
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 112
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const 1
-  i8x16.splat
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.const 1
-  i32.eq
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 113
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
   i32.const 0
   i8x16.splat
   i32.const 1
@@ -866,170 +588,6 @@
    call $~lib/builtins/abort
    unreachable
   end
-  v128.const i32x4 0x7fff7ffe 0x7fff7fff 0x7fff7fff 0x7fff7fff
-  i32.const 2
-  i16x8.splat
-  i16x8.add_saturate_s
-  i32.const 32767
-  i16x8.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 157
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0xfffffffe 0xffffffff 0xffffffff 0xffffffff
-  i32.const 2
-  i16x8.splat
-  i16x8.add_saturate_u
-  i32.const -1
-  i16x8.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 163
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x80008001 0x80008000 0x80008000 0x80008000
-  i32.const 2
-  i16x8.splat
-  i16x8.sub_saturate_s
-  i32.const -32768
-  i16x8.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 169
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x00000001 0x00000000 0x00000000 0x00000000
-  i32.const 2
-  i16x8.splat
-  i16x8.sub_saturate_u
-  i32.const 0
-  i16x8.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 175
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const 1
-  i16x8.splat
-  i32.const 1
-  i16x8.shl
-  i32.const 2
-  i16x8.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 181
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const -2
-  i16x8.splat
-  i32.const 1
-  i16x8.shr_s
-  i32.const -1
-  i16x8.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 182
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const -1
-  i16x8.splat
-  i32.const 1
-  i16x8.shr_u
-  i32.const 32767
-  i16x8.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 183
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x00000001 0x00000000 0x00000000 0x00000000
-  i16x8.any_true
-  i32.const 0
-  i32.ne
-  i32.const 1
-  i32.eq
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 184
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const 1
-  i16x8.splat
-  i16x8.all_true
-  i32.const 0
-  i32.ne
-  i32.const 1
-  i32.eq
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 185
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
   i32.const 0
   i16x8.splat
   i32.const 1
@@ -1325,153 +883,65 @@
    i32.const 24
    i32.const 211
    i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  local.get $2
-  i32x4.extract_lane 0
-  i32.const 2
-  i32.eq
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 216
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  local.get $2
-  i32x4.extract_lane 3
-  i32.const -2147483648
-  i32.eq
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 217
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  local.get $2
-  i32.const 5
-  i32x4.replace_lane 3
-  v128.const i32x4 0x00000002 0x00000003 0x00000004 0x00000005
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 218
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  local.get $0
-  local.get $1
-  v8x16.shuffle 0 1 2 3 4 5 6 7 24 25 26 27 28 29 30 31
-  v128.const i32x4 0x00000001 0x00000002 0x00000001 0x00000001
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 223
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const 1
-  i32x4.splat
-  i32.const 1
-  i32x4.shl
-  i32.const 2
-  i32x4.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 228
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const -2
-  i32x4.splat
-  i32.const 1
-  i32x4.shr_s
-  i32.const -1
-  i32x4.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
+   call $~lib/builtins/abort
+   unreachable
+  end
+  local.get $2
+  i32x4.extract_lane 0
+  i32.const 2
+  i32.eq
   i32.eqz
   if
    i32.const 0
    i32.const 24
-   i32.const 229
+   i32.const 216
    i32.const 2
    call $~lib/builtins/abort
    unreachable
   end
-  i32.const -1
-  i32x4.splat
-  i32.const 1
-  i32x4.shr_u
-  i32.const 2147483647
-  i32x4.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
+  local.get $2
+  i32x4.extract_lane 3
+  i32.const -2147483648
+  i32.eq
   i32.eqz
   if
    i32.const 0
    i32.const 24
-   i32.const 230
+   i32.const 217
    i32.const 2
    call $~lib/builtins/abort
    unreachable
   end
-  v128.const i32x4 0x00000001 0x00000000 0x00000000 0x00000000
-  i32x4.any_true
+  local.get $2
+  i32.const 5
+  i32x4.replace_lane 3
+  v128.const i32x4 0x00000002 0x00000003 0x00000004 0x00000005
+  i8x16.eq
+  i8x16.all_true
   i32.const 0
   i32.ne
-  i32.const 1
-  i32.eq
   i32.eqz
   if
    i32.const 0
    i32.const 24
-   i32.const 231
+   i32.const 218
    i32.const 2
    call $~lib/builtins/abort
    unreachable
   end
-  i32.const 1
-  i32x4.splat
-  i32x4.all_true
+  local.get $0
+  local.get $1
+  v8x16.shuffle 0 1 2 3 4 5 6 7 24 25 26 27 28 29 30 31
+  v128.const i32x4 0x00000001 0x00000002 0x00000001 0x00000001
+  i8x16.eq
+  i8x16.all_true
   i32.const 0
   i32.ne
-  i32.const 1
-  i32.eq
   i32.eqz
   if
    i32.const 0
    i32.const 24
-   i32.const 232
+   i32.const 223
    i32.const 2
    call $~lib/builtins/abort
    unreachable
@@ -1661,42 +1131,6 @@
    call $~lib/builtins/abort
    unreachable
   end
-  f32.const -1.5
-  f32x4.splat
-  i32x4.trunc_sat_f32x4_s
-  i32.const -1
-  i32x4.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 247
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  f32.const -1.5
-  f32x4.splat
-  i32x4.trunc_sat_f32x4_u
-  i32.const 0
-  i32x4.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 252
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
  )
  (func $simd/test_i64x2 (; 5 ;) (type $FUNCSIG$v)
   (local $0 v128)
@@ -1849,130 +1283,6 @@
    call $~lib/builtins/abort
    unreachable
   end
-  i64.const 1
-  i64x2.splat
-  i32.const 1
-  i64x2.shl
-  i64.const 2
-  i64x2.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 284
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i64.const -2
-  i64x2.splat
-  i32.const 1
-  i64x2.shr_s
-  i64.const -1
-  i64x2.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 285
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i64.const -1
-  i64x2.splat
-  i32.const 1
-  i64x2.shr_u
-  i64.const 9223372036854775807
-  i64x2.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 286
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  v128.const i32x4 0x00000001 0x00000000 0x00000000 0x00000000
-  i64x2.any_true
-  i32.const 0
-  i32.ne
-  i32.const 1
-  i32.eq
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 287
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i64.const 1
-  i64x2.splat
-  i64x2.all_true
-  i32.const 0
-  i32.ne
-  i32.const 1
-  i32.eq
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 288
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  f64.const -1.5
-  f64x2.splat
-  i64x2.trunc_sat_f64x2_s
-  i64.const -1
-  i64x2.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 289
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  f64.const -1.5
-  f64x2.splat
-  i64x2.trunc_sat_f64x2_u
-  i64.const 0
-  i64x2.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 294
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
  )
  (func $simd/test_f32x4 (; 6 ;) (type $FUNCSIG$v)
   (local $0 v128)
@@ -2351,58 +1661,6 @@
    call $~lib/builtins/abort
    unreachable
   end
-  v128.const i32x4 0x40800000 0x41100000 0x41800000 0x41c80000
-  f32x4.sqrt
-  v128.const i32x4 0x40000000 0x40400000 0x40800000 0x40a00000
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 339
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const -1
-  i32x4.splat
-  f32x4.convert_i32x4_s
-  f32.const -1
-  f32x4.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 340
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i32.const -1
-  i32x4.splat
-  f32x4.convert_i32x4_u
-  f32.const 4294967296
-  f32x4.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 345
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
  )
  (func $simd/test_f64x2 (; 7 ;) (type $FUNCSIG$v)
   (local $0 v128)
@@ -2781,58 +2039,6 @@
    call $~lib/builtins/abort
    unreachable
   end
-  v128.const i32x4 0x00000000 0x40100000 0x00000000 0x40220000
-  f64x2.sqrt
-  v128.const i32x4 0x00000000 0x40000000 0x00000000 0x40080000
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 390
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i64.const -1
-  i64x2.splat
-  f64x2.convert_i64x2_s
-  f64.const -1
-  f64x2.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 391
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
-  i64.const -1
-  i64x2.splat
-  f64x2.convert_i64x2_u
-  f64.const 18446744073709551615
-  f64x2.splat
-  i8x16.eq
-  i8x16.all_true
-  i32.const 0
-  i32.ne
-  i32.eqz
-  if
-   i32.const 0
-   i32.const 24
-   i32.const 396
-   i32.const 2
-   call $~lib/builtins/abort
-   unreachable
-  end
  )
  (func $simd/test_v8x16 (; 8 ;) (type $FUNCSIG$v)
   (local $0 v128)
diff --git a/tests/compiler/std/date.optimized.wat b/tests/compiler/std/date.optimized.wat
index 8aea8ac9ea..66ad05beb2 100644
--- a/tests/compiler/std/date.optimized.wat
+++ b/tests/compiler/std/date.optimized.wat
@@ -20,11 +20,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -69,20 +64,26 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $0
-  i32.const 8
+  local.tee $1
+  i32.const 16
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $0
+  local.get $1
   i32.const 16
   i32.sub
-  local.tee $1
+  local.tee $0
+  i32.const 16
+  i32.store
+  local.get $0
+  i32.const -1
+  i32.store offset=4
+  local.get $0
   i32.const 3
   i32.store offset=8
-  local.get $1
+  local.get $0
   i32.const 8
   i32.store offset=12
-  local.get $0
+  local.get $1
  )
  (func $start:std/date (; 5 ;) (type $FUNCSIG$v)
   (local $0 i32)
diff --git a/tests/compiler/std/date.untouched.wat b/tests/compiler/std/date.untouched.wat
index 1f4794ccb9..5624c16711 100644
--- a/tests/compiler/std/date.untouched.wat
+++ b/tests/compiler/std/date.untouched.wat
@@ -29,14 +29,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -90,6 +82,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -100,25 +93,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/std/new.optimized.wat b/tests/compiler/std/new.optimized.wat
index ad42bcf0fe..3e2d84ab44 100644
--- a/tests/compiler/std/new.optimized.wat
+++ b/tests/compiler/std/new.optimized.wat
@@ -12,11 +12,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -61,20 +56,26 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $0
-  i32.const 8
+  local.tee $1
+  i32.const 16
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $0
+  local.get $1
   i32.const 16
   i32.sub
-  local.tee $1
+  local.tee $0
+  i32.const 16
+  i32.store
+  local.get $0
+  i32.const -1
+  i32.store offset=4
+  local.get $0
   i32.const 3
   i32.store offset=8
-  local.get $1
+  local.get $0
   i32.const 8
   i32.store offset=12
-  local.get $0
+  local.get $1
  )
  (func $std/new/AClass#constructor (; 2 ;) (type $FUNCSIG$i) (result i32)
   (local $0 i32)
diff --git a/tests/compiler/std/new.untouched.wat b/tests/compiler/std/new.untouched.wat
index f9a4f9c1fd..e5cc939cfe 100644
--- a/tests/compiler/std/new.untouched.wat
+++ b/tests/compiler/std/new.untouched.wat
@@ -20,14 +20,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -81,6 +73,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -91,25 +84,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/std/object-literal.optimized.wat b/tests/compiler/std/object-literal.optimized.wat
index 32bc1fb1db..20c99ede7e 100644
--- a/tests/compiler/std/object-literal.optimized.wat
+++ b/tests/compiler/std/object-literal.optimized.wat
@@ -16,11 +16,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -62,6 +57,7 @@
  (func $~lib/rt/stub/__alloc (; 2 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -71,25 +67,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $~lib/string/String#get:length (; 3 ;) (type $FUNCSIG$ii) (param $0 i32) (result i32)
   local.get $0
diff --git a/tests/compiler/std/object-literal.untouched.wat b/tests/compiler/std/object-literal.untouched.wat
index 61a0219051..ebe13abbf6 100644
--- a/tests/compiler/std/object-literal.untouched.wat
+++ b/tests/compiler/std/object-literal.untouched.wat
@@ -22,14 +22,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -83,6 +75,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -93,25 +86,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/std/operator-overloading.optimized.wat b/tests/compiler/std/operator-overloading.optimized.wat
index 664e4f8f01..8f90298127 100644
--- a/tests/compiler/std/operator-overloading.optimized.wat
+++ b/tests/compiler/std/operator-overloading.optimized.wat
@@ -83,11 +83,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -132,20 +127,26 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $1
-  i32.const 8
+  local.tee $2
+  i32.const 16
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $1
+  local.get $2
   i32.const 16
   i32.sub
-  local.tee $2
+  local.tee $1
+  i32.const 16
+  i32.store
+  local.get $1
+  i32.const -1
+  i32.store offset=4
+  local.get $1
   local.get $0
   i32.store offset=8
-  local.get $2
+  local.get $1
   i32.const 8
   i32.store offset=12
-  local.get $1
+  local.get $2
  )
  (func $std/operator-overloading/Tester#constructor (; 3 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
diff --git a/tests/compiler/std/operator-overloading.untouched.wat b/tests/compiler/std/operator-overloading.untouched.wat
index 72b3266b2b..e7d5a02fd6 100644
--- a/tests/compiler/std/operator-overloading.untouched.wat
+++ b/tests/compiler/std/operator-overloading.untouched.wat
@@ -90,14 +90,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -151,6 +143,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -161,25 +154,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
diff --git a/tests/compiler/std/static-array.optimized.wat b/tests/compiler/std/static-array.optimized.wat
index d8f7839ba3..b604d86c82 100644
--- a/tests/compiler/std/static-array.optimized.wat
+++ b/tests/compiler/std/static-array.optimized.wat
@@ -56,11 +56,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -102,6 +97,7 @@
  (func $~lib/rt/stub/__alloc (; 3 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -111,25 +107,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $~lib/memory/memory.copy (; 4 ;) (type $FUNCSIG$viii) (param $0 i32) (param $1 i32) (param $2 i32)
   (local $3 i32)
@@ -321,31 +329,38 @@
   if
    i32.const 0
    i32.const 472
-   i32.const 40
+   i32.const 43
    i32.const 2
    call $~lib/builtins/abort
    unreachable
   end
-  global.get $~lib/rt/stub/offset
   local.get $0
   i32.const 16
   i32.sub
-  local.tee $2
-  i32.load offset=12
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $4
-  local.get $0
-  i32.add
-  i32.eq
-  local.set $3
+  local.tee $3
+  i32.load
+  local.set $2
+  local.get $3
+  i32.load offset=4
+  i32.const -1
+  i32.ne
+  if
+   i32.const 0
+   i32.const 472
+   i32.const 46
+   i32.const 13
+   call $~lib/builtins/abort
+   unreachable
+  end
   local.get $1
-  local.get $4
+  local.get $2
   i32.gt_u
   if
-   local.get $3
+   global.get $~lib/rt/stub/offset
+   local.get $0
+   local.get $2
+   i32.add
+   i32.eq
    if
     local.get $1
     i32.const 1073741808
@@ -353,47 +368,69 @@
     if
      unreachable
     end
-    local.get $0
     local.get $1
+    i32.const 15
+    i32.add
+    i32.const -16
+    i32.and
+    local.tee $2
+    local.get $0
     i32.add
     call $~lib/rt/stub/maybeGrowMemory
+    local.get $3
+    local.get $2
+    i32.store
    else    
     local.get $1
-    local.get $4
+    i32.const 15
+    i32.add
+    i32.const -16
+    i32.and
+    local.tee $4
+    local.get $2
     i32.const 1
     i32.shl
-    local.tee $3
-    local.get $1
-    local.get $3
+    local.tee $2
+    local.get $4
+    local.get $2
     i32.gt_u
     select
-    local.get $2
+    local.get $3
     i32.load offset=8
     call $~lib/rt/stub/__alloc
     local.tee $2
     local.get $0
-    local.get $4
+    local.get $3
+    i32.load offset=12
     call $~lib/memory/memory.copy
     local.get $2
     local.tee $0
     i32.const 16
     i32.sub
-    local.set $2
+    local.set $3
    end
   else   
-   local.get $3
+   global.get $~lib/rt/stub/offset
+   local.get $0
+   local.get $2
+   i32.add
+   i32.eq
    if
-    local.get $0
     local.get $1
-    i32.add
     i32.const 15
     i32.add
     i32.const -16
     i32.and
+    local.tee $2
+    local.get $0
+    i32.add
     global.set $~lib/rt/stub/offset
+    local.get $3
+    local.get $2
+    i32.store
    end
   end
-  local.get $2
+  local.get $3
   local.get $1
   i32.store offset=12
   local.get $0
diff --git a/tests/compiler/std/static-array.untouched.wat b/tests/compiler/std/static-array.untouched.wat
index 815a946ab7..69aa2c85d7 100644
--- a/tests/compiler/std/static-array.untouched.wat
+++ b/tests/compiler/std/static-array.untouched.wat
@@ -76,14 +76,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -137,6 +129,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -147,25 +140,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2
@@ -1428,7 +1435,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  (local $6 i32)
   local.get $0
   i32.const 0
   i32.ne
@@ -1444,7 +1450,7 @@
   if
    i32.const 0
    i32.const 472
-   i32.const 40
+   i32.const 43
    i32.const 2
    call $~lib/builtins/abort
    unreachable
@@ -1454,25 +1460,30 @@
   i32.sub
   local.set $2
   local.get $2
-  i32.load offset=12
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
+  i32.load
   local.set $3
-  local.get $0
-  local.get $3
-  i32.add
-  global.get $~lib/rt/stub/offset
+  local.get $2
+  i32.load offset=4
+  i32.const -1
   i32.eq
-  local.set $4
+  i32.eqz
+  if
+   i32.const 0
+   i32.const 472
+   i32.const 46
+   i32.const 13
+   call $~lib/builtins/abort
+   unreachable
+  end
   local.get $1
   local.get $3
   i32.gt_u
   if
-   local.get $4
+   local.get $0
+   local.get $3
+   i32.add
+   global.get $~lib/rt/stub/offset
+   i32.eq
    if
     local.get $1
     i32.const 1073741808
@@ -1480,48 +1491,77 @@
     if
      unreachable
     end
-    local.get $0
     local.get $1
+    i32.const 15
+    i32.add
+    i32.const 15
+    i32.const -1
+    i32.xor
+    i32.and
+    local.set $3
+    local.get $0
+    local.get $3
     i32.add
     call $~lib/rt/stub/maybeGrowMemory
+    local.get $2
+    local.get $3
+    i32.store
    else    
     local.get $1
-    local.tee $5
+    i32.const 15
+    i32.add
+    i32.const 15
+    i32.const -1
+    i32.xor
+    i32.and
+    local.tee $4
     local.get $3
     i32.const 1
     i32.shl
-    local.tee $6
+    local.tee $5
+    local.get $4
     local.get $5
-    local.get $6
     i32.gt_u
     select
+    local.set $3
+    local.get $3
     local.get $2
     i32.load offset=8
     call $~lib/rt/stub/__alloc
-    local.set $5
-    local.get $5
+    local.set $4
+    local.get $4
     local.get $0
-    local.get $3
+    local.get $2
+    i32.load offset=12
     call $~lib/memory/memory.copy
-    local.get $5
+    local.get $4
     local.tee $0
     i32.const 16
     i32.sub
     local.set $2
    end
   else   
-   local.get $4
+   local.get $0
+   local.get $3
+   i32.add
+   global.get $~lib/rt/stub/offset
+   i32.eq
    if
-    local.get $0
     local.get $1
-    i32.add
     i32.const 15
     i32.add
     i32.const 15
     i32.const -1
     i32.xor
     i32.and
+    local.set $3
+    local.get $0
+    local.get $3
+    i32.add
     global.set $~lib/rt/stub/offset
+    local.get $2
+    local.get $3
+    i32.store
    end
   end
   local.get $2
diff --git a/tests/compiler/std/symbol.optimized.wat b/tests/compiler/std/symbol.optimized.wat
index fec5e67cfb..0b2c961ca9 100644
--- a/tests/compiler/std/symbol.optimized.wat
+++ b/tests/compiler/std/symbol.optimized.wat
@@ -69,11 +69,6 @@
   (local $1 i32)
   (local $2 i32)
   local.get $0
-  i32.const 15
-  i32.add
-  i32.const -16
-  i32.and
-  local.tee $0
   memory.size
   local.tee $2
   i32.const 16
@@ -115,6 +110,7 @@
  (func $~lib/rt/stub/__alloc (; 3 ;) (type $FUNCSIG$iii) (param $0 i32) (param $1 i32) (result i32)
   (local $2 i32)
   (local $3 i32)
+  (local $4 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -124,25 +120,37 @@
   global.get $~lib/rt/stub/offset
   i32.const 16
   i32.add
-  local.tee $2
-  local.get $0
-  i32.const 1
+  local.tee $3
   local.get $0
-  i32.const 1
+  i32.const 15
+  i32.add
+  i32.const -16
+  i32.and
+  local.tee $2
+  i32.const 16
+  local.get $2
+  i32.const 16
   i32.gt_u
   select
+  local.tee $4
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
-  local.get $2
+  local.get $3
   i32.const 16
   i32.sub
-  local.tee $3
+  local.tee $2
+  local.get $4
+  i32.store
+  local.get $2
+  i32.const -1
+  i32.store offset=4
+  local.get $2
   local.get $1
   i32.store offset=8
-  local.get $3
+  local.get $2
   local.get $0
   i32.store offset=12
-  local.get $2
+  local.get $3
  )
  (func $~lib/memory/memory.fill (; 4 ;) (type $FUNCSIG$vii) (param $0 i32) (param $1 i32)
   (local $2 i32)
diff --git a/tests/compiler/std/symbol.untouched.wat b/tests/compiler/std/symbol.untouched.wat
index 4d2acb2831..d90a4709e6 100644
--- a/tests/compiler/std/symbol.untouched.wat
+++ b/tests/compiler/std/symbol.untouched.wat
@@ -94,14 +94,6 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
-  local.get $0
-  i32.const 15
-  i32.add
-  i32.const 15
-  i32.const -1
-  i32.xor
-  i32.and
-  local.set $0
   memory.size
   local.set $1
   local.get $1
@@ -155,6 +147,7 @@
   (local $3 i32)
   (local $4 i32)
   (local $5 i32)
+  (local $6 i32)
   local.get $0
   i32.const 1073741808
   i32.gt_u
@@ -165,25 +158,39 @@
   i32.const 16
   i32.add
   local.set $2
-  local.get $2
   local.get $0
+  i32.const 15
+  i32.add
+  i32.const 15
+  i32.const -1
+  i32.xor
+  i32.and
   local.tee $3
-  i32.const 1
+  i32.const 16
   local.tee $4
   local.get $3
   local.get $4
   i32.gt_u
   select
+  local.set $5
+  local.get $2
+  local.get $5
   i32.add
   call $~lib/rt/stub/maybeGrowMemory
   local.get $2
   i32.const 16
   i32.sub
-  local.set $5
+  local.set $6
+  local.get $6
   local.get $5
+  i32.store
+  local.get $6
+  i32.const -1
+  i32.store offset=4
+  local.get $6
   local.get $1
   i32.store offset=8
-  local.get $5
+  local.get $6
   local.get $0
   i32.store offset=12
   local.get $2