Ciro Santilli A computer is a highly layered system, and so you have to decide which layers you are the most interested in studying.
Although the layer are somewhat independent, they also sometimes interact, and when that happens it usually hurts your brain. E.g., if compilers were perfect, no one optimizing software would have to know anything about microarchitecture. But if you want to go hardcore enough, you might have to learn some lower layer.
It must also be said that like in any industry, certain layers are hidden in commercial secrecy mysteries making it harder to actually learn them. In computing, the lower level you go, the more closed source things tend to become.
But as you climb down into the abyss of low level hardcoreness, don't forget that making usefulness is more important than being hardcore: Figure 1. "xkcd 378: Real Programmers".
First, the most important thing you should know about this subject: cirosantilli.com/linux-kernel-module-cheat/should-you-waste-your-life-with-systems-programming
Here's a summary from low-level to high-level:
- semiconductor physical implementation this level is of course the most closed, but it is fun to try and peek into it from any openings given by commercials and academia:
- photolithography, and notably photomask design
- register transfer level
- interactive Verilator fun: Is it possible to do interactive user input and output simulation in VHDL or Verilog?
- more importantly, and much harder/maybe impossible with open source, would be to try and set up a open source standard cell library and supporting software to obtain power, performance and area estimates
- Are there good open source standard cell libraries to learn IC synthesis with EDA tools? on Quora
- the most open source ones are some initiatives targeting FPGAs, e.g. symbiflow.github.io/, www.clifford.at/icestorm/
- qflow is an initiative targeting actual integrated circuits
- microarchitecture: a good way to play with this is to try and run some minimal userland examples on gem5 userland simulation with logging, e.g. see on the Linux Kernel Module Cheat:This should be done at the same time as books/website/courses that explain the microarchitecture basics.This is the level of abstraction that Ciro Santilli finds the most interesting of the hardware stack. Learning it for actual CPUs (which as of 2020 is only partially documented by vendors) could actually be useful in hardcore software optimization use cases.
- instruction set architecture: a good approach to learn this is to manually write some userland assembly with assertions as done in the Linux Kernel Module Cheat e.g. at:
- github.com/cirosantilli/linux-kernel-module-cheat/blob/9b6552ab6c66cb14d531eff903c4e78f3561e9ca/userland/arch/x86_64/add.S
- cirosantilli.com/linux-kernel-module-cheat/x86-userland-assembly
- learn a bit about calling conventions, e.g. by calling C standar library functions from assembly:
- you can also try and understand what some simple C programs compile to. Things can get a bit hard though when
-O3is used. Some cute examples:
- executable file format, notably executable and Linkable Format. Particularly important is to understand the basics of:
- address relocation: How do linkers and address relocation work?
- position independent code: What is the -fPIE option for position-independent executables in GCC and ld?
- how to observe which symbols are present in object files, e.g.:
- how C++ uses name mangling What is the effect of extern "C" in C++?
- how C++ template instantiation can help reduce link time and size: Explicit template instantiation - when is it used?
- operating system. There are two ways to approach this:
- learn about the Linux kernel Linux kernel. A good starting point is to learn about its main interfaces. This is well shown at Linux Kernel Module Cheat:
- system calls
- write some system calls in
- pure assembly:
- C GCC inline assembly:
- write some system calls in
- learn about kernel modules and their interfaces. Notably, learn about to demystify special files such
/dev/randomand so on: - learn how to do a minimal Linux kernel disk image/boot to userland hello world: What is the smallest possible Linux implementation?
- learn how to GDB Step debug the Linux kernel itself. Once you know this, you will feel that "given enough patience, I could understand anything that I wanted about the kernel", and you can then proceed to not learn almost anything about it and carry on with your life
- system calls
- write your own (mini-) OS, or study a minimal educational OS, e.g. as in:
- learn about the Linux kernel Linux kernel. A good starting point is to learn about its main interfaces. This is well shown at Linux Kernel Module Cheat:
- programming language
This is a general principle of software/hardware design that Ciro feels holds wide applicability.
The most extreme case of this is of course the integrated circuit itself, in which it is essentially impossible (?) to observe the specific value of some indidual wire at some point.
Somewhat on the other extreme, we have high level programming languages running on top of an operating system: at this point, you can just GDB step debug your program, print the value of any variable/memory location, and fully understand anything that you want. Provided that you manage to easily reach that point of interest.
And for anything in between we have various intermediate levels of complication. The most notable perhaps being developing the operating system itself. At this level, you can't so easily step debug (although techniques do exist). For early boot or bootloaders for example, you might want to use JTAG for example on real hardware.
In parallel to this, there is also another very important pair of closely linked tradeoffs:
- the lower level at which something is implemented, the faster it runs
- emulation gives you observability back, at the cost of slower runtime
Emulation also has another potential downside: unless you are very careful at implementing things correctly, your model might not be representative of the real thing. Also, there may be important tradeoffs between how much the model looks like the real thing, and how fast it runs. For example, QEMU's use of binary translation allows it to run orders of magnitude faster than gem5. However, you are unable to make any predictions about system performance with QEMU, since you are not modelling key elements like the cache or CPU pipeline.
Instrumentation is another technique that has can be considered to achieve greater observability.
Instrumentation basically means adding loggers/print statements to certain points of interest of your hardware/software.
Instrumentation tends to slow execution down a bit, but way less than emulation.
The downside is that if the instrumentation does not provide you the data you need to debug, there's not much you can do, you will need to modify it, i.e. you don't get full visibility from instrumention.
This is unlike emulation that provides full observability.
The term loosely refers to certain layers of the computer abstraction layers hierarchy, usually high level hardware internals like CPU pipeline, caching and the memory system. Basically exactly what gem5 models.