diff --git a/src/assembler.h b/src/assembler.h
index 6d6033e38f..2ce04b44e7 100644
--- a/src/assembler.h
+++ b/src/assembler.h
@@ -273,10 +273,13 @@ class Assembler {
 
     virtual bool reserved(int register_) = 0;
 
+    virtual unsigned frameFootprint(unsigned footprint) = 0;
     virtual unsigned argumentFootprint(unsigned footprint) = 0;
     virtual unsigned argumentRegisterCount() = 0;
     virtual int argumentRegister(unsigned index) = 0;
 
+    virtual unsigned stackAlignmentInWords() = 0;
+
     virtual bool matchCall(void* returnAddress, void* target) = 0;
 
     virtual void updateCall(UnaryOperation op, bool assertAlignment,
@@ -322,7 +325,6 @@ class Assembler {
   virtual Architecture* arch() = 0;
 
   virtual void saveFrame(unsigned stackOffset, unsigned baseOffset) = 0;
-  virtual void restoreFrame(unsigned stackOffset, unsigned baseOffset) = 0;
   virtual void pushFrame(unsigned argumentCount, ...) = 0;
   virtual void allocateFrame(unsigned footprint) = 0;
   virtual void popFrame() = 0;
@@ -331,6 +333,8 @@ class Assembler {
                                    int framePointerSurrogate) = 0;
   virtual void popFrameAndPopArgumentsAndReturn(unsigned argumentFootprint)
   = 0;
+  virtual void popFrameAndUpdateStackAndReturn(unsigned stackOffsetFromThread)
+  = 0;
 
   virtual void apply(Operation op) = 0;
 
diff --git a/src/compile.cpp b/src/compile.cpp
index 20cbca17bf..3d2457869a 100644
--- a/src/compile.cpp
+++ b/src/compile.cpp
@@ -378,7 +378,7 @@ alignedFrameSize(MyThread* t, object method)
     (localSize(t, method)
      - methodParameterFootprint(t, method)
      + codeMaxStack(t, methodCode(t, method))
-     + t->arch->argumentFootprint(MaxNativeCallFootprint));
+     + t->arch->frameFootprint(MaxNativeCallFootprint));
 }
 
 unsigned
@@ -4835,6 +4835,17 @@ invokeNative2(MyThread* t, object method)
 
     default: abort(t);
     }
+
+    if (t->arch->argumentFootprint(footprint)
+        > t->arch->stackAlignmentInWords())
+    {
+      t->stack = static_cast<uintptr_t*>(t->stack)
+        + (t->arch->argumentFootprint(footprint)
+           - t->arch->stackAlignmentInWords());
+    }
+
+    t->stack = static_cast<uintptr_t*>(t->stack)
+      - t->arch->frameReturnAddressSize();
   } else {
     result = 0;
   }
@@ -6119,9 +6130,7 @@ compileThunks(MyThread* t, Allocator* allocator, MyProcessor* p,
     Assembler::Constant proc(&(nativeContext.promise));
     a->apply(LongCall, BytesPerWord, ConstantOperand, &proc);
   
-    a->popFrame();
-
-    a->apply(Return);
+    a->popFrameAndUpdateStackAndReturn(difference(&(t->stack), t));
 
     a->endBlock(false)->resolve(0, 0);
   }
diff --git a/src/powerpc.cpp b/src/powerpc.cpp
index 19c217a3e8..dc1e315ebb 100644
--- a/src/powerpc.cpp
+++ b/src/powerpc.cpp
@@ -65,6 +65,7 @@ inline int sth(int rs, int ra, int i) { return D(44, rs, ra, i); }
 inline int sthx(int rs, int ra, int rb) { return X(31, rs, ra, rb, 407, 0); }
 inline int stw(int rs, int ra, int i) { return D(36, rs, ra, i); }
 inline int stwu(int rs, int ra, int i) { return D(37, rs, ra, i); }
+inline int stwux(int rs, int ra, int i) { return X(31, rs, ra, rb, 183, 0); }
 inline int stwx(int rs, int ra, int rb) { return X(31, rs, ra, rb, 151, 0); }
 inline int add(int rt, int ra, int rb) { return XO(31, rt, ra, rb, 0, 266, 0); }
 inline int addc(int rt, int ra, int rb) { return XO(31, rt, ra, rb, 0, 10, 0); }
@@ -907,9 +908,15 @@ moveAndUpdateRM(Context* c, unsigned srcSize UNUSED, Assembler::Register* src,
 {
   assert(c, srcSize == BytesPerWord);
   assert(c, dstSize == BytesPerWord);
-  assert(c, dst->index == NoRegister);
 
-  issue(c, stwu(src->low, dst->base, dst->offset));
+  if (dst->index == NoRegister) {
+    issue(c, stwu(src->low, dst->base, dst->offset));
+  } else {
+    assert(c, dst->offset == 0);
+    assert(c, dst->scale == 1);
+    
+    issue(c, stwux(src->low, dst->base, dst->index));
+  }
 }
 
 void
@@ -2034,6 +2041,26 @@ class MyAssembler: public Assembler {
     return_(&c);
   }
 
+  virtual void popFrameAndUpdateStackAndReturn(unsigned stackOffset) {
+    popFrame();
+
+    Register tmp1(0);
+    Memory stackSrc(StackRegister, 0);
+    moveMR(&c, BytesPerWord, &stackSrc, BytesPerWord, &tmp1);
+
+    Register tmp2(arch_->returnLow());
+    Memory newStackSrc(ThreadRegister, stackOffsetFromThread);
+    moveMR(&c, BytesPerWord, &stackSrc, BytesPerWord, &tmp2);
+
+    Register stack(StackRegister);
+    subR(&c, BytesPerWord, &tmp2, &stack, &tmp2);
+
+    Memory stackDst(StackRegister, 0, tmp2.low);
+    moveAndUpdateRM(&c, BytesPerWord, &tmp1, BytesPerWord, &stackDst);
+
+    return_(&c);
+  }
+
   virtual void apply(Operation op) {
     arch_->c.operations[op](&c);
   }
diff --git a/src/x86.cpp b/src/x86.cpp
index b33b418263..8bdd62e4e5 100644
--- a/src/x86.cpp
+++ b/src/x86.cpp
@@ -2062,7 +2062,7 @@ class MyArchitecture: public Assembler::Architecture {
     }
   }
 
-  virtual unsigned stackPadding(unsigned footprint) {
+  virtual unsigned frameFootprint(unsigned footprint) {
     return max(footprint > argumentRegisterCount() ?
                footprint - argumentRegisterCount() : 0,
                StackAlignmentInWords);
@@ -2097,6 +2097,10 @@ class MyArchitecture: public Assembler::Architecture {
     }
   }
 
+  virtual unsigned stackAlignmentInWords() {
+    return StackAlignmentInWords;
+  }
+
   virtual bool matchCall(void* returnAddress, void* target) {
     uint8_t* instruction = static_cast<uint8_t*>(returnAddress) - 5;
     int32_t actualOffset; memcpy(&actualOffset, instruction + 1, 4);
@@ -2346,18 +2350,6 @@ class MyAssembler: public Assembler {
           BytesPerWord, MemoryOperand, &baseDst);
   }
 
-  virtual void restoreFrame(unsigned stackOffset, unsigned baseOffset) {
-    Register stack(rsp);
-    Memory stackDst(rbx, stackOffset);
-    apply(Move, BytesPerWord, MemoryOperand, &stackDst,
-          BytesPerWord, RegisterOperand, &stack);
-
-    Register base(rbp);
-    Memory baseDst(rbx, baseOffset);
-    apply(Move, BytesPerWord, MemoryOperand, &baseDst,
-          BytesPerWord, RegisterOperand, &base);
-  }
-
   virtual void pushFrame(unsigned argumentCount, ...) {
     struct {
       unsigned size;
@@ -2483,6 +2475,20 @@ class MyAssembler: public Assembler {
     }
   }
 
+  virtual void popFrameAndUpdateStackAndReturn(unsigned stackOffsetFromThread)
+  {
+    popFrame();
+
+    Register returnAddress(rcx);
+    popR(&c, BytesPerWord, &returnAddress);
+
+    Register stack(rsp);
+    Memory stackSrc(rbx, stackOffsetFromThread);
+    moveMR(&c, BytesPerWord, &stackSrc, BytesPerWord, &stack);
+
+    jumpR(&c, BytesPerWord, &returnAddress);
+  }
+
   virtual void apply(Operation op) {
     arch_->c.operations[op](&c);
   }