Skip to content

πŸ—οΈ Architectural Design Decisions

A.1 Always use modern C++

libhal uses the modern C++. Meaning that libhal is will follow the most modern and available compilers available. When a sufficient number of features have become available in both GCC & Clang and are determined to be useful to libhal libhal will increment its major number to indicate that it has upgraded compiler versions.

This decision exists to escape the issues of vendor and toolchain lock in thats prevalant in the C++ and embedded industry. With sufficient testing, upgrading compilers shouldn't result in bugs in applications.

A.2 Interface Design Choices

Interfaces MUST follow this layout:

  • Use #pragma once at the start of the file: Simpler than an include guard
  • All virtual functions must be private & each virtual functions is accompanied by a public API that is used to call the virtual API
  • ~~The return type of each API MUST be a result<T> where T is a structure.~~ Amendment 1.0: Return types should never be a structure with the expectation that it can be grown in the future. This is an ABI break.

Pragma once is needed to ensure files are included once. Its less error prone then hand writing include guards.

The reasons for a private virtual with public API can be found in this article (TODO: add link to article).

Returning a structure for each API means that, in the future, if the return type needs to be extended, it can be done without breaking down stream libraries. For example:

class adc {
  struct read_t { // V1
    float percentage;
  };
  struct read_t { // V2
    float percentage;
    // Optional field that is default initialized to std::nullopt indicating
    // that it defaults to not exist
    std::optional<uint8_t> bit_resolution = std::nullopt;
  };
};

Given that the field bit_resolution is an optional, code looking for it can determine if it is available or not, and code that never used it can ignore it.

A.2.1 No utility methods in interfaces (UFCS)

Utility functions shall not exist in interface definitions. For example, hal::i2c could have a hal::i2c::write() and hal::i2c::read() function implemented in its interface.

This has the effect of reducing the number of headers in the interface files and dependencies. This, in turn, results in an interface that is minimal, clean, and simple.

The major purpose of this is to keep compile times down as much possible for each interface. This also ensures that the "pay-for-what-you-use" model is followed. No need to pay for a utility you never planned to use.

The final reason is in preparation for UFCS (Unified Function Call Syntax). UFCS is a proposal for C++23 and C++26. It did not get into C++23 but is slated for review in 26. For more details see this page What is unified function call syntax anyway?.

A.3 Using tweak files over macros

Tweak files were used as an alternative to MACROS. MACROs can be quite problematic in many situations and are advised against in the core C++ guidelines. The benefits of tweak files can be found here.

A.4 ❌ ~~Header Only Implementations~~ ➑️ (See Amendment A.21)

libhal libraries and drivers are, in general, header-only. libhal uses header only implementations in order to enable the broadest set of package managers, build system and projects to use it.

The strongest reason for a header-only approach is due to the fact that libhal libraries never intend to be distributed in prebuilt binaries. Conan is designed to ship with prebuilt binaries or build against the host machine. These settings can be altered, but you still end up with a single global prebuilt binary for a driver does not make sense when that driver could be used in a variety of environments such as the host device for host side tests, a specific target device, and a target device that is in the family of that specific target device.

For example, lets consider liblpc40xx. If you are building to target the lpc4078 chip then that prebuilt ought to be built with usage of FPU registers enabled. But if you use that same prebuilt with the lpc4074, you'll find that the program crashes because the 74 variant does not have an FPU. You can attempt make a prebuilt binary for ever possible build variation that an embedded engineer may want, but you'll always come up short. The better approach is to simply build the library each time, thus ensuring that the build flags are considered each time.

If compile-times are a concern, there are reasonably easy methods for managing this. See Handling Long Compile Times.

A.5 Encapsulated Memory Mapped Classes

Target drivers that use Memory-Mapped-IO usually come with a vendor generated header file that describes each peripheral as a structure type, along with bit mask MACROs, and MACROs that result in pointers to each peripheral in memory. The main problem using these headers files causes is naming conflicts. Many of these vendor generated headers work with both C and C++. Meaning that namespaces are not utilized. And many do not expect that they will be used in an environment where another vendor generated header file will exists. So no care is taken to ensure that the names of the types are unique. This WILL cause linker errors as the linker sees both GPIO_TypeDef from an STM library and GPIO_TypeDef from an LPC library that aren't the same.

Because of this we have style S.x Encapsulated Memory Mapped classes guideline.

A.6 Using hal::function_ref over std::function&

std::function has all of the flexibility and functionality needed, but it has the potential to allocate and requires potentially expensive copy operations when passed by value.

hal::function_ref is a non-owning version of the std::function, with a size of just two pointers. hal::function_ref fits most use cases in that class functions that take them only need them for the duration of the function and do not need to own them for later.

!!! info hal::function_ref is an alias for tl:function_ref which comes from the project TartanLlama/function_ref.

A.7 Using virtual (runtime) polymorphism

Polymorphism is critical for libhal to reach the goals of flexible and easy of use. Static based polymorphism, by its nature, is inflexible at runtime and can be quite complicated to work with.

Runtime polymorphism, or the usage of virtual enables a broader scope of flexibility and isolation between drivers and application logic. The only downside to using virtual polymorphism is the cost of a virtual function call. But the actual cost of making a virtual function call is usually tiny in comparison to the work performed in the actual API call. In most cases the call latency and lack of inlining of a virtual call isn't an important factor in most applications.

And over all, along with the broad amount of flexibility comes the ease of use. Virtual polymorphism for interfaces is very easy to perform and has a ton of language support.

A.8 Strongly Leverage Package Managers

Finding and integration libraries into C++ programs is a pain. Doing the same thing for embedded is doubly so, especially if there is vendor IDE lock in. libhal seeks to escape this by using the available package managers and indexes.

Libhal was designed around and split up into parts that each come together via these package managers. The purpose of this design is to achieve:

  • Stable version and release control for each library
  • Can be easily found the indexes
  • Ease of integration

A.9 Foundation & Interface Stability

libhal-util, libhal-mock and libhal-soft were all apart of libhal originally, but due to the constant changes and API breaks in those categories of code, the version number of libhal would increment constantly, shifting the foundation of the ecosystem. To prevent constant churn and API breaks libhal was split into those 4 libraries.

The goal is to keep the version number for libhal constant for long periods of time to prevent breaking down stream libraries, drivers, and applications.

A.10 libhal driver directory

One of the libhal repos will contain a directory of libhal libraries that extend it along with which interfaces it implements and what type of library it is.

Official libhal libraries must go into the directory. Developers outside of the libhal organization can also contribute to and opt into this directory by making a PR to the repo containing the directory.

The purpose of this is to make finding and exploring the available set of drivers easier for the end developer by having them all in one place.

A.11 Github Actions & Remote Workflows

libhal uses github and github action "workflow_dispatch" to allow other repos to reuse libhal's continuous integration steps. The actions are configurable via input parameters to allow libraries to customize and control how the CI works.

libhal's CI attempts to use as many tools as reasonable to make sure that the C++ source code follows the style guide, C++ core guidelines and retains a certain level of quality. All offical libhal libaries must opt in to the common libhal/libhal workflow.

This helps to ensure that all projects are held to the same standard and quality. The workflow files can be found in libhal/libhal/.github/workflows.

A.12 Boost.UT as our unit testing framework

Boost.UT was chosen for its lack of macros, stunning compile time performance, and its ease of use.

A.13 ❌ ~~Boost.LEAF for error handling~~ ➑️ (See Amendment A.22)

One major issue with any project is handling errors. Because the libhal interfaces can be used in such broad environments, it is hard to determine what the BEST error type in advance could work for all users. Some use error codes, some use std::expected<T, E>, and some use exceptions.

Error codes are problematic as they tend to lack details and context around an error. Sometimes the documentation along with the error code provides all of the necessary context, but many times more context is needed.

std::expected<T, E> seems like a better alternative to error codes, but... is it really? What should E be? An error code? What if we have it be an error code and a const string. What if we want a file name and function name? What about a line number? What about 16 bytes for holding context information about the error? That should be enough, right? What about- what about- what about? ... wait, how big is this error type? 32 bytes? Wasn't this supposed to be light weight? Unfortunately, std::expected is not a good choice for interfaces with extremely broad and unknowable of error states. This forces the error type to be massive to accommodate everything and everyone.

Exceptions somewhat fix this issue but are still lacking. The benefit of exceptions is that you can throw just about anything, meaning the developer can provide loads of information in the thrown object. But exceptions fail on 4 counts:

  • Exceptions tend to not be available for embedded systems, either due to a toolchain not compiling with them enabled or because a project has strict requirements that forbid exceptions.
  • When exceptions do occur, the amount of time it takes to reach its catch can take a long time, longer than what real time applications can handle.
  • Normally requires heap allocation
  • Exceptions can only throw one type and the cost of those thrown exceptions are always paid for.

Boost.LEAF has the following properties:

  • Portable single-header format, no dependencies.
  • Tiny code size when configured for embedded development.
  • No dynamic memory allocations, even with very large payloads.
  • Deterministic unbiased efficiency on the "happy" path and the "sad" path.
  • Error objects are handled in constant time, independent of call stack depth.
  • Can be used with or without exception handling.
  • Can throw more than 1 error at a time

All of these features are critical for libhal to have the performance for real time applications.

The last feature is important for debugging, bug reports, and context specific error handling. Boost.LEAF gives the driver the choice to emit several error types and allows the user to pick out which one they would like to opt to catch if any of them. This can be used to capture an error code as well as s snapshot of the register map of a peripheral, the object's current state or even a debug message.

A.14 ~~Using Statement Expressions with HAL_CHECK()~~ ➑️ (See Amendment A.22)

HAL_CHECK() is the only MACRO in libhal. It exists because there is nothing like Rust's ? operator which either unwraps a value or returns an error from the current function. The "Statement Expression" only works with GCC & Clang which is one of the reasons why libhal only supports those compilers. Compare the following two expressions:

// 1. Using statement expressions
auto percentage = HAL_CHECK(adc.read()).percentage;

// 2. Without using statement expressions
HAL_CHECK(adc_read_temporary, adc.read());
auto percentage = adc_read_temporary.percentage;

The second option looks very unnatural and require explanation. On the other hand users who have never seen HAL_CHECK() in action have an immediate idea of how it works in the first section of the code. Portability to other compilers was sacrificed in order to make the code easier to read, understand, and write.

A.15 libhal WILL NOT use fixed point

Because fixed point will NOT result in better performance or space savings compared to SOFTWARE floating point. Team did venture to use fixed point throughout the entire code base and when we felt that the fixed point code reached a point where it was usable everywhere, we benchmarked it and got these:

double_time            = 8921794
[i64 +Round]fixed_time = 4558238 (best fixed point option)
[soft]float_time       = 1424913
[i64 -Round]fixed_time = 1410720 (precision issues)
[i32 +Round]fixed_time = 815107  (will easily overflow)
[hard]float_time       = 110089  (not always available)
[i32 -Round]fixed_time = 95085   (will not actually work)

Here is an old gist of the example: kammce/fixed_v_float.cpp

The above metrics were for a program that run a map function to map an input number from one range to another range. The numbers on the right hand side are the number of cycles of a Arm Cortex M4F DWT counter. Fixed point 32-bit integers is enough for a representation but to handle arithmetic like multiplication, 64-bit integers were needed. Those 64-bit operations resulted in computation time approaching double floating point. If a system used 32-bit floats, the 32-bit fixed point would be ~4x slower. If a system used double floating point in software mode, it will only be ~2x slower than 32-bit fixed point. Fixed point, over all, is more expensive in terms of space and time.

If you don't believe the metrics measured here, you can also check fpm performance metrics. Notice how fpm fairs far worse for anything that isn't addition/subtraction.

See these articles for more details:

A.16 libhal does NOT use a units library

Unit libraries have the potential to really help prevent an entire category of unit based errors, it is also extremely difficult and annoying to use.

Th article Unit of measurement libraries, their popularity and suitability goes into detail about the usability issues faced by unit libraries. Because, at the time of writing libhal there is not a unit library that is easy to use and concise, libhal decided to simply stick with 32-bit floats and helper UDLs.

A.17 ~~Always return hal::result<T> from every API~~ ➑️ (See Amendment A.22)

Every interface in libhal returns a hal::result<T> type.

The return types should be a result<T> because the implementation could be an abstraction for anything. As an example, it could come from an I2C to PWM generator and if something goes wrong with the i2c communication, the information must be emitted from the function.

A.18 Using inplace_function/hal::callback for interrupt callbacks

There are interfaces such as hal::can, hal::interrupt_pin, and hal::timer that all have APIs for setting a callback.

Because those callbacks could be lambdas, function objects, pure functions, or other callable types, we need a polymorphic type erased function type that can take any callable type as input and call it when its operator() is called.

The options for these callbacks are:

  • std::function
    • PROS
      • Part of the standard library
      • Can take any callable type without restrictions
    • CONS
      • Allocating (compiler implementations will use SBO but the size of those buffers are not specified in the standard and should not be relied upon)
      • Can be quite large in size (40 bytes on 32-bit arm)
  • function_ref
    • PROS
      • Very lightweight (very fast construction)
      • Very small size (2 pointers in size)
    • CONS
      • For this to work as a callback, the callable passed to the function_ref must have a lifetime that is greater than the object implementing the interface.
  • inplace_function
    • PROS
      • Works and behaves just like std::function
    • CONS
      • Fixed callable size limit

std::function is automatically out because it is allocating. Using std::function for any interface API would ensure that applications that disallow dynamic allocations after boot or in general could never use them.

function_ref has two great PROS but the largets CON is lifetime issues that are really easy to fall into. Specifically something like this:

obj.on_event([&single_capture]() {
  // does a thing  ...
});

The lambda is actually a temporary! So after this call it is out of scope and no longer exists. If the reference to temporary is stored and called later, the code WILL suffer from a "stack use after scope" violation which is undefined behavior.

inplace_function has all of the features of std::function but with limited size. Due to this, constructing an inplace_function is deterministic and relatively light weight.

A.19 hal::callback sizing

hal::callback is an alias to inplace_function with a buffer size of 2 pointers (sizeof(std::intptr_t) * 2). This size was chosen in order to be small and easily storable. Two pointers worth of size should be enough to hold a pointer to this in classes as well a pointer to some sort of state object.

The size of the callback object was not choosen in order to improve the performance of calling callback setting class functions. Even with the small size of hal::callback, its too large to take advantage of register based parameter passing. Thus the size of 2 pointers was mostly to help in keeping the memory footprint of the callback small. In most cases, setting an callback is something that is either done once or done very infrequently, and thus does not get much of a benefit from higher performance function calls.

A.20 ~~Why functions that setup events do not return hal::status~~ ➑️ (See Amendment A.22)

Functions like hal::can::on_receive() and hal::interrupt_pin::on_trigger() return void and not hal::status like other APIs. Thus these functions cannot return an error and are considered "infallible". There infallibility guarantee makes constructing drivers using these interfaces easier. It also eliminates the need for drivers to concern themselves with handling errors from these APIs.

This guarantee is easily made, because having any one of these APIs fail IS A bug and not something that a developer should or could be responsible with handling. These APIs MUST be implemented as target library peripheral drivers because setting interrupts is something that only target and processor libraries can do. Setting up and configuring interrupts is only possible if the processor supports it. Being apart of a target library means that they know exactly the set of possible configurations that are allowed. This also means that constructing a target peripheral with interrupt customization can be include compile time checks as well.

A.21 Critical importance of providing prebuilt binaries

This amends architectural component A.4 and pivots away from header only libraries over to prebuilt binaries.

The following are the reasons why supporting, producing and distibuting prebuilt binaries is critical to libhal.

  • ⏱️ Faster compilation times: When a project grows in size, compiling a C++ codebase can become time-consuming. By using precompiled binaries, the compiler has less work to do because portions of the code have already been compiled. It doesn't need to parse and compile the same headers over and over again. Developers will refuse to use libhal if it results in extremely long compilation times. They will tire of the project and seek faster to compile alternatives.
  • βœ… Consistency: With precompiled binaries, you can be sure that the same code will be compiled in the same way, which can reduce inconsistencies between different environments or build configurations. libhal will not be taken seriously if libhal didn't use prebuilt binaries and also support producing them. Distributing consistent code in a consistent form will be a requirement for many organizations seeking to use libhal.
  • πŸ“¦ Code distribution and updates: When distributing C++ code, rather than give out the full source code or require users to compile it themselves, providing a precompiled binary is more user-friendly. Updates to the program can also be handled by replacing the old binaries with the new ones, without requiring the end-user to handle the compilation process. The philosophy of conan regarding library packages is that, at their core, C++ libraries are header files and prebuilt binaries.
  • πŸ”’ Protection of Intellectual Property: Precompiled binaries are much harder to reverse-engineer than raw source code. If a developer wants to write an application and distribute the binary, but the C++ code contains proprietary algorithms or trade secrets that you don't want to expose to the public, distributing it as a precompiled binary can provide an additional layer of protection.

In order to provide the best experience for our developers as well as necessary features of the platform, libhal emphasizes prebuilt libraries over header only.

Note

Why the sudden change from A.4 to this? Supporting header-only made bringing up libhal much easier. The process of implementing prebuilt binaries required adding additional architecture settings, required the use of conan profiles to make the arm-gnu-toolchain available globally for all packages, figuring out how to insert architectural compiler flags and ensure they propogate to all packages took time and was a complex process. And header-only libraries allowed the initial version of libhal to bypass all of this.

A.22 Use C++ Exception for Error handling

libhal uses C++ exceptions for these reasons:

  1. Removes runtime cost of calling functions that error codes, result types, and sentinel variables.
  2. Reduces code size greatly by:
  3. Amortizing the error handling mechanism to a single set of functions rather than distributing the workload across all functions that can error.
  4. Unwind information is a compressed ISA that only needs 8 bits per instruction in ARM (32 and 64), where as normal ARM or Thumb-2 which require 16 to 32 bits of information, resulting in code bloat.
  5. GCC's Language Specific Data Area (LSDA) is a further compressed data structure for determining if a frame contains a suitable catch block or cleanup routine.

Issues regarding memory allocations can be overcome by simply overriding/ wrapping __cxa_allocate_exception and using a statically allocated buffer.In the case of multiple threads, each thread can be given its own memory buffer through TLS or a circular pool of blocks with an std::atomic<int> for every thread to use.

Exception throw runtime can be reduced by optimizing __cxa_throw's implementation.

The runtime is determinist because the grand majority of our code is statically linked. This means our code is not concerned with shared library locking the exception table down with a mutex to update it. This mutex lock will blocking threads from continuing to unwind, making the time non-determinist. Once you remove that, your unwind time is deterministic.

The issue with massive code bloat when you enable -fexceptions only comes from the linux version of std::terminate which uses <iostream> and some other locale related stuff. Simply overriding this with your own implmentation fixes this issue and reduces code size by ~100kB.

A link to a paper written by Khalil Estell will be linked here to describe in detail why C++ exception handling is the superior error handling mechanism for embedded systems and software in general when performance and code size are critical concerns.