fix asm instruction, make scratch allocator able to work per frame

alloc.d

@@ -15,8 +15,9 @@ const DEFAULT_ALIGNMENT = (void *).sizeof * 2;
 
 struct Scratch
 {
-	Arena arena;
-	bool  init;
+	Arena[] arenas;
+	u64     index;
+	bool    init;
 }
 
 struct ArenaPool
@@ -188,25 +189,34 @@ Free(T)(T* ptr)
 }
 
 void
-ResetScratch(u64 size)
+ResetScratch(u64 size, u64 count)
 {
-	if(!g_scratch.init)
+	with(g_scratch)
 	{
-		g_scratch.arena = CreateArena(size);
-		g_scratch.init = true;
-	}
+		if(!init)
+		{
+			arenas = AllocArray!(Arena)(count);
+			for(u64 i = 0; i < count; i += 1)
+			{
+				arenas[i] = CreateArena(size);
+			}
 
-	Reset(&g_scratch.arena);
+			g_scratch.init = true;
+		}
+
+		index = index % count;
+		Reset(&arenas[index]);
+	}
 }
 
 T*
 ScratchAlloc(T)()
 {
-	return Alloc!(T)(&g_scratch.arena);
+	return Alloc!(T)(&g_scratch.arenas[g_scratch.index]);
 }
 
 T[]
 ScratchAlloc(T)(u64 count)
 {
-	return AllocArray!(T)(&g_scratch.arena, count);
+	return AllocArray!(T)(&g_scratch.arenas[g_scratch.index], count);
 }

util.d

@@ -836,7 +836,7 @@ MemCpy(void* dst_p, void* src_p, u64 length)
 				movdqu XMM3, [R8+48];
 		
 				add R9, i;
-				movdqu [R9+00], XMM0;
+				movups [R9+00], XMM0;
 				movups [R9+16], XMM1;
 				movups [R9+32], XMM2;
 				movups [R9+48], XMM3;