When people say that Twitter is a cesspool of conspiracy and abuse, I don’t recognise it based on my experience. My Twitter timeline is all jokes and geeky chat¹, and that’s where this post takes its cue:

When I started learning assembler, no site ever mentioned what registers were good for. Wish it had said:

CPU talking to a RAM chip is slow, registers are a bit of memory built into the CPU in which you load numbers from RAM, do several calculations, and only THEN write back.

I said that this was a RISC-centric approach and was challenged to come up with a better definition.

It’s a harder question to answer than I initially thought. Every time I came up with a pithy definition, I thought of an exception. And with every clarification I got further and further away from 280 characters.

With no character limit, what is a register?

Wikipedia says that it’s “a quickly accessible location available to a computer’s processor,” which is the definition that I was arguing against. Thanks, Wikipedia.

Nevertheless, I maintain that I’m right. Let’s dig into the definition further.

It wasn’t the “quick access” bit I didn’t like. Registers are faster than main memory². They’re also faster than on-die caches³.

The RISC-y bit was in the second sentence: loading bits of memory, do a few calculations, write back.

To explain why that’s not always true we have to take a quick tour of CPU design history. By necessity I’ve missed out many details or made sweeping generalisations⁴. I don’t think this detracts from my point.

First, a quick aside: for the sake of completeness, I should point out that there’s nothing sacred about registers. There are CPU architectures that do not have them, primarily stack-based but there may be others. Let’s ignore them for now.

There have been a few waves of CPU and instruction set architecture design, and the use of registers has changed over that time.

In The Beginning, making a CPU at all was a challenge. Getting more than a few thousand transistors on a single chip was hard! The first chips designers optimised for what was possible rather than to make things easy for programmers.

To minimise the number of transistors, there were few registers, and those that were present had predefined uses. One could be used for adding or subtracting (the accumulator). Some could be used to store memory addresses. Others might record the status of other operations. But you couldn’t use the address registers for arithmetic or vice versa.

The answer to the question “what is a register for” at this point is that it saves transistors. By wiring up the logic circuits to a single register and having few of them anyway, you could have enough space to do something.

As it became easier to add transistors to a slice of silicon, CPU designers started to make things easier for programmers. To get the best out of the limited hardware, many developers wrote code in assembler⁷. Therefore, making it easier for programmers meant adding new, more powerful hardware instructions.

The first CPUs might have had an instruction to “copy address n from memory into a register” and another to “add register 2 to the accumulator.” The next generation built on those foundations with instructions like “add the value at address n to the accumulator and write to address m.” The complexity of instructions grew over time.

Working on these early machines was hard partly because they had few registers and the instructions were simple. The new instructions made things easier by being able to work directly with the values in memory.

These new instructions were not magic, however. Having a single instruction to do the work didn’t make copying data from memory quicker. Developers trying to eke out the best performance had an increasingly difficult time figuring out the best combination of instructions. Is this single instruction that takes twenty cycles faster than these other three instructions that nominally does the same thing?

Around the same time, writing code in higher level languages became increasingly popular. The funny thing with compilers and interpreters is that it’s easier to write and optimise them when you use a limited set of instructions. All those esoteric instructions that were designed for people were rarely used when computers wrote the assembler code.

CPU designers figured out that they could make real-world code execute more quickly by heavily optimising a small number of instructions.

This led to completely different CPU designs, called RISC, or reduced instruction set chips. Rather than have a small number of special purpose registers and a large number of complex instructions, RISC insisted on large numbers of general purpose registers and few instructions⁸. The reduction in the instruction count was partly achieved by making arithmetic operations only work on registers. If you wanted to add two numbers stored in memory, you first had to load them into registers, add them together and write them back out to memory.

At this point, the answer to the question “what is a register for” changed. It became a sensible option to throw away the transistors used to implement many of the complex instructions and use them for general purpose registers. Lacking instructions that worked directly on memory, the definition of a register became “temporary storage for fast computations” — pretty much what we started with.

Since then, the original designs, with lots of instructions and a small number of registers (CISC), and the newer one, with lots of registers and few instructions (RISC), have to some extent merged. RISC chips have become less “pure” and have more special purpose instructions. CISC chips have gained more registers⁹ and, internally at least, have taken on many of the attributes traditionally attributed to RISC chips¹⁰.

Let’s loop back to the original question. Are registers a bit of memory built into the CPU in which you load numbers from RAM, do several calculations, and only then write back?

We’ve seen how registers are not necessary. We’ve see that their importance has waxed and waned. But, if we had to distill the answer down to a single word or sentence, what would that be?

On current hardware, on most machines, much of the time, the answer is: yes, they are a bit of memory built into the CPU in which you load numbers from RAM, do several calculations, and only then write back.

Was I being pedantic for no reason?