VM fixes for Win64 calling convention

vm/cpu-x86.32.S

@@ -10,14 +10,18 @@ and the callstack top is passed in EDX */
 #define DS_REG %esi
 #define RETURN_REG %eax
 
+#define NV_TEMP_REG %rbx
+
 #define CELL_SIZE 4
 #define STACK_PADDING 12
 
 #define PUSH_NONVOLATILE \
 	push %ebx ; \
+	push %ebp ; \
 	push %ebp
 
 #define POP_NONVOLATILE \
+	pop %ebp ; \
 	pop %ebp ; \
 	pop %ebx
 

vm/cpu-x86.64.S

@@ -7,6 +7,8 @@
 #define CELL_SIZE 8
 #define STACK_PADDING 56
 
+#define NV_TEMP_REG %rbp
+
 #ifdef WINDOWS
 
 	#define ARG0 %rcx
@@ -20,9 +22,11 @@
 		push %rdi ; \
 		push %rsi ; \
 		push %rbx ; \
+		push %rbp ; \
 		push %rbp
 
 	#define POP_NONVOLATILE \
+		pop %rbp ; \
 		pop %rbp ; \
 		pop %rbx ; \
 		pop %rsi ; \
@@ -41,9 +45,11 @@
 		push %rbx ; \
 		push %rbp ; \
 		push %r12 ; \
+		push %r13 ; \
 		push %r13
 
 	#define POP_NONVOLATILE \
+		pop %r13 ; \
 		pop %r13 ; \
 		pop %r12 ; \
 		pop %rbp ; \

vm/cpu-x86.S

@@ -1,20 +1,21 @@
 DEF(F_FASTCALL void,c_to_factor,(CELL quot)):
 	PUSH_NONVOLATILE
-	push ARG0
+	mov ARG0,NV_TEMP_REG
-
-	/* Save stack pointer */
-	lea -CELL_SIZE(STACK_REG),ARG0      
 
 	/* Create register shadow area for Win64 */
-	sub $32,STACK_REG                    
+	sub $32,STACK_REG
+
+	/* Save stack pointer */
+	lea -CELL_SIZE(STACK_REG),ARG0
 	call MANGLE(save_callstack_bottom)
-	add $32,STACK_REG
 
 	/* Call quot-xt */
-	mov (STACK_REG),ARG0
+	mov NV_TEMP_REG,ARG0
 	call *QUOT_XT_OFFSET(ARG0)
 
-	pop ARG0
+	/* Tear down register shadow area */
+	add $32,STACK_REG
+
 	POP_NONVOLATILE
 	ret