# Stephen Smith's Blog

Musings on Machine Learning…

# Introduction

We think of computer processors executing a set of instructions one at a time in sequential order. As programmers this is exactly what we expect the computer to do and if the computer decided to execute our carefully written code in a different order then this terrifies us. We would expect our program to fail, producing wrong results or crashing. However we see manufacturers claiming their processors execute instructions out-of-order and that this is a feature that improves performance. In this article, we’ll look at what is really going on here and how it can benefit us, without causing too much fear.

# Disclaimer

ARM defines the Instruction Set Architecture (ISA), which defines the Assembly Language instruction set. ARM provides some reference implementations, but individual manufacturers can take these, customize these or develop their own independent implementation of the ARM instruction set. As a result the internal workings of ARM processors differs from manufacturer to manufacturer. A main point of difference is in performance optimizations. Apple is very aggressive in this regard, which is why the ARM processors in iPads and iPhones beat the competition. This means the level of out-of-order execution differs from manufacturer to manufacturer, further this is much more prevalent in newer ARM chips. As a result, the examples in this article will apply to a selection of ARM chips but not all.

# A Couple of Simple Cases

Consider the following small bit of code to multiply two numbers then load another number from memory and add it to the result of the multiplication:

MUL R3, R4, R5 @ R3 = R4 * R5
LDR R6, [R7]   @ Load R6 with the memory pointed to by R7
ADD R3, R6     @ R3 = R3 + R6

The ARM Processor is a RISC processor and its goal is to execute each instruction in 1 clock cycle. However multiplication is an exception and takes several clock cycles longer due to the loop of shifting and adding it has to perform internally. The load instruction doesn’t rely on the result of the multiplication and doesn’t involve the arithmetic unit. Thus it’s fairly simple for the ARM Processor to see this and execute the load while the multiply is still churning away. If the memory location is in cache, chances are the LDR will complete before the MUL and hence we say the instructions executed out-of-order. The ADD instruction then needs the results from both the MUL and LDR instruction, so it needs to wait for both of these to complete before executing it’s addition.

Consider another example of three LDR instructions:

LDR R1, [R4] @ memory in swap file
LDR R2, [R5] @ memory not in cache
LDR R3, [R6] @ memory in cache

Here the memory being loaded by the first instruction, has been swapped out of memory to secondary storage, so loading it is going to be slow. The second memory location is in regular memory. DDR4 memory, like that used in the new Raspberry Pi 4, is pretty fast, but not as fast as the CPU and it is also loading instructions to process, hence this second LDR might take a couple of cycles to execute. It makes a request to the memory controller and its request is queued with everything else going on. The third instruction, assumes the memory is in the CPU cache and hence processed immediately, so this instruction really does take only 1 clock cycle.

The upshot is that these three LDR instructions could well complete in reverse order.

Newer ARM processors can look ahead through the instructions looking for independent instructions to execute, the size of this pool will determine how out-of-order things can get. The important point is that instructions that have dependencies can’t start and that to the programmer, it looks like his code is executing in order and that all this magic is transparent to the correct execution of the program.

Since the CPU is executing all these instructions at once, you might wonder what the value of the program counter register (PC) is? This register has a very precisely defined value, since it is used for PC relative addressing. So the PC can’t be affected by out-of-order execution.

# Coprocessors

All newer ARM processors include floating-point coprocessors and NEON vector coprocessors. The instructions that execute on these usually take a few instructions cycles to execute. If the instructions that follow a coprocessor instruction are regular ARM instructions and don’t rely on the results of coprocessor operations, then they can continue to execute in parallel to the coprocessor. This is a handy way to get more code parallelism going, keeping all aspects of the CPU busy. Intermixing coprocessor and regular instructions is another great way to leverage out-of-order instructions to get better performance.

# Compilers and Code Generation

This indicates that if a compiler code generator or an Assembly Language program rearranges some of their instructions, they can get more things happening at once in parallel giving the program better performance. ARM Holdings contributes to the GNU Compiler Collection (GCC) to fully utilize the optimization present in their reference implementations. In the ARM specific options for GCC, you can select the ARM processor version that matches your target and get more advanced optimizations. Since Apple creates their own development tools under XCode, they can add optimizations specific to their custom ARM implementations.

As Assembly Language programmers, if we want to get the absolute best performance we might consider re-arranging some of our instructions so that instructions that are independent of each other are in a row and hopefully can be executed in parallel. This can require quite a bit of testing to reverse engineer the exact out-of-order instruction capability of your particular target ARM processor model. As always with performance optimizations, you must test the performance to prove you are improving things, and not just making your code more cryptic.

# Interrupts

This all sounds great, but what happens when an interrupt happens? This could be a timer interrupt to say your time-slice is up and another process gets to use the ARM Core, or it could be that more data needs to be read from the Wifi or a USB device.

Here the ARM CPU designer has a choice, they can forget about the work-in-progress and handle the interrupt quickly, or they can wait a couple of cycles to let work-in-progress complete and then handle the interrupt. Either way they have to allow the interrupt handler to save the current context and then restore the context to continue execution. Typically interrupt handlers do this by saving all the CPU and coprocessor registers to the system stack, doing their work and then restoring state.

When you see an ARM processor advertised as designed for real-time or industrial use, this typically means that it handles interrupts quickly with minimal delay. In this case, the work-in-progress is discarded and will be redone after the interrupt is finished. For ARM processors designed for general purpose computing, this usually means that user performance is more important than being super responsive to interrupts and hence they can let some of the work-in-progress complete before servicing the interrupt. For general purpose computing this is ok, since the attached devices like USB, ethernet and such have buffers that can hold enough contents to wait for the CPU to get around to them.

# A Step Too Far and Spectre

Hardware designers went even further with branch prediction, where if a conditional branch instruction needs to wait for a condition code to be set, they don’t wait but keep going assuming one branch direction (perhaps based on the result from the last time this code executed) and keep going. The problem here is that at this point, the CPU has to save the current state, incase it needs to go back when it guesses wrong. This CPU state was saved in a CPU cache that was only used for this, but had no security protection, resulting in the Spectre attack that figured out a way to get at this data. This caused data leakage across processes or even across virtual machines. The whole spectre debacle showed that great care has to be taken with these sorts of optimizations.

# Heat, the Ultimate Gotcha

Suppose your your ARM processor has four CPU cores and you write a brilliant Assembly language program that deploys to use all four cores and fully exploits out-of-order execution. Your program is now using every bit of the ARM CPU, each core is intermixing regular ARM, floating point and NEON instructions You have intermixed your ARM instructions to get the arithmetic unit operating in parallel to the memory unit. This will be the fastest implementation yet. Then you run your program, it gets off to a great start, but then suddenly slows down to a crawl. What happened?

The enemy of parallel processing on a single chip is heat. Everything the CPU does generates a little heat. The more things you get going at once the more heat will be generated by the CPU. Most ARM based computers like the Raspberry Pi assume you won’t be running the CPU so hard, and only provide heat dissipation for a more standard load. This is why Raspberry Pis usually do so badly playing high-res videos. They can do it, as long as they don’t overheat, which typically doesn’t take long.

This leaves you a real engineering problem. You need to either add more cooling to your target device, or you have to deliberately reduce the CPU usage of your program, where perhaps paradoxically you get more work done using two cores rather than four, because you won’t be throttled due to overheating.

# Summary

This was a quick overview of out-of-order instructions. Hopefully you don’t find these scary and keep in mind the potential benefits as you write your code. As newer ARM processors come to market, we’ll be seeing larger and larger pools of instructions executed in parallel, where the ability for instructions to execute out-of-order will have even greater benefits.

If you are interested in machine code or Assembly Language programming, be sure to check out my book: “Raspberry Pi Assembly Language Programming” from Apress. It is available on all major booksellers or directly from Apress here.

Written by smist08

November 15, 2019 at 11:11 am

This site uses Akismet to reduce spam. Learn how your comment data is processed.