Re: [PATCH 1/1] vsprintf: optimize decimal conversion (again)

[Denys Vlasenko <vda.linux@googlemail.com>]

On Tuesday 27 March 2012 01:18, Denys Vlasenko wrote:
> > I don't think Linux runs on anything with BITS_PER_LONG_LONG != 64...
> > 
> > BTW, what about CPUs with slow 32x32 multiplication and/or slow 64-bit
> > division?
> 
> Without 32x32->64 multiply, the best we can generate is 4 decimal digits:
> we produce next digit by approximating x/10 with (x * 0xcccd) >> 19,
> and the first x where it gives wrong result is 81920 if multiply result
> is truncated to 32 bits.
> With it, we can generate 9 digits using (x * 0x1999999a) >> 32.
> 
> Regrading "slow 64-bit division" - after this patch, 32-bit machines
> wouldn't use it at all. Only 64-bit machines will perform 64-bit
> division, one per 9 decimal digits (thus, at most three divisions
> per one long_long->string conversion).
> 
> In fact, with small change to #ifdefs, all machines with long long <= 64
> bits can use division-less routine. It might be a good thing to try...

Well, apparently it's not a good idea for my Phenom II in 64-bit mode:
It's bigger:

   text    data     bss     dec     hex filename
   2395     448       0    2843     b1b test_div64.o
   2507     448       0    2955     b8b test_nodiv.o

And slower:
Conversions per second:
div64: 8:42660000 123:31472000 123456:21748000 12345678:19140000 123456789:20980000 2^32:16948000 2^64:9480000
nodiv: 8:40532000 123:30276000 123456:21172000 12345678:18672000 123456789:13440000 2^32:13440000 2^64:8992000

-- 
vda