To: djgpp AT delorie DOT com
Subject: Re: fixed point math: NEED HELP QUICK!
Message-ID: <19970131.205344.8463.2.chambersb@juno.com>
References: <Pine DOT LNX DOT 3 DOT 95 DOT 970131164737 DOT 140G-100000 AT capslock DOT com>
From: chambersb AT juno DOT com (Benjamin D Chambers)
Date: Fri, 31 Jan 1997 23:52:33 EST


On Fri, 31 Jan 1997 16:52:03 -0500 (EST) mharris AT blackwidow DOT saultc DOT on DOT ca
writes:
>On Tue, 28 Jan 1997, Benjamin D Chambers wrote:
>
>> 
>> On Tue, 28 Jan 1997 20:30:34 -0500 DJ Delorie <dj AT delorie DOT com> 
>writes:
>> >
>> >> function.  Anyway, here's my problem;  To my understanding the 
>> >following
>> >> are 100% equal:
>> >> 
>> >> other_stuff = stuff / 256;
>> >> other_stuff = stuff << 8;
>> >
>> >The second multiplies.  Try ">> 8" instead.
>> >
>> >
>> Also, just to be clear, the first gets translated into the second 
>when
>> compiling with optimizations on (-O2).
>> In fact, I believe *ALL* integer multiplication/division by a 
>constant
>> gets converted to shifts when using -O2, so you might as well use 
>the
>> first - if for nothing else then for readability's sake.
>
>If *ALL* integer mults/divs are converted to shifts, then what would
>be the equivalent for:
>
>other_stuff = stuff * 57
>other_stuff = stuff << ???
>
>The point I'm trying to make is that no, only mul/div's by powers of 2
>would be converted to shifts.  Others such as multiplying by numbers
>whose factors are of powers of 2 can also be optimized in this way.
>Numbers that don't fit (such as 57) can't be implemented via shifting
>and be optimal.
Uh, yeah?
Write 57 in binary...
111001
On a 486, that's 3 shifts and 3 adds, that's 9 cycles... Can you make a
mul do that?
On a pentium, you should be able to do that in 4 cycles (with 2 stalls on
the V-pipe, so throw in some more code...).  That don't beat the pants
off an integer mul?

The catch is that you _HAVE_ to know the multiplier at compile time, so
it can only be done sometimes.

...Chambers