Re: Help Me Please :)

[r_c_chapman@my-deja.com]

> Certain types of optimization assist in this goal. For example,
> with gnat, we usually find that -O1 assembly code is far easier
> to follow than -O0 code, simply because there is so much less
> of it.

I concur - results from our paper in Ada Europe 99 confirm this.

Brief summary:

On the original SHOLIS project, we had SPARK83 targetting 68040,
and we found the optimizer made very little difference to run-time
and/or object-code understandability, but the Alsys compiler
has hand-written code selection which makes the "unoptimized" code
quite reasonable in those terms.

I ported SHOLIS to SPARK95, and used GNAT/68k to recompile.
At -O0, GNAT produces more code than Alsys, and it's actually
harder to understand, since there are so many redundant loads
and stored littered about the place.  At -O1, the GNAT code
is far smaller and easier reading.  This is no surprise, since
GCC's code selection is table-driven.

Interestingly, we did not do full-scale object-code verification
(OCV) on SHOLIS, owing to the long track-record and stability
of the Alsys compiler.  (well..and the simplicity of SPARK...)

On more recent targets (e.g. PowerPC), it's really important
to have _some_ optimizations on (particularly decent
register allocation) since the difference between
a register, a cache hit, and a main-memory access is so
much more pronounced on these targets.  We tried
running code from VADSCross/PowerPC with -O0 on another project -
the results were disappointing, so we reverted to the _default_
level of -O4 for that compiler, which proved fine (this was also
a SIL4 project.)

High-level optimization that perform large-scale reorganisation
of code are definitely a no-no, but I can live with local improvements.

Overall, I recommend you simply avoid having to do full-on
OCV - it's very hard work!
 - Rod Chapman
   Praxis Critical Systems
   rod@praxis-cs.co.uk




Sent via Deja.com http://www.deja.com/
Before you buy.