Custom Search

Sunday, March 17, 2013

Idiomatic C and Arduino

The few of you who read this blog regularly will be wondering about "C". Not Python, not some new-fangled functional language, but good old-fashioned C, the workhorse language of the last 30 years or so. That's because I'm talking Arduino, and C - sorta kinda - is the language of choice for that platform.

For those not familiar with it, Arduino is an open source hardware platform built around the ATMega micro-controller CPUs. The boards - whose design is covered by a Creative Commons Attribution Share-Alike license - provide the micro-controller and support circuitry to make using external devices easy, and have a standardized form factor for adding expansion hardware. There is also an IDE for writing code, for the C/C++ framework to run on the controller. The boards - or clones of them - are available for as little as $15 shipped

These are 8 bit CPUs with memory sizes (flash, eeprom, and ram) measured in K, or maybe 10s of K. There is no way the run time system of a modern language will run on them, though there is a forth system or two available, as well as some custom interpreters. You can use a modern language to control the hardware if you run it on the desktop with some kind of serial link between the devices. But normally you cross-compile a small language - like C or a DSL for hardware systems like Atom - for the thing. C isn't much of a reach - these systems aren't much less powerful than the machines C was developed on, though just enough so that a C compiler probably won't fit.

The project

As mentioned, the Arduino boards are designed to make it easy to use external hardware, and the standard examples start with LEDs. They're cheap, easy to hook up, and hard to screw up.

So, you give a geek a bunch of LEDs that can be controlled by a computer, and the first thing they're going to do is build a binary clock, unless they have a grid of LEDs, in which case they'll play life on them. Because that is, after all cool - unless you're not a geek, in which case it's geeky.

Once I noticed that there were wearable Arduino platforms available, I figured a binary wrist watch was a natural. So I built a prototype:

I'm not going to discuss the code here, but it has been added to the blog repository. What I want to discuss are the implementations of the binary clocks I found.

The code

None of the implementations felt like idiomatic C to me. Possibly the problem is with me - I learned C long enough ago that =+ was a valid operator, and this code is more C++ than C. Possibly it's the way hardware people (which is where Arduino's roots are) approach C. Possibly it's the result of learning programming with relatively modern languages instead of things like C and it's predecessors. But we're going to look at one function: display_value(uint8_t *, uint16_t)

display_value accepts a pointer to an array of pin numbers (Arduino output devices, which for this purpose can be set to either 0 or 1 to turn an LED off or on), and a 16 bit unsigned value to display on those LEDs. A common variation is to use a starting pin number and assume the pins are assigned in order, but this doesn't work well with lots of LEDs, as some pins have special capabilities, and wiring constraints may cause you to want to use pins in strange orders. The array of pin numbers allows them to be assigned to whatever pins are convenient.

One common version of the function (usually just coded in line) looks like so:
void
display_value(uint8_t *pins, uint16_t value) {

  for (int8_t i = 0; i < 8; i++) {
    digitalWrite(pins[i], value & (1 << i)) ;
  }
}

The digitalWrite function sets an output pin to the given value. This version is bad on the ATmega processors because they don't have multi-bit shift/rotate operations - you have to do them one bit at a time. So the 1 << i turns into an inner loop. The Arduino framework includes bit twiddling functions that would presumable avoid this, but they don't seem to be in common use, and aren't recommended in the language documentation about >> and <<. In fact, that the processors needs a loop to do variable multi-bit shifts isn't mentioned, either.

The second form avoids that, by using a mask variable that it modifies with each pass through the loop:
void
display_value(uint8_t *pins, uint16_t value) {

  for (uint8_t bit = 0x8, i = 0; bit; i++, bit >>= 1) {
    digitalWrite(pins[i], value & bit) ;
  }
}

This version doesn't have the inner loop for the shift, and when disassembled takes about half the instructions of the first version, and a bit less stack space.

The version I wrote as "natural" C changes the interface slightly, using a terminal zero in the array for the last element, then doing pointer arithmetic instead of an explicit index:
void
display_value(uint8_t *pins, uint16_t value) {

  for (; *pins; pins++, value >>= 1) {
    digitalWrite(*pins, value & 1) ;
  }
}

When disassembled, this one is slightly shorter and uses slightly less stack than the second version. However, if it were open coded, it would require another variable, or two, depending on whether or not value could be destroyed.

As I said, this feels natural to me. I learned to do array processing in C with pointer arithmetic. I can see that that might no longer be fashionable - after all, it is error-prone. However, that it allows me to avoid introducing two new variables seems to more than make up for that, as those introduce there own errors.

What I'm wondering is if there's some good reason to avoid the third version for Arduino projects, or modern C in general. Disassembling it didn't turn up any. Anyone else have an opinion?