When Intel expanded the 8086 architecture to 32-bit in 1985, they extended the 16-bit registers present to 32-bit registers. ax became eax, but it was still possible to use the low 16 bits of eax as ax just like before. Their choice was that performing operations on the low 16 bits did not change the high 16 bits of the register.

AMD expanded the 32-bit architecture to 64-bit in 2003. This was again a superset of the original, making it backwards compatible. They extended the 32-bit registers to 64-bit, and eax became rax. Again it was possible to to perform operations on the low 32 bits, but doing so clears the high 32 bits of the register.

Operations that output to a 32-bit subregister are automatically zero-extended to the entire 64-bit register. Operations that output to 8-bit or 16-bit subregisters are not zero-extended (this is compatible x86 behavior). [source]

Now both choices work as far as backwards compatibility goes, and as long as we as programmers are aware of what happens, neither is a problem.

When building the aPLib compression library, I use Visual C++ to generate assembly listings, which I then perform some changes on with a Perl script, before assembling the object files. While working on the recently released 64-bit version, I ran into a problem – the debug build of the library worked fine, but the release build did not.

Bugs like this are often caused by some improper memory usage, so I spent a day trying to track down the problem without much luck. Somehow the contents of a register was corrupted.

Looking through the code in HIEW I finally found the cause; a seemingly random instruction that wrote to the 32-bit part of a register, thereby clearing the high 32 bits. Then it dawned on me.

Visual C++ emits padding macros into assembly listings to align code and improve performance. These macros, npad, are defined in a file called listing.inc which resides in the Visual C++ include folder. But there is no 64-bit version of this file!

Let’s have a look then:

;; LISTING.INC

;; non destructive nops
npad macro size
if size eq 1
  nop
else
 if size eq 2
   mov edi, edi
 else
   ...

And there we have it. An instruction like mov edi, edi is safe to use as padding in 32-bit code, because moving the register to itself has no effect. But if you insert it in 64-bit code, it all of a sudden has an effect – the high 32 bits of rdi are cleared.

I have reported the problem to Microsoft and they say it will be addressed in a future release.