diff --git a/src/alloy/backend/x64/lowering/lowering_sequences.cc b/src/alloy/backend/x64/lowering/lowering_sequences.cc
index 45bbaa32e..576ad65a0 100644
--- a/src/alloy/backend/x64/lowering/lowering_sequences.cc
+++ b/src/alloy/backend/x64/lowering/lowering_sequences.cc
@@ -231,7 +231,7 @@ void IssueCall(X64Emitter& e, FunctionInfo* symbol_info, uint32_t flags) {
 
   // Actually jump/call to rax.
   if (flags & CALL_TAIL) {
-    e.add(e.rsp, StackLayout::GUEST_STACK_SIZE);
+    e.add(e.rsp, (uint32_t)e.stack_size());
     e.jmp(e.rax);
   } else {
     e.call(e.rax);
@@ -250,7 +250,7 @@ void IssueCallIndirect(X64Emitter& e, Value* target, uint32_t flags) {
 
   // Actually jump/call to rax.
   if (flags & CALL_TAIL) {
-    e.add(e.rsp, StackLayout::GUEST_STACK_SIZE);
+    e.add(e.rsp, (uint32_t)e.stack_size());
     e.jmp(e.rax);
   } else {
     e.call(e.rax);
@@ -397,6 +397,7 @@ table->AddSequence(OPCODE_CALL_EXTERN, [](X64Emitter& e, Instr*& i) {
     e.mov(e.r8, (uint64_t)symbol_info->extern_arg0());
     e.mov(e.r9, (uint64_t)symbol_info->extern_arg1());
     TransitionToHost(e);
+    ReloadRDX(e);
   }
   i = e.Advance(i);
   return true;
diff --git a/src/alloy/backend/x64/lowering/op_utils.inl b/src/alloy/backend/x64/lowering/op_utils.inl
index 3e0ed6789..5ec86b65d 100644
--- a/src/alloy/backend/x64/lowering/op_utils.inl
+++ b/src/alloy/backend/x64/lowering/op_utils.inl
@@ -28,7 +28,7 @@ namespace {
 void LoadEflags(X64Emitter& e) {
 #if STORE_EFLAGS
   e.mov(e.eax, e.dword[e.rsp + STASH_OFFSET]);
-  e.push(e.ax);
+  e.push(e.rax);
   e.popf();
 #else
   // EFLAGS already present.
@@ -37,7 +37,7 @@ void LoadEflags(X64Emitter& e) {
 void StoreEflags(X64Emitter& e) {
 #if STORE_EFLAGS
   e.pushf();
-  e.pop(e.word[e.rsp + STASH_OFFSET]);
+  e.pop(e.qword[e.rsp + STASH_OFFSET]);
 #else
   // EFLAGS should have CA set?
   // (so long as we don't fuck with it)
@@ -84,7 +84,7 @@ void MovMem64(X64Emitter& e, RegExp& addr, uint64_t v) {
 void CallNative(X64Emitter& e, void* target) {
   e.mov(e.rax, (uint64_t)target);
   e.call(e.rax);
-  e.mov(e.rcx, e.qword[e.rsp + StackLayout::RCX_HOME]);
+  e.mov(e.rcx, e.qword[e.rsp + StackLayout::GUEST_RCX_HOME]);
   e.mov(e.rdx, e.qword[e.rcx + 8]); // membase
 }
 
diff --git a/src/alloy/backend/x64/x64_code_cache.cc b/src/alloy/backend/x64/x64_code_cache.cc
index 2b2bf322d..7282c2e23 100644
--- a/src/alloy/backend/x64/x64_code_cache.cc
+++ b/src/alloy/backend/x64/x64_code_cache.cc
@@ -267,10 +267,7 @@ void X64CodeChunk::AddTableEntry(uint8_t* code, size_t code_size,
     unwind_code.OpInfo = stack_size / 8 - 1;
   } else {
     // TODO(benvanik): take as parameters?
-    uint8_t prolog_size = 17;
-
-    // This doesn't work, for some reason.
-    XEASSERTALWAYS();
+    uint8_t prolog_size = 7;
 
     // http://msdn.microsoft.com/en-us/library/ddssxxy8.aspx
     UNWIND_INFO* unwind_info = (UNWIND_INFO*)(buffer + unwind_info_offset);
@@ -284,7 +281,7 @@ void X64CodeChunk::AddTableEntry(uint8_t* code, size_t code_size,
     // http://msdn.microsoft.com/en-us/library/ck9asaa9.aspx
     size_t co = 0;
     auto& unwind_code = unwind_info->UnwindCode[co++];
-    unwind_code.CodeOffset = 17; // end of instruction + 1 == offset of next instruction
+    unwind_code.CodeOffset = 7; // end of instruction + 1 == offset of next instruction
     unwind_code.UnwindOp = UWOP_ALLOC_LARGE;
     unwind_code.OpInfo = 0;
     unwind_code = unwind_info->UnwindCode[co++];
diff --git a/src/alloy/backend/x64/x64_emitter.cc b/src/alloy/backend/x64/x64_emitter.cc
index 5aee9fb59..c8ddaddb4 100644
--- a/src/alloy/backend/x64/x64_emitter.cc
+++ b/src/alloy/backend/x64/x64_emitter.cc
@@ -64,14 +64,15 @@ int X64Emitter::Emit(
   }
 
   // Fill the generator with code.
-  int result = Emit(builder);
+  size_t stack_size = 0;
+  int result = Emit(builder, stack_size);
   if (result) {
     return result;
   }
 
   // Copy the final code to the cache and relocate it.
   out_code_size = getSize();
-  out_code_address = Emplace(StackLayout::GUEST_STACK_SIZE);
+  out_code_address = Emplace(stack_size);
 
   // Stash source map.
   if (debug_info_flags & DEBUG_INFO_SOURCE_MAP) {
@@ -99,7 +100,7 @@ void* X64Emitter::Emplace(size_t stack_size) {
 
 #define XEALIGN(value, align) ((value + align - 1) & ~(align - 1))
 
-int X64Emitter::Emit(HIRBuilder* builder) {
+int X64Emitter::Emit(HIRBuilder* builder, size_t& out_stack_size) {
   // These are the registers we will not be using. All others are fare game.
   const uint32_t reserved_regs =
       GetRegBit(rax)  | // scratch
@@ -125,7 +126,7 @@ int X64Emitter::Emit(HIRBuilder* builder) {
   // Calculate stack size. We need to align things to their natural sizes.
   // This could be much better (sort by type/etc).
   auto locals = builder->locals();
-  size_t stack_offset = 0;
+  size_t stack_offset = StackLayout::GUEST_STACK_SIZE;
   for (auto it = locals.begin(); it != locals.end(); ++it) {
     auto slot = *it;
     size_t type_size = GetTypeSize(slot->type);
@@ -134,6 +135,9 @@ int X64Emitter::Emit(HIRBuilder* builder) {
     slot->set_constant(stack_offset);
     stack_offset += type_size;
   }
+  // Ensure 16b alignment.
+  stack_offset -= StackLayout::GUEST_STACK_SIZE;
+  stack_offset = XEALIGN(stack_offset, 16);
 
   // Function prolog.
   // Must be 16b aligned.
@@ -147,11 +151,13 @@ int X64Emitter::Emit(HIRBuilder* builder) {
   //     X64CodeCache, which dynamically generates exception information.
   //     Adding or changing anything here must be matched!
   const bool emit_prolog = true;
-  const size_t stack_size = StackLayout::GUEST_STACK_SIZE;
+  const size_t stack_size = StackLayout::GUEST_STACK_SIZE + stack_offset;
+  XEASSERT((stack_size + 8) % 16 == 0);
+  out_stack_size = stack_size;
+  stack_size_ = stack_size;
   if (emit_prolog) {
-    mov(qword[rsp + 8 * 2], rdx);
-    mov(qword[rsp + 8 * 1], rcx);
-    sub(rsp, stack_size);
+    sub(rsp, (uint32_t)stack_size);
+    mov(qword[rsp + StackLayout::GUEST_RCX_HOME], rcx);
   }
 
   auto lowering_table = backend_->lowering_table();
@@ -187,9 +193,8 @@ int X64Emitter::Emit(HIRBuilder* builder) {
   // Function epilog.
   L("epilog");
   if (emit_prolog) {
-    add(rsp, stack_size);
-    mov(rcx, qword[rsp + 8 * 1]);
-    mov(rdx, qword[rsp + 8 * 2]);
+    mov(rcx, qword[rsp + StackLayout::GUEST_RCX_HOME]);
+    add(rsp, (uint32_t)stack_size);
   }
   ret();
 
diff --git a/src/alloy/backend/x64/x64_emitter.h b/src/alloy/backend/x64/x64_emitter.h
index c5bc51e05..4962dab14 100644
--- a/src/alloy/backend/x64/x64_emitter.h
+++ b/src/alloy/backend/x64/x64_emitter.h
@@ -145,9 +145,11 @@ public:
 
   void MarkSourceOffset(hir::Instr* i);
 
+  size_t stack_size() const { return stack_size_; }
+
 protected:
   void* Emplace(size_t stack_size);
-  int Emit(hir::HIRBuilder* builder);
+  int Emit(hir::HIRBuilder* builder, size_t& out_stack_size);
 
 protected:
   runtime::Runtime* runtime_;
@@ -168,6 +170,8 @@ protected:
 
   size_t    source_map_count_;
   Arena     source_map_arena_;
+
+  size_t    stack_size_;
 };
 
 
diff --git a/src/alloy/backend/x64/x64_thunk_emitter.cc b/src/alloy/backend/x64/x64_thunk_emitter.cc
index 0bd7239f6..7fc6fab60 100644
--- a/src/alloy/backend/x64/x64_thunk_emitter.cc
+++ b/src/alloy/backend/x64/x64_thunk_emitter.cc
@@ -38,7 +38,8 @@ HostToGuestThunk X64ThunkEmitter::EmitHostToGuestThunk() {
   mov(qword[rsp + 8 * 1], rcx);
   sub(rsp, stack_size);
 
-  mov(qword[rsp + 56], rbx);
+  mov(qword[rsp + 48], rbx);
+  mov(qword[rsp + 56], rcx);
   mov(qword[rsp + 64], rbp);
   mov(qword[rsp + 72], rsi);
   mov(qword[rsp + 80], rdi);
@@ -74,7 +75,8 @@ HostToGuestThunk X64ThunkEmitter::EmitHostToGuestThunk() {
   movaps(xmm14, ptr[rsp + 256]);
   movaps(xmm15, ptr[rsp + 272]);*/
 
-  mov(rbx, qword[rsp + 56]);
+  mov(rbx, qword[rsp + 48]);
+  mov(rcx, qword[rsp + 56]);
   mov(rbp, qword[rsp + 64]);
   mov(rsi, qword[rsp + 72]);
   mov(rdi, qword[rsp + 80]);
@@ -104,7 +106,8 @@ GuestToHostThunk X64ThunkEmitter::EmitGuestToHostThunk() {
   mov(qword[rsp + 8 * 1], rcx);
   sub(rsp, stack_size);
 
-  mov(qword[rsp + 56], rbx);
+  mov(qword[rsp + 48], rbx);
+  mov(qword[rsp + 56], rcx);
   mov(qword[rsp + 64], rbp);
   mov(qword[rsp + 72], rsi);
   mov(qword[rsp + 80], rdi);
@@ -120,7 +123,8 @@ GuestToHostThunk X64ThunkEmitter::EmitGuestToHostThunk() {
   mov(r8, r9);
   call(rax);
 
-  mov(rbx, qword[rsp + 56]);
+  mov(rbx, qword[rsp + 48]);
+  mov(rcx, qword[rsp + 56]);
   mov(rbp, qword[rsp + 64]);
   mov(rsi, qword[rsp + 72]);
   mov(rdi, qword[rsp + 80]);
@@ -128,7 +132,7 @@ GuestToHostThunk X64ThunkEmitter::EmitGuestToHostThunk() {
   mov(r13, qword[rsp + 96]);
   mov(r14, qword[rsp + 104]);
   mov(r15, qword[rsp + 112]);
-  
+
   add(rsp, stack_size);
   mov(rcx, qword[rsp + 8 * 1]);
   mov(rdx, qword[rsp + 8 * 2]);
diff --git a/src/alloy/backend/x64/x64_thunk_emitter.h b/src/alloy/backend/x64/x64_thunk_emitter.h
index 6559ab9a5..a9f27650b 100644
--- a/src/alloy/backend/x64/x64_thunk_emitter.h
+++ b/src/alloy/backend/x64/x64_thunk_emitter.h
@@ -25,15 +25,18 @@ namespace x64 {
  * ----------------------------
  * NOTE: stack must always be 16b aligned.
  *
+ * Thunk stack:
  *  +------------------+
  *  | arg temp, 3 * 8  | rsp + 0
  *  |                  |
  *  |                  |
  *  +------------------+
- *  | scratch, 24b     | rsp + 32
+ *  | scratch, 16b     | rsp + 32
  *  |                  |
  *  +------------------+
- *  | rbx              | rsp + 56
+ *  | rbx              | rsp + 48
+ *  +------------------+
+ *  | rcx / context    | rsp + 56
  *  +------------------+
  *  | rbp              | rsp + 64
  *  +------------------+
@@ -90,17 +93,30 @@ namespace x64 {
  *  |                  |
  *  +------------------+
  *
+ * Guest stack:
+ *  +------------------+
+ *  | arg temp, 3 * 8  | rsp + 0
+ *  |                  |
+ *  |                  |
+ *  +------------------+
+ *  | scratch, 32b     | rsp + 32
+ *  |                  |
+ *  +------------------+
+ *  | rcx / context    | rsp + 64
+ *  +------------------+
+ *    ... locals ...
+ *  +------------------+
+ *  | (return address) |
+ *  +------------------+ 
+ *
  */
 
 class StackLayout {
 public:
-  const static size_t GUEST_STACK_SIZE = 120;
-
   const static size_t THUNK_STACK_SIZE = 120;
 
-  const static size_t RETURN_ADDRESS = 120;
-  const static size_t RCX_HOME = 128;
-  const static size_t RDX_HOME = 136;
+  const static size_t GUEST_STACK_SIZE = 72;
+  const static size_t GUEST_RCX_HOME = 64;
 };