Efficient C for ARM

Local Variable Types

The ARM data processing operations always operate on 32-bit quantities. You should therefore:

  • Use a 32-bit data type (e.g. int) for local variables.
  • Avoid char and short for local variables, even if you’re manipulating a char or short value.

The exception to this is when you require wrap-around or modulo arithmetic (e.g. 255+1 = 0).

Examples

This example code calculates a simple checksum on a packet of 64 words:

int checksum1(const int *data)
{
  char i;
  int  sum = 0;

  for (i = 0; i < 64; i++)
    sum += data[i];

  return sum;
}

Let’s look at the annotated compiler output:

checksum1
        MOV     r2,r0             ; R2 = data
        MOV     r0,#0             ; sum = 0
        MOV     r1,#0             ; i = 0
loop
        LDR     r3,[r2,r1,LSL #2] ; R3 = data[i]
        ADD     r1,r1,#1          ; R1 = i+1
        AND     r1,r1,#0xff       ; i = (char)R1   *** UNNECESSARY ***
        CMP     r1,#0x40          ; compare i to 64
        ADD     r0,r3,r0          ; sum += R3
        BCC     loop              ; if (i<64) loop
        MOV     pc,r14            ; return sum

The compiler is emitting an AND r1,r1,#0xff instruction even though it should know that i never exceeds 64. If we change i from char to unsigned int the AND disappears: it’s no longer necessary to account for wrap-around.

Remember that this isn’t just a saving of one instruction or cycle. It saves 64 instructions: one for each iteration.

This is an inner loop. Optimisations to inner loops are highly beneficial.

Remarks

You might think char is an efficient choice for i; using less stack space or register space than an int might. On the ARM, this is wrong:

  • Stack entries are at least 32 bits wide.
  • Registers are 32 bits wide.

To compute the modification of i correctly the compiler must account for the case where i will wrap around which you get for ‘free’ with ints, but not with chars.

Navigate