Some ideas / notes about printf HOT 8 OPEN

I have another proposition of optimization.
In _ntoa_long and _ntoa_long_long you can make different processing path based on base of number. If its base is power of 2 you can replace division and modulo operation with simple bitwise right shift and and operation. It will improve performance a lot on most microcontrolers and even more on those without hardware division support.
Another optimization (in those functions) is to use look up table while obtaining if output should be a digit or a letter. It will reduce the if statement and all folowing calculations with single memory access.
I propose something like this:

const char LUT[] = "0123456789ABCDEF";
const char lut[] = "0123456789abcdef";
// internal itoa for 'long' type
static size_t _ntoa_long(out_fct_type out, char* buffer, size_t idx, size_t maxlen, unsigned long value, bool negative, unsigned long base, unsigned int prec, unsigned int width, unsigned int flags)
{
  char buf[PRINTF_NTOA_BUFFER_SIZE];
  size_t len = 0U;

  // no hash for 0 values
  if (!value) 
  {
    flags &= ~FLAGS_HASH;
  }

  // write if precision != 0 and value is != 0
  if (!(flags & FLAGS_PRECISION) || value) 
  {
    if(base == 10)
    {
      do
      {
        const char digit = (char) (value % base);
        buf[len++] = LUT[digit];
        value /= base;
      } while(value && (len < PRINTF_NTOA_BUFFER_SIZE));
    }
    else
    {
      char* lutUsed = (flags & FLAGS_UPPERCASE ? LUT : lut);
      unsigned long baseBits = (base == 16 ? 4 : (base == 2 ? 2 : 3));
      do
      {
        const char digit = (char) (value & (base - 1));
        buf[len++] = lutUsed[digit];
        value >>= baseBits;
      } while(value && (len < PRINTF_NTOA_BUFFER_SIZE));
    }
  }

  return _ntoa_format(out, buffer, idx, maxlen, buf, len, negative, (unsigned int)base, prec, width, flags);
}

from printf.

from printf.

(I am just a downstream observer, not a participant.)

I recognize those kinds of ARM optimizations, and they sound good. This is great.

However I’d also like to encourage you to optimize for code size as well. And I can help if you like.

from printf.

Hi @ledvinap , thanks a lot for your detailed investigation and the according results. Amazing!!
I´m on Easter holiday in the moment and have a look into all issues when returning home.
Happy Easter!

from printf.

Hi @ledvinap, I am presuming this means that out_fct_type would be removed and only out_object.fn would remain? This makes sense because fctprintf is the most general overload and all other variants can pass through the same struct, so there should be no need to have two separate function pointers.

from printf.

That would actually be a good idea for the the base-10 branch also. I would perhaps explicitly write value % 10 instead of value % base, although I believe most optimizing compilers will understand that base is 10 at that point.

But gcc and clang will use multiplication and shifting instead of division whenever the divisor is a known integer (example here, the latter calls __aeabi_idivmod).

from printf.

@ledvinap : Hello Petr,

I think these different suggestions should be split into separate issues; and for each issue, a PR. I realize this repository hasn't seen much traction, but still.

from printf.

@legier : See also issue #116 .

from printf.