Re: Code size of Ada ?

["Robert I. Eachus" <eachus@mitre.org>]

Robert Dewar wrote:
 
> No, I was not talking about short sequences here. I have written
> many large programs in assembly in the past (probably I have
> written about half a million lines of delivered commercial
> assembly code). For an example of a fairly large program
> written in very aggressive assembly language throughout, about
> 20K SLOC, look at the 360 implemnetation of SPITBOL (note that
> is a COMPLETE, very fast SNOBOL-4 compiler which fits in about
> 80K bytes of code -- it compiles at many millions of lines
> a minute, using the TRT instruction extensively for lexical
> analysis).
> 
> But I certainly do NOT recommend this style of programming,
> an ASM program of this size, written in this manner, is barely
> maintainable, and the number of people who can work with such
> programs is minimal.
> 
> I think it is important to understand that using higher level
> languages does have a cost. The objective is to keep the cost
> at a reasonable level (many current Java implementations fail
> to meet this criterion). With current machines, we can afford
> to give up a small factor of inefficiency, but not a large
> factor :-)

    I've written programs of that size in assembler as well, although
the limiting factor used to be memory size.  In fact I go back far
enough that I
started writing in machine code.  I have a copy of the first assembler
manual (GP--Generalized Programming--for the Univac I) in my attic, and
it was a big step forward.

    My point was that programming that results in good generated code
requires knowing exactly what constitutes good machine code, and that is
a relatively rare skill these days.  My style now, when I need this
degree of efficiency is to write in a high level language and then fix
any bogus code either in the compiler optimizer, by reorganizing the
high-level code so that the compiler does find the correct idiom, or by
writing some routines in assembler.  For example, on one Ada 83 project,
writing a seventeen instruction loop in a 25 KSLOC program as a code
insert improved performance by 25%.  Or for another example, in the
random number generator I wrote, on most hardware computing U squared
mod N, for U and N integer, can be done in two instructions (excluding
moving the data into registers) even though the intermediate value is
double word size. Writing this as a code insert can speed things up
dramatically because a lot of the computational tricks used are no
longer needed.

-- 

                                        Robert I. Eachus

with Standard_Disclaimer;
use  Standard_Disclaimer;
function Message (Text: in Clever_Ideas) return Better_Ideas is...