Re: [PATCH 1/5] compat/bswap.h: Fix build on cygwin, MinGW and msvc

From: SZEDER Gábor <szeder@ira.uka.de>
Date: Nov 8, 2013, 00:45 UTC
Message-ID: <20131108004550.GA16843@goldbirke>
In-Reply-To: <527C0D4A.7070101@ramsay1.demon.co.uk>

Hi,

On Thu, Nov 07, 2013 at 09:59:38PM +0000, Ramsay Jones wrote:

Show 11 quoted lines

> +static inline uint64_t default_bswap64(uint64_t val)
> +{
> +	return (((val & (uint64_t)0x00000000000000ffULL) << 56) |
> +		((val & (uint64_t)0x000000000000ff00ULL) << 40) |
> +		((val & (uint64_t)0x0000000000ff0000ULL) << 24) |
> +		((val & (uint64_t)0x00000000ff000000ULL) <<  8) |
> +		((val & (uint64_t)0x000000ff00000000ULL) >>  8) |
> +		((val & (uint64_t)0x0000ff0000000000ULL) >> 24) |
> +		((val & (uint64_t)0x00ff000000000000ULL) >> 40) |
> +		((val & (uint64_t)0xff00000000000000ULL) >> 56));
> +}

This got me thinking. To swap 8 bytes this function performs 8 bitwise shifts, 8 bitwise ANDs and 7 bitwise ORs plus uses 8 64bit constants. We could do better than that:

static inline uint64_t hacked_bswap64(uint64_t val)
{
	uint64_t tmp = val << 32 | val >> 32;
	return (((tmp & (uint64_t)0xff000000ff000000ULL) >> 24) |
		((tmp & (uint64_t)0x00ff000000ff0000ULL) >>  8) |
		((tmp & (uint64_t)0x0000ff000000ff00ULL) <<  8) |
		((tmp & (uint64_t)0x000000ff000000ffULL) << 24));
}

This performs only 6 shifts, 4 ANDs, 4 ORs and uses 4 64bit constants.

bswap64ing 1000000000 64bit ints with default_bswap64() compiled with -O2 takes:

  real    0m1.808s
  user    0m1.796s
  sys     0m0.000s

The same with hacked_bswap64():

  real    0m0.823s
  user    0m0.816s
  sys     0m0.000s

I doubt that in normal usage git would spend enough time bswap64ing to make this noticeable, but it was a fun micro-optimization on a wet Thursday evening nevertheless :)

Best, Gábor