Like floating point, division is also emulated.

It’s implemented by two runtime library functions. In ARM’s tools they’re called:

• __rt_sdiv for signed values.
• __rt_udiv for unsigned values.

Slow:

• 10-100× slower than equivalent shift operations (20-140 cycles) depending on the values.

It’s not always so bad for divisions by a constant:

• Often optimised away by the compiler.
• If you divide by a power of two the compiler may be able to emit a shift instead.

Possible solutions:

• Multiply by a fixed-point reciprocal then shift.
• Avoid where possible by algebraic substitution.
  • e.g. rewrite (x ÷ y) > z as x > (z × y)
• Use unsigned operands to avoid sign fix-up operations.
• Avoid modulo with a wrap-around counter.
• Use lookup tables for small problems.

Modulus is implemented by the same routine as division.

• The routine produces both results simultaneously.
• One or the other can be computed for free.