Does ObjectAda optimize at all??

Does ObjectAda optimize at all??

[dennison]

I tried asking their tech support this, but they told me all the engineers are
"going to a function" today.

I'm looking for information on what optimizations ObjectAda (on PC/WinNT) can
perform. The only parameters their compiler and IDE provide are for turning
off checks (which are my least favorite optimization). There's nothing about
register optimization, loop unrolling, hoisting, etc.

I can't find anything about it in their printed documentation. Their online
help has three references; two for turning off checks and one reference to
annex M where they state that they ignore "pragma optimize".

Since I have never got an "optimized away" message from their debugger, I
figured no optimization was being performed. But surely ObjectAda can do
optimizations, right? Right???

T.E.D.

-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp   Create Your Own Free Member Forum

Re: Does ObjectAda optimize at all??

[Jerry van Dijk]

dennison@telepath.com wrote:

: I'm looking for information on what optimizations ObjectAda (on PC/WinNT) can
: perform. The only parameters their compiler and IDE provide are for turning
: off checks (which are my least favorite optimization). There's nothing about
: register optimization, loop unrolling, hoisting, etc.

There are no documented optimization options in either the IDE of commandline 
compilers.

-- 
-- Jerry van Dijk  | email: jdijk@acm.org
-- Leiden, Holland | member Team-Ada
-- Ada & Win32: http://stad.dsl.nl/~jvandyk

Re: Does ObjectAda optimize at all??

[Tarjei Tj�stheim Jensen]

dennison@telepath.com wrote in message <6qcl1i$qij$1@nnrp1.dejanews.com>...
>I tried asking their tech support this, but they told me all the engineers are
>"going to a function" today.
>
>I'm looking for information on what optimizations ObjectAda (on PC/WinNT) can
>perform. The only parameters their compiler and IDE provide are for turning
>off checks (which are my least favorite optimization). There's nothing about
>register optimization, loop unrolling, hoisting, etc.
>
>I can't find anything about it in their printed documentation. Their online
>help has three references; two for turning off checks and one reference to
>annex M where they state that they ignore "pragma optimize".


If nobody answers, try getting hold of the strings program and use it on the
compiler. By perusing the output there is a chance you may discover the
neccessary parameters.


Greetings,

Re: Does ObjectAda optimize at all??

[dennison]

In article <ExACx5.9v@jvdsys.nextjk.stuyts.nl>,
  jerry@jvdsys.nextjk.stuyts.nl (Jerry van Dijk) wrote:
> dennison@telepath.com wrote:
>
> : I'm looking for information on what optimizations ObjectAda (on PC/WinNT)
can
> : perform. The only parameters their compiler and IDE provide are for turning
> : off checks (which are my least favorite optimization). There's nothing about
> : register optimization, loop unrolling, hoisting, etc.
>
> There are no documented optimization options in either the IDE of commandline
> compilers.

Well, yes. That's a very good paraphrase of what I said. What I want to know
is are there any optimizations going on at all (and if so, what)? Surely a
supposedly production-quality compiler must do something.

T.E.D.

-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp   Create Your Own Free Member Forum

Re: Does ObjectAda optimize at all??

[Robert Dewar]

T.E.D. said

<<>I can't find anything about it in their printed documentation. Their online
>help has three references; two for turning off checks and one reference to
>annex M where they state that they ignore "pragma optimize".
>>


Note that most compilers ignore pragma Optimize. This "feature" of Ada 83
represented a naive and not very helpful view of what optimization is all
about (*). Remember that the issue is not what optimizations a compiler
performs but what the quality of the generated code is. This is particularly
important for an Ada compiler, where the classical optimizations that you
read about say in the Dragon book are only a small part of the total issues
involved in generating efficient code.

That being said it is of course interesting to ask what optimization 
flexibility is provided, and most certainly a compiler with no control
over optimization levels is going to be unsuitable for some kinds of
applications (e.g. safety critical applications often want a minimum
of optimization to improve tracability of code for certification
purposes, whereas obviously other applications want the fastest
code possible and are most certainly willing to sacrifice 
object to source tracability.

Note incidentally that GNAT in addition to the -O0/1/2/3 levels that specify
general optimization control (and the related -gnatn and -gnatN switches)
also recognizes all the (rather large set of) specific optimization control
switches provided by gcc (e.g. -funroll-loops and -funroll-all-loops). I
would guess that most of these switches get VERY little use, but who knows?

(*) What GNAT does with pragma Optimize is warn you if you use this pragma
and then compile without optimization, but it is much too late for the
pragma to actually affect optimization, since the GCC driver needs to know
optimization levels and requirements *before* it starts compiling the code.

Robert Dewar
Ada Core Technologies

Re: Does ObjectAda optimize at all??

[Dave Wood]

[Please reply via email, as I rarely have a
chance to check this group any more.]

It depends on which ObjectAda we're talking
about.  Generally, the real-time/embedded/safety 
critical versions have a lot more in the way of 
compiler optimization switches than does
the Windows native.  For that matter, so does
the Unix native.  This is true primarily for
practical and logistical reasons.  OA/Win native
doesn't give you a lot of options, relying
on the combination of some hidden front-end
optimizations, an inherently good code 
generator (though getting long in the teeth),
and build-time stuff like unused subprogram
elimination.

Optimization options for the sake of optimization
options aren't high on the list for that product,
as (by and large) the users of that product are
longing for other kinds of improvements and we
need to prioritize accordingly.

At some point, we'll roll out a brand new code
generator that will allow a longer list of back-end
optimizations (which will have importance for
embedded Intel targets, but I'm doubtful it will
provide much more than psychological effect for
Win native where there are so many other factors
at play), but for now we find the most
bang for the buck for the typical native user 
by improving the performance of the runtime.  
For example, the new 7.1.2 release (which 
will be shipping in 3-4 weeks) has improved 
performance of some benchmarks at the level of 
an order of magnitude, just" by optimizing the 
runtime.

-- Dave Wood, Aonix
-- Product Manager, ObjectAda for Windows

p.s.  I don't know what "function" the engineers
were at when you called, since that implies they
have a real life but we like to avoid that by keeping
them busy round-the-clock.  :)  But, I might suggest
you join our user group mailing list.  It's a very
useful information source for our customers and more
likely to be tracked by our engineers than c.l.a.
You'll find instructions for joining at:

http://www.aonix.com/Support/ada_index.html

Re: Does ObjectAda optimize at all??

[dennison]

In article <35CE22D9.113EAEEC@cts.com>,
  Dave Wood <dpw@cts.com> wrote:
> [Please reply via email, as I rarely have a
> chance to check this group any more.]

Unfortunately, Dejanews does not give me that option, so its a bit difficult
to do. I'll hack it my posting this w/ "send me a copy", and forwarding the
copy to you.

> practical and logistical reasons.  OA/Win native
> doesn't give you a lot of options, relying
> on the combination of some hidden front-end
> optimizations, an inherently good code
> generator (though getting long in the teeth),
> and build-time stuff like unused subprogram
> elimination.

I noticed. What I'd like to know is what it *is* doing to optimize things? The
thing that got me curious was that I never got any "optimized away" messages
when I tried to look at things in the debugger. I don't know much about i386
instructions, but I do know that the Pentium has a base of RISC instructions.
You have to wonder about the efficiency of any compiler that doesn't optimize
variables away to registers on a RISC processor.

> Optimization options for the sake of optimization
> options aren't high on the list for that product,
> as (by and large) the users of that product are
> longing for other kinds of improvements and we
> need to prioritize accordingly.

When I called tech support, he said he never had a question about
optimizations before. I found that rather shocking, but if its true then you
are obviously right to concentrate on other things.

I suppose the typical response to slow code on a Wintel platform is "buy a
faster box".

> p.s.  I don't know what "function" the engineers
> were at when you called, since that implies they
> have a real life but we like to avoid that by keeping
> them busy round-the-clock.  :)  But, I might suggest

Cool. I was getting jealous. :-)

> you join our user group mailing list.  It's a very
> useful information source for our customers and more
> likely to be tracked by our engineers than c.l.a.

I'll look into it. Thanks.

-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp   Create Your Own Free Member Forum

Re: Does ObjectAda optimize at all??

[Tom Moran]

There is (was?) a suite of Ada benchmarking programs for comparing
compilers.  Have you tried that?  I presume what you really care about
is the size/speed of the generated code, not which optimizations it
does/does not use.

Re: Does ObjectAda optimize at all??

[dennison]

In article <35d4797b.244364@SantaClara01.news.InterNex.Net>,
  tmoran@bix.com (Tom Moran) wrote:
> There is (was?) a suite of Ada benchmarking programs for comparing
> compilers.  Have you tried that?  I presume what you really care about

I'm intrigued. Where?

> is the size/speed of the generated code, not which optimizations it
> does/does not use.
>

Speed, yes. However, the actual optimizations performed *are* a concern. Many
developers have an unfortunate tendency to make complete trash out of their
source in an attempt to "optimize" it. (Manual loop unrolling, huge procedures
with repeated code, reused variables, global variables, etc.) If I know what
our compiler can optimize, I can stop these folks from making an unreadable
mess of thier source code for the sake of speed.

However, if ObjectAda doesn't do variable optimization, loop unrolling,
subprogram inlining, etc., then these folks are right to do this, even though
they drasticly worsen the maintainability of the code. That is why I think it
is a disservice to your customers to not provide optimizations, even if they
never specificly ask for them.

T.E.D.

-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp   Create Your Own Free Member Forum

Re: Does ObjectAda optimize at all??

[Steve Doiel]

Tom Moran wrote in message
<35d4d400.23421382@SantaClara01.news.InterNex.Net>...
>>making an unreadable
>>mess of thier source code for the sake of speed.
>Someone said "Premature optimization is the root of all evil."

In "The Practical Guide to Structured Systems Design" by Meiler Page - Jones
they have a chapter on optimization.  I work by the philosophy presented
there:

It's easier to get a working system efficient than it is to get an efficient
system working.

Simplicity is a virtue that bears it's own rewards.

Optimize a system only if it fails a performance test.

Optimize only the parts of the system worth optimizing (10% of the code runs
90% of the time).


It has been my experience that making a clean system that is not heavily
optimized often reveals high level changes that lead to greater gains than
low level optimizations.  BTW: I am one of those programmers who has had to
go in and try to get a system running that was too heavily optimized.  I've
even seen heavily optimized code that is only run once at boot time on a
system that is rarely restarted.

FWIW
SteveD

Re: Does ObjectAda optimize at all??

[Tom Moran]

>making an unreadable
>mess of thier source code for the sake of speed.
Someone said "Premature optimization is the root of all evil." 

Re: Does ObjectAda optimize at all??

[Larry Kilgallen]

In article <6r2a41$prp$1@nnrp1.dejanews.com>, dennison@telepath.com writes:

> However, if ObjectAda doesn't do variable optimization, loop unrolling,
> subprogram inlining, etc., then these folks are right to do this, even though
> they drasticly worsen the maintainability of the code. That is why I think it
> is a disservice to your customers to not provide optimizations, even if they
> never specificly ask for them.

That argument is somewhat weakened in the case at hand, where the
ObjectAda customer base _has_ been asking for many other things,
faster than Aonix can deliver.  To make a judgement that what the
customers have been asking for is "less important" than any other
particular addition gets very subjective.

Larry Kilgallen

Re: Does ObjectAda optimize at all??

[Robert Dewar]

T.E.D. says

<<However, if ObjectAda doesn't do variable optimization, loop unrolling,
subprogram inlining, etc., then these folks are right to do this, even though
they drasticly worsen the maintainability of the code. That is why I think it
is a disservice to your customers to not provide optimizations, even if they
never specificly ask for them.
>>

In my experience, programmers often do not know enough to know whether
such manual optimizations help or hurt. They often end up not only
trashing the code, but also making it run nmore slowly. For example,
it takes a careful analysis of icache usage to know whether loop
unrolling is helpful.

At the very least, you shoudl only indulge in this kind of low level
optimziation of your coded if (a) profiling shows that there is a
potential for improvement and (b) testing shows the change is in fact helpful

Re: Does ObjectAda optimize at all??

[Tom Moran]

>> There is (was?) a suite of Ada benchmarking programs for comparing
>> compilers.  Have you tried that?  I presume what you really care about

>I'm intrigued. Where?
Look at http://info.acm.org/sigada/WG/piwg/aces_20.html
(and let us know your findings!).

Re: Does ObjectAda optimize at all??

[dennison]

In article <1998Aug14.222830.1@eisner>,
  Kilgallen@eisner.decus.org.nospam wrote:
> In article <6r2a41$prp$1@nnrp1.dejanews.com>, dennison@telepath.com writes:
>
> > However, if ObjectAda doesn't do variable optimization, loop unrolling,
> > subprogram inlining, etc., then these folks are right to do this, even
though
> > they drasticly worsen the maintainability of the code. That is why I think
it
> > is a disservice to your customers to not provide optimizations, even if they
> > never specificly ask for them.
>
> That argument is somewhat weakened in the case at hand, where the
> ObjectAda customer base _has_ been asking for many other things,
> faster than Aonix can deliver.  To make a judgement that what the
> customers have been asking for is "less important" than any other
> particular addition gets very subjective.

True. I'm talking ideal la-la-land, rather than the real world. There are lots
of things that "ought" to be in a compiler. You have a finite amount of effort
to put into it, and thus even important things must be prioritized. I
understand that.

But the message I was getting was that optimizations are unimportant because
no- one has specifically asked for them. That is *not* true.

Unfortunatly, I now seem to be sharing an argument with the GC folks. Most
unfortunate. Perhaps I'd best bow out while I'm behind. :-)

T.E.D.

-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp   Create Your Own Free Member Forum

Re: Does ObjectAda optimize at all??

[dennison]

In article <dewar.903282165@merv>,
  dewar@merv.cs.nyu.edu (Robert Dewar) wrote:
> In my experience, programmers often do not know enough to know whether
> such manual optimizations help or hurt. They often end up not only
> trashing the code, but also making it run nmore slowly. For example,
> it takes a careful analysis of icache usage to know whether loop
> unrolling is helpful.

Exactly. I think I'm familiar enough with my own ignorance to aviod this
personally. But its awfully hard to stop other folks from doing this in their
own code when they know:

  o  The code in question is called several times from a high-frequency
routine  o  The complier is unable to do the optimization for them.  o	If
they manually "optimize" the code later, they are liable to  accidentally
introduce new bugs (at a very late stage in development).

Thus I think it vital that developers have some knowledge of and faith in the
optimization abilities of their compiler. If a compiler doesn't do much in the
way of optimization, it clearly falls down in this respect.

T.E.D.

-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp   Create Your Own Free Member Forum

Re: Does ObjectAda optimize at all??

[dewar]

T.E.D. says

<<Exactly. I think I'm familiar enough with my own ignorance to aviod this 
personally. But its awfully hard to stop other folks from doing this in their
 own code when they know:

  o  The code in question is called several times from a high-frequency
  routine  o  The complier is unable to do the optimization for them.  o  If
  they manually "optimize" the code later, they are liable to  accidentally
  introduce new bugs (at a very late stage in development). >>


This is missing the most critical criterion, namely that it will actually
make a difference to do the transformation. As I say, often programmers
just don't know it will make a difference. For example, loop unrolling
may or may not help.

A programmer who tries to improve their program with inadequate knowledge
is a completely in the dark. Yes incompetent programmers might do anything
(you always seem to be worrying about them), but there is no reason to think
that knowing anything about the compiler will in positively affect their
behavior if they are this incompetent!

-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp   Create Your Own Free Member Forum

Optimizations' effect on software quality (was: Does ObjectAda optimize at all??)

[dennison]

In article <6rngqo$dep$1@nnrp1.dejanews.com>,
  dewar@gnat.com wrote:

> is a completely in the dark. Yes incompetent programmers might do anything
> (you always seem to be worrying about them), but there is no reason to think
> that knowing anything about the compiler will in positively affect their
> behavior if they are this incompetent!

I'm referring to real software developers that I deal with on a daily basis.
Some of them are neophites, yes. But everyone here was at one point as
well. As a matter of fact, if we take a definition of competence that includes
"knowing not to try to hand-optimize code up front without making sure that it
actually significantly helps", we would label "incompetent" nearly every
software engineer I have ever worked with. I think that's a bit extreme.

If you work with nothing but expert engineers, you should consider yourself
fortunate. I have been on a couple of projects like that, and they are a joy
to work. But generally I deal with a lot of engineers who are new to Ada and
haven't spent a lot of time learning what I would consider sound software
engineering principles. You may be able to dismiss such folks as
"incompetent", but I don't have that luxury.

Its my responsibility to help such people along, with the poor powers of
logic and persuasion that I have at my disposal. For this particular issue,
I'm used to using they compiler's optimization features to strengthen my
argument. My own ameteur psycology theory is: Yes, the real point is that you
shouldn't blindly perform "optimizations" that might not help. While the
logic of that is generally convincing, accepting that when they have already
started hand-optimizing code means accepting that they aren't superior
engineers, which they have a lot of trouble doing. However if they can also
tell themselves, "well, the compiler can do that for me later", then it
becomes an issue of saving them work. That's much more palatable. Consider it
the sugar that helps the medicine to go down.

But there's a bit more to it that just that. Suppose I know my compiler
doesn't do subprogram inlining at all. Am I going to break everything up into
lots of tiny modular one-use routines? Certianly not to the extent I would
have otherwise. If I know that my compiler is incapable of moving
loop-invariant code outside of a loop, I'm probably not going to put any
(that I notice) inside a loop, even if it would have made the source code a
little bit clearer and less sensitive to change.

So lack of optimizations in the compiler will have a real effect on the
readability and maintainability of the resultant code. The effect will
probably vary by developer. It could be striking, or it could be subtle. But
it will be there.

--
T.E.D.

-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/rg_mkgrp.xp   Create Your Own Free Member Forum