diff --git a/xpcom/reflect/xptcall/src/md/unix/xptcinvoke_arm.cpp b/xpcom/reflect/xptcall/src/md/unix/xptcinvoke_arm.cpp
index beb5deac344..169115145fa 100644
--- a/xpcom/reflect/xptcall/src/md/unix/xptcinvoke_arm.cpp
+++ b/xpcom/reflect/xptcall/src/md/unix/xptcinvoke_arm.cpp
@@ -60,8 +60,65 @@
 
 // Remember that these 'words' are 32bit DWORDS
 
+static PRUint32
+invoke_count_words(PRUint32 paramCount, nsXPTCVariant* s)
+{
+    PRUint32 result = 0;
+    for(PRUint32 i = 0; i < paramCount; i++, s++)
+    {
+        if(s->IsPtrData())
+        {
+            result++;
+            continue;
+        }
+        switch(s->type)
+        {
+        case nsXPTType::T_I8     :
+        case nsXPTType::T_I16    :
+        case nsXPTType::T_I32    :
+            result++;
+            break;
+        case nsXPTType::T_I64    :
+            result+=VAR_STACK_SIZE_64;
+            break;
+        case nsXPTType::T_U8     :
+        case nsXPTType::T_U16    :
+        case nsXPTType::T_U32    :
+            result++;
+            break;
+        case nsXPTType::T_U64    :
+            result+=VAR_STACK_SIZE_64;
+            break;
+        case nsXPTType::T_FLOAT  :
+            result++;
+            break;
+        case nsXPTType::T_DOUBLE :
+            result+=VAR_STACK_SIZE_64;
+            break;
+        case nsXPTType::T_BOOL   :
+        case nsXPTType::T_CHAR   :
+        case nsXPTType::T_WCHAR  :
+            result++;
+            break;
+        default:
+            // all the others are plain pointer types
+            result++;
+            break;
+        }
+    }
 
-static PRUint32*
+#ifdef __ARM_EABI__
+    /* Ensure stack is always aligned to doubleword boundary; we take 3 words
+     * off the stack to r1-r3 later, so it must always be on _odd_ word 
+     * boundary after this */
+    if (result % 2 == 0)
+        result++;
+#endif
+
+    return result;
+}
+
+static void
 invoke_copy_to_stack(PRUint32* d, PRUint32 paramCount, nsXPTCVariant* s)
 {
     for(PRUint32 i = 0; i < paramCount; i++, d++, s++)
@@ -98,7 +155,6 @@ invoke_copy_to_stack(PRUint32* d, PRUint32 paramCount, nsXPTCVariant* s)
             break;
         }
     }
-    return d;
 }
 
 extern "C" {
@@ -107,6 +163,7 @@ extern "C" {
         PRUint32 Index;         
         PRUint32 Count;         
         nsXPTCVariant* params;  
+        PRUint32 fn_count;     
         PRUint32 fn_copy;      
     };
 }
@@ -122,6 +179,7 @@ NS_InvokeByIndex(nsISupports* that, PRUint32 methodIndex,
     my_params.Count = paramCount;
     my_params.params = params;
     my_params.fn_copy = (PRUint32) &invoke_copy_to_stack;
+    my_params.fn_count = (PRUint32) &invoke_count_words;
 
 /* This is to call a given method of class that.
  * The parameters are in params, the number is in paramCount.
@@ -149,20 +207,22 @@ NS_InvokeByIndex(nsISupports* that, PRUint32 methodIndex,
  */
  
   __asm__ __volatile__(
+    "ldr	r1, [%1, #12]	\n\t"	/* prepare to call invoke_count_words	*/
+    "ldr	ip, [%1, #16]	\n\t"	/* r0=paramCount, r1=params		*/
     "ldr	r0, [%1,  #8]	\n\t"
-    "mov	r4, r0, lsl #3	\n\t"	/* This is the amount of bytes needed.	*/
-#ifdef __ARM_EABI__
-    "add        r4, r4, #4      \n\t"
-#endif
+    "mov	lr, pc		\n\t"	/* call it...				*/
+    "mov	pc, ip		\n\t"
+    "mov	r4, r0, lsl #2	\n\t"	/* This is the amount of bytes needed.	*/
     "sub	sp, sp, r4	\n\t"	/* use stack space for the args...	*/
-    "add        r0, %1, #8      \n\t"
-    "ldmia      r0, {r1, r2, ip} \n\t"
     "mov	r0, sp		\n\t"	/* prepare a pointer an the stack	*/
-    "blx        ip              \n\t"
-    "sub        r4, r0, sp      \n\t"
-    
-    "ldmia      %1, {r0, r2}    \n\t"
+    "ldr	r1, [%1,  #8]	\n\t"	/* =paramCount				*/
+    "ldr	r2, [%1, #12]	\n\t"	/* =params				*/
+    "ldr	ip, [%1, #20]	\n\t"	/* =invoke_copy_to_stack		*/
+    "mov	lr, pc		\n\t"	/* copy args to the stack like the	*/
+    "mov	pc, ip		\n\t"	/* compiler would.			*/
+    "ldr	r0, [%1]	\n\t"	/* =that				*/
     "ldr	r1, [r0, #0]	\n\t"	/* get that->vtable offset		*/
+    "ldr	r2, [%1, #4]	\n\t"
     "mov	r2, r2, lsl #2	\n\t"	/* a vtable_entry(x)=8 + (4 bytes * x)	*/
 #if defined(__GXX_ABI_VERSION) && __GXX_ABI_VERSION >= 100 /* G++ V3 ABI */
     "ldr        ip, [r1, r2]    \n\t"   /* get method adress from vtable        */
@@ -177,7 +237,8 @@ NS_InvokeByIndex(nsISupports* that, PRUint32 methodIndex,
     "addle	sp, sp, r4	\n\t"	/*      and restore stack pointer	*/
     "movle	r4, #0		\n\t"	/*	a mark for restoring sp		*/
     "ldr	r0, [%1, #0]	\n\t"	/* get (self)				*/
-    "blx        ip              \n\t"
+    "mov	lr, pc		\n\t"	/* call mathod				*/
+    "mov	pc, ip		\n\t"
     "add	sp, sp, r4	\n\t"	/* restore stack pointer		*/
     "mov	%0, r0		\n\t"	/* the result...			*/
     : "=r" (result)