Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!think.com!paperboy!hsdndev!cmcl2!lanl!cochiti.lanl.gov!jlg From: jlg@cochiti.lanl.gov (Jim Giles) Newsgroups: comp.arch Subject: Re: new instructions Message-ID: <24343@lanl.gov> Date: 23 May 91 16:11:27 GMT References: <1991May22.001620.751@craycos.com> <1991May23.084258.5062@kithrup.COM> <24216@lanl.gov> Sender: news@lanl.gov Organization: Los Alamos National Laboratory Lines: 22 In article , peter@ficc.ferranti.com (peter da silva) writes: |> [...] |> But the question isn't whether you can do these operations in software |> anywhere near as fast as they can be done in hardware, but rather whether |> they can be done fast enough. That is, how much faster would applications |> that make heavy use of these instructions run if you sped them up by a |> factor of 10 (which is about what I would have guessed the speed difference |> to be)? If you speed up 3% of the program by a factor of 10, it's pretty |> much lost in the noise. Exactly so. The only thing I was saying is that the software mechanisms are _not_ done "quite quickly" compared to a hardware implementation of the same thing. Whether is is worthwhile to put an instruction into the hardware is dependent on two things: 1) the commonality of the operation, and 2) the amount by which the hardware is faster. Sean Fagan was claiming the the speedup was insignificant - it isn't. Now, if you want to claim the instruction is rarely needed, that's a completely different issue. It also depends on two things: a) what is the application domain, and b) what features _are_ present in the hardware (individual instructions cannot be selected in isolation from the rest of the instruction set). J. Giles