Nick Desaulniers

The enemy's gate is down

What's in a Word?

| Comments

Recently, there some was some confusion between myself and a coworker over the definition of a “word.” I’m currently working on a blog post about data alignment and figured it would be good to clarify some things now, that we can refer to later.

Having studied computer engineering and being quite fond of processor design, when I think of a “word,” I think of the number of bits wide a processor’s general purpose registers are (aka word size). This places hard requirements on the largest representable number and address space. A 64 bit processor can represent 264-1 (1.8x1019) as the largest unsigned long integer, and address up to 264-1 (16 EiB) different addresses in memory.

Further, word size limits the possible combinations of operations the processor can perform, length of immediate values used, inflates the size of binary files and memory needed to store pointers, and puts pressure on instruction caches.

Word size also has implications on loads and stores based on alignment, as we’ll see in a follow up post.

When I think of 8 bit computers, I think of my first microcontroller: an Arduino with an Atmel AVR processor. When I think of 16 bit computers, I think of my first game console, a Super Nintendo with a Ricoh 5A22. When I think of 32 bit computers, I think of my first desktop with Intel’s Pentium III. And when I think of 64 bit computers, I think modern smartphones with ARMv8 instruction sets. When someone mentions a particular word size, what are the machines that come to mind for you?

So to me, when someone’s talking about a 64b processor, to that machine (and me) a word is 64b. When we’re referring to a 8b processor, a word is 8b.

Now, some confusion.

Back in my previous blog posts about x86-64 assembly, JITs, or debugging, you might have seen me use instructions that have suffixes of b for byte (8b), w for word (16b), dw for double word (32b), and qw for quad word (64b) (since SSE2 there’s also double quadwords of 128b).

Wait a minute! How suddenly does a “word” refer to 16b on a 64b processor, as opposed to a 64b “word?”

In short, historical baggage. Intel’s first hit processor was the 4004, a 4b processor released in 1971. It wasn’t until 1979 that Intel created the 16b 8086 processor.

The 8086 was created to compete with other 16b processors that beat it to the market, like the Zilog Z80 (any Gameboy emulator fans out there? Yes, I know about the Sharp LR35902). The 8086 was the first design in the x86 family, and it allowed for the same assembly syntax from the earlier 8008, 8080, and 8085 to be reassembled for it. The 8086’s little brother (8088) would be used in IBM’s PC, and the rest is history. x86 would become one of the most successful ISAs in history.

For backwards compatibility, it seems that both Microsoft’s (whose success has tracked that of x86 since MS-DOS and IBM’s PC) and Intel’s documentation refers to words still as being 16b. This allowed 16b PE32+ executables to be run on 32b or even 64b newer versions of Windows, without requiring recompilation of source or source code modification.

This isn’t necessarily wrong to refer to a word based on backwards compatibility, it’s just important to understand the context in which the term “word” is being used, and that there might be some confusion if you have a background with x86 assembly, Windows API programming, or processor design.

So the next time someone asks: why does Intel’s documentation commonly refer to a “word” as 16b, you can tell them that the x86 and x86-64 ISAs have maintained the notion of a word being 16b since the first x86 processor, the 8086, which was a 16b processor.

Side Note: for an excellent historical perspective programming early x86 chips, I recommend Michael Abrash’s Graphics Programming Black Book. For instance he talks about 8086’s little brother, the 8088, being a 16b chip but only having an 8b bus with which to access memory. This caused a mysterious “cycle eater” to prevent fast access to 16b variables, though they were the processor’s natural size. Michael also alludes to alignment issues we’ll see in a follow up post.