Re: [Algorithms] Shifting or Multiplication....

[Conor S. <cs...@tp...>]

>On plain Pentiums, shifting was 0.5 cycle in the U pipe only (so there was
a 0.5 cycle latency if in the V pipe). Integer multiplications were about
~10 cycles. Nowadays, even if rules have changed, you shouldn't discard
bitshifts :

I believe that it was actually 3 cycles. But you could execute 6 adds in the
time to do 1 mul. But you are pretty much right.

I think the LEA takes the same time - But there was some advantage or other
to it :)

Conor Stokes

1) you never know where your program will run in the end
2) using a bitshift for masks and similar things is conceptually clearer
anyway. Keep your code clean!
3) you probably code in C/C++, and most of the time the compiler handles
this for you anyway (*)
4) there's more to optimization than this :)
5) you can't do the same as SHLD / SHRD with muls, ah!

(*) As a reminder, even "slow" things like:

    int Value;
    ...
    Value*=5;

...are cleverly compiled using the good old LEA instruction (that is:
without shifts or muls). Nowadays I would leave all those things to the
compiler.


Pierre Terdiman      *   Home: p.t...@wa...
Coder in the dark    *   Zappy's Lair:  www.codercorner.com


----- Original Message -----
From: Alex Morano
To: gda...@li...
Sent: Saturday, December 23, 2000 9:38 PM
Subject: [Algorithms] Shifting or Multiplication....


    I was wondering if anyone could clear up something. Algorithmically, is
bitshifting or using multiplication faster?? Now, I know this sounds like a
subjective question, as some will undoubtadly say, it depends on the
compiler and what it translates the code into. Lets take it down to the ASM
level.

    Shifting, to me, looks like a 1 cycle operation.

    In this age, if the CPU can handle most MUL's almost as rapidly, is
there a place for bitshifts? If so, then for what kinds of algo
applications? Do we still need them to mask off bits in pixels colors?