diff --git a/macros/smartalign.mac b/macros/smartalign.mac
new file mode 100644
index 00000000..1a631edc
--- /dev/null
+++ b/macros/smartalign.mac
@@ -0,0 +1,255 @@
+;
+; Smart alignment macros
+;
+USE: smartalign
+
+%imacro alignmode 1-2.nolist
+  %ifnempty %2
+    %xdefine __ALIGN_JMP_THRESHOLD__ %2
+  %else
+    %define __ALIGN_JMP_THRESHOLD__ 17
+  %endif
+  %ifidni %1,nop
+    %define __ALIGN_16BIT_1B__ 0x90
+    %define __ALIGN_16BIT_2B__ 0x90,0x90
+    %define __ALIGN_16BIT_3B__ 0x90,0x90,0x90
+    %define __ALIGN_16BIT_4B__ 0x90,0x90,0x90,0x90
+    %define __ALIGN_16BIT_5B__ 0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_16BIT_6B__ 0x90,0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_16BIT_7B__ 0x90,0x90,0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_16BIT_8B__ 0x90,0x90,0x90,0x90,0x90,0x90,0x90,0x90
+
+    %define __ALIGN_32BIT_1B__ 0x90
+    %define __ALIGN_32BIT_2B__ 0x90,0x90
+    %define __ALIGN_32BIT_3B__ 0x90,0x90,0x90
+    %define __ALIGN_32BIT_4B__ 0x90,0x90,0x90,0x90
+    %define __ALIGN_32BIT_5B__ 0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_32BIT_6B__ 0x90,0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_32BIT_7B__ 0x90,0x90,0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_32BIT_8B__ 0x90,0x90,0x90,0x90,0x90,0x90,0x90,0x90
+
+    %define __ALIGN_64BIT_1B__ 0x90
+    %define __ALIGN_64BIT_2B__ 0x90,0x90
+    %define __ALIGN_64BIT_3B__ 0x90,0x90,0x90
+    %define __ALIGN_64BIT_4B__ 0x90,0x90,0x90,0x90
+    %define __ALIGN_64BIT_5B__ 0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_64BIT_6B__ 0x90,0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_64BIT_7B__ 0x90,0x90,0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_64BIT_8B__ 0x90,0x90,0x90,0x90,0x90,0x90,0x90,0x90
+  %elifidni %1,generic
+    %define __ALIGN_16BIT_1B__ 0x90
+    %define __ALIGN_16BIT_2B__ 0x90,0x90
+    %define __ALIGN_16BIT_3B__ 0x90,0x90,0x90
+    %define __ALIGN_16BIT_4B__ 0x90,0x90,0x90,0x90
+    %define __ALIGN_16BIT_5B__ 0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_16BIT_6B__ 0x90,0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_16BIT_7B__ 0x90,0x90,0x90,0x90,0x90,0x90,0x90
+    %define __ALIGN_16BIT_8B__ 0x90,0x90,0x90,0x90,0x90,0x90,0x90,0x90
+
+    %define __ALIGN_32BIT_1B__ 0x90
+    %define __ALIGN_32BIT_2B__ 0x89,0xf6
+    %define __ALIGN_32BIT_3B__ 0x8d,0x76,0x00
+    %define __ALIGN_32BIT_4B__ 0x8d,0x74,0x26,0x00
+    %define __ALIGN_32BIT_5B__ 0x90,0x8d,0x74,0x26,0x00
+    %define __ALIGN_32BIT_6B__ 0x8d,0xb6,0x00,0x00,0x00,0x00
+    %define __ALIGN_32BIT_7B__ 0x8d,0xb4,0x26,0x00,0x00,0x00,0x00
+    %undef  __ALIGN_32BIT_8B__
+
+    %define __ALIGN_64BIT_1B__ 0x90
+    %define __ALIGN_64BIT_2B__ 0x66,0x90
+    %define __ALIGN_64BIT_3B__ 0x66,0x66,0x90
+    %define __ALIGN_64BIT_4B__ 0x66,0x66,0x66,0x90
+    %define __ALIGN_64BIT_5B__ 0x66,0x66,0x90,0x66,0x90
+    %define __ALIGN_64BIT_6B__ 0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_64BIT_7B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_64BIT_8B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x66,0x90
+  %elifidni %1,k8
+    %define __ALIGN_16BIT_1B__ 0x90
+    %define __ALIGN_16BIT_2B__ 0x66,0x90
+    %define __ALIGN_16BIT_3B__ 0x66,0x66,0x90
+    %define __ALIGN_16BIT_4B__ 0x66,0x66,0x66,0x90
+    %define __ALIGN_16BIT_5B__ 0x66,0x66,0x90,0x66,0x90
+    %define __ALIGN_16BIT_6B__ 0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_16BIT_7B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_16BIT_8B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x66,0x90
+
+    %define __ALIGN_32BIT_1B__ 0x90
+    %define __ALIGN_32BIT_2B__ 0x66,0x90
+    %define __ALIGN_32BIT_3B__ 0x66,0x66,0x90
+    %define __ALIGN_32BIT_4B__ 0x66,0x66,0x66,0x90
+    %define __ALIGN_32BIT_5B__ 0x66,0x66,0x90,0x66,0x90
+    %define __ALIGN_32BIT_6B__ 0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_32BIT_7B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_32BIT_8B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x66,0x90
+
+    %define __ALIGN_64BIT_1B__ 0x90
+    %define __ALIGN_64BIT_2B__ 0x66,0x90
+    %define __ALIGN_64BIT_3B__ 0x66,0x66,0x90
+    %define __ALIGN_64BIT_4B__ 0x66,0x66,0x66,0x90
+    %define __ALIGN_64BIT_5B__ 0x66,0x66,0x90,0x66,0x90
+    %define __ALIGN_64BIT_6B__ 0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_64BIT_7B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_64BIT_8B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x66,0x90
+  %elifidni %1,k7
+    %define __ALIGN_16BIT_1B__ 0x90
+    %define __ALIGN_16BIT_2B__ 0x66,0x90
+    %define __ALIGN_16BIT_3B__ 0x66,0x66,0x90
+    %define __ALIGN_16BIT_4B__ 0x66,0x66,0x66,0x90
+    %define __ALIGN_16BIT_5B__ 0x66,0x66,0x90,0x66,0x90
+    %define __ALIGN_16BIT_6B__ 0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_16BIT_7B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_16BIT_8B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x66,0x90
+
+    %define __ALIGN_32BIT_1B__ 0x90
+    %define __ALIGN_32BIT_2B__ 0x8b,0xc0
+    %define __ALIGN_32BIT_3B__ 0x8d,0x04,0x20
+    %define __ALIGN_32BIT_4B__ 0x8d,0x44,0x20,0x00
+    %define __ALIGN_32BIT_5B__ 0x8d,0x44,0x20,0x00,0x90
+    %define __ALIGN_32BIT_6B__ 0x8d,0x80,0x00,0x00,0x00,0x00
+    %define __ALIGN_32BIT_7B__ 0x8d,0x04,0x05,0x00,0x00,0x00,0x00
+    %undef  __ALIGN_32BIT_8B__
+
+    %define __ALIGN_64BIT_1B__ 0x90
+    %define __ALIGN_64BIT_2B__ 0x66,0x90
+    %define __ALIGN_64BIT_3B__ 0x66,0x66,0x90
+    %define __ALIGN_64BIT_4B__ 0x66,0x66,0x66,0x90
+    %define __ALIGN_64BIT_5B__ 0x66,0x66,0x90,0x66,0x90
+    %define __ALIGN_64BIT_6B__ 0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_64BIT_7B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_64BIT_8B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x66,0x90
+  %elifidni %1,p6
+    %define __ALIGN_16BIT_1B__ 0x90
+    %define __ALIGN_16BIT_2B__ 0x66,0x90
+    %define __ALIGN_16BIT_3B__ 0x66,0x66,0x90
+    %define __ALIGN_16BIT_4B__ 0x66,0x66,0x66,0x90
+    %define __ALIGN_16BIT_5B__ 0x66,0x66,0x90,0x66,0x90
+    %define __ALIGN_16BIT_6B__ 0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_16BIT_7B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x90
+    %define __ALIGN_16BIT_8B__ 0x66,0x66,0x66,0x90,0x66,0x66,0x66,0x90
+
+    %define __ALIGN_32BIT_1B__ 0x90
+    %define __ALIGN_32BIT_2B__ 0x66,0x90
+    %define __ALIGN_32BIT_3B__ 0x0f,0x1f,0x00
+    %define __ALIGN_32BIT_4B__ 0x0f,0x1f,0x40,0x00
+    %define __ALIGN_32BIT_5B__ 0x0f,0x1f,0x44,0x00,0x00
+    %define __ALIGN_32BIT_6B__ 0x66,0x0f,0x1f,0x44,0x00,0x00
+    %define __ALIGN_32BIT_7B__ 0x0f,0x1f,0x80,0x00,0x00,0x00,0x00
+    %define __ALIGN_32BIT_8B__ 0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
+
+    %define __ALIGN_64BIT_1B__ 0x90
+    %define __ALIGN_64BIT_2B__ 0x66,0x90
+    %define __ALIGN_64BIT_3B__ 0x0f,0x1f,0x00
+    %define __ALIGN_64BIT_4B__ 0x0f,0x1f,0x40,0x00
+    %define __ALIGN_64BIT_5B__ 0x0f,0x1f,0x44,0x00,0x00
+    %define __ALIGN_64BIT_6B__ 0x66,0x0f,0x1f,0x44,0x00,0x00
+    %define __ALIGN_64BIT_7B__ 0x0f,0x1f,0x80,0x00,0x00,0x00,0x00
+    %define __ALIGN_64BIT_8B__ 0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
+  %else
+    %error unknown alignment mode: %1
+  %endif
+  %xdefine __ALIGNMODE__ %1,__ALIGN_JMP_THRESHOLD__
+%endmacro
+
+%unimacro align 1-2+.nolist
+%imacro align 1-2+.nolist
+  %ifnempty %2
+    times ($$-$) % (%1) %2
+  %else
+    %push
+    %assign %$pad ($$-$) % %1
+    %if %$pad >= __ALIGN_JMP_THRESHOLD__
+      jmp %%end
+      ; We can't re-use %$pad here as $ will have changed!
+      times ($$-$) % %1 db 90h
+%%end:
+    %else
+      %if __BITS__ == 16
+        %ifdef __ALIGN_16BIT_8B__
+          %rep %$pad / 8
+            db __ALIGN_16BIT_8B__
+          %endrep
+          %assign %$pad %$pad % 8
+        %else
+          %rep %$pad / 7
+            db __ALIGN_16BIT_7B__
+          %endrep
+          %assign %$pad %$pad % 7
+        %endif
+        %if %$pad == 1
+          db __ALIGN_16BIT_1B__
+        %elif %$pad == 2
+          db __ALIGN_16BIT_2B__
+        %elif %$pad == 3
+          db __ALIGN_16BIT_3B__
+        %elif %$pad == 4
+          db __ALIGN_16BIT_4B__
+        %elif %$pad == 5
+          db __ALIGN_16BIT_5B__
+        %elif %$pad == 6
+          db __ALIGN_16BIT_6B__
+        %elif %$pad == 7
+          db __ALIGN_16BIT_7B__
+        %endif
+      %elif __BITS__ == 32
+        %ifdef __ALIGN_32BIT_8B__
+          %rep %$pad / 8
+            db __ALIGN_32BIT_8B__
+          %endrep
+          %assign %$pad %$pad % 8
+        %else
+          %rep %$pad / 7
+            db __ALIGN_32BIT_7B__
+          %endrep
+          %assign %$pad %$pad % 7
+        %endif
+        %if %$pad == 1
+          db __ALIGN_32BIT_1B__
+        %elif %$pad == 2
+          db __ALIGN_32BIT_2B__
+        %elif %$pad == 3
+          db __ALIGN_32BIT_3B__
+        %elif %$pad == 4
+          db __ALIGN_32BIT_4B__
+        %elif %$pad == 5
+          db __ALIGN_32BIT_5B__
+        %elif %$pad == 6
+          db __ALIGN_32BIT_6B__
+        %elif %$pad == 7
+          db __ALIGN_32BIT_7B__
+        %endif
+      %elif __BITS__ == 64
+        %ifdef __ALIGN_64BIT_8B__
+          %rep %$pad / 8
+            db __ALIGN_64BIT_8B__
+          %endrep
+          %assign %$pad %$pad % 8
+        %else
+          %rep %$pad / 7
+            db __ALIGN_64BIT_7B__
+          %endrep
+          %assign %$pad %$pad % 7
+        %endif
+        %if %$pad == 1
+          db __ALIGN_64BIT_1B__
+        %elif %$pad == 2
+          db __ALIGN_64BIT_2B__
+        %elif %$pad == 3
+          db __ALIGN_64BIT_3B__
+        %elif %$pad == 4
+          db __ALIGN_64BIT_4B__
+        %elif %$pad == 5
+          db __ALIGN_64BIT_5B__
+        %elif %$pad == 6
+          db __ALIGN_64BIT_6B__
+        %elif %$pad == 7
+          db __ALIGN_64BIT_7B__
+        %endif
+      %else
+        %error "Invalid __BITS__ value"
+      %endif
+    %endif
+    %pop
+  %endif
+%endmacro
+
+	alignmode generic
diff --git a/test/smartalign.asm b/test/smartalign.asm
new file mode 100644
index 00000000..65bdfcf4
--- /dev/null
+++ b/test/smartalign.asm
@@ -0,0 +1,23 @@
+%use smartalign
+
+	bits 32
+
+	alignmode nop
+	add ax,ax
+	align 16
+
+	alignmode generic
+	add ax,ax
+	align 16
+
+	alignmode k7
+	add ax,ax
+	align 16
+
+	alignmode k8
+	add ax,ax
+	align 16
+
+	alignmode p6
+	add ax,ax
+	align 16