🚀 go-pugleaf

In article <1195@aw.sei.cmu.edu> firth@sei.cmu.edu (Robert Firth) writes:
>
>Postscript
>----------
>
>Nearly all machines have trouble with arrays of things smaller than
>the addressing unit.  And many of them have trouble with arrays that
>aren't some special multiple of the addressing unit.  In high-level
>languages, such arrays are pretty common, and the idea that byte
>and word arrays are special cases is obsolescent.

One of the most fundamental computer engineering principles is efficient 
implementation of the most common cases as special cases when there's a suf-
ficient performance win to be had, and this certainly applies here.

>
>It seems to me about time the hardware designers built the right
>indexing primitive, which takes
>
>	(array root, component index, component size)
>
>and where the component size can be anything from 1 bit to HUGE, and 
>the hardware does the scaling as appropriate, either multiplying or
>dividing the index.

Are we in a time warp back to the mid-70s?  This flies in the face of all the
research that led us to RISC.  The standard argument against this kind of
thing is:
1. For a given amount of hardware, it'll be slower than simpler indexing.
2. For a given speed, it'll need more hardware.
3. If you build the fastest machine you can with this indexing scheme, it will
   be slower than the fastest machine you can build with simple indexing.
4. If you special case the first couple of sigmas via normal byte or word
   addressing, you're left with so few cases it no longer makes sense to
   waste the hardware on it instead of generating special code.  

Russell Williams
..{ucbvax!sun,lll-lcc!styx,altos86,bridge2}!elxsi!rw