Fix some corner cases where exceptions are thrown through +initialize.

Test/objc_msgSend.m

@@ -53,12 +53,22 @@ Class TestCls;
 + (void)initialize
 {
 	[self printf: "Format %s %d %f%c", "string", 42, 42.0, '\n'];
+	@throw self;
 }
 + nothing { return 0; }
 @end
 int main(void)
 {
 	TestCls = objc_getClass("Test");
+	int exceptionThrown = 0;
+	@try {
+		objc_msgSend(TestCls, @selector(foo));
+	} @catch (id e)
+	{
+		assert((TestCls == e) && "Exceptions propagate out of +initialize");
+		exceptionThrown = 1;
+	}
+	assert(exceptionThrown);
 	assert((id)0x42 == objc_msgSend(TestCls, @selector(foo)));
 	objc_msgSend(TestCls, @selector(nothing));
 	objc_msgSend(TestCls, @selector(missing));

dtable.c

@@ -621,6 +621,9 @@ static void remove_dtable(InitializingDtable* meta_buffer)
 {
 	LOCK(&initialize_lock);
 	InitializingDtable *buffer = meta_buffer->next;
+	// Install the dtable:
+	meta_buffer->class->dtable = meta_buffer->dtable;
+	buffer->class->dtable = buffer->dtable;
 	// Remove the look-aside buffer entry.
 	if (temporary_dtables == meta_buffer)
 	{
@@ -728,10 +731,4 @@ PRIVATE void objc_send_initialize(id object)
 	// insert it into a global list, even though it's a temporary variable,
 	// because we will clean it up after this function.
 	initializeSlot->method((id)class, initializeSel);
-
-	// Install the real dtable for both the class and the metaclass.  We can do
-	// this without the lock, because now both ways of accessing the dtable are
-	// safe.  We only need the lock when we remove the cached version.
-	meta->dtable = dtable;
-	class->dtable = class_dtable;
 }

objc_msgSend.x86-64.S

@@ -72,7 +72,6 @@
 	push  %rcx
 	push  %r8
 	push  %r9
-	//push  %r9                            # Fudge to ensure stack alignment stays at 16 bytes, or SSE dies
 	
 	sub $0x98, %rsp
 	movups	%xmm0, 0x80(%rsp)
@@ -99,7 +98,7 @@
 	mov   %rdx, %rsi                     # move _cmd to where the callee expects it to be
 .endif
 
-	.cfi_def_cfa_offset 80
+	.cfi_adjust_cfa_offset 0xD8
 	call  slowMsgLookup                  # Call the slow lookup function
 	mov   %rax, %r10                     # Load the returned IMP
 
@@ -117,7 +116,6 @@
 	movups	0x10(%rsp), %xmm7
 	add   $0x98, %rsp
 
-	//pop   %r9                            # Fudge to ensure stack alignment stays at 16 bytes, or SSE dies
 	pop   %r9
 	pop   %r8
 	pop   %rcx