Re: WinNT ADA compilers comparison

["Thierry Lelegard" <lelegard@club-internet.fr>]

> From the user point of view, I think that reliability comes first 

Sure. Performance comes after reliablity.

> More than that, increase performance by 5 doesn't mean a lot.
> After a user action :
> he will find not acceptable the difference between 1s and 5s
> he will not care or even notice the difference between 1ms and 5ms
> (And I saw some developpers spending days for the second improvement, while
> not even care about the first ..., usually the user is not very happy)

There are two different points here:
1. The way the application (source code) is written
2. The way the compiler generates the binary code.

Spending days optimizing the less relevant part of an application
is the first point. I agree that this is useless.

But, on a code generator point a view, there is no difference
in optimizing from 5s to 1s and from 5ms to 1ms. The first case
is most of the time (roughly) a loop of 1000 iterations of the
second case in the source code.

The quality and speed of the generated code is an instrinsic
quality of a compiler. To address the full range of applications,
a compiler must make it right. Let me briefly describe two cases
I experienced where it made all the difference:

- Years ago, I spent two days optimizing a few ms in a matrix
calculation. Why? It was in a Sonar software and the extra
ms missed the next hit.

- Currently, we sell software products to customers. These
software process huge amount of data and we must produce
new versions on a regular basis. Our customers have invested
a lot of money in large computers which are consequently
accurately sized for our software. If the generated code
becomes poor, then the customers must invest in larger
systems, which is not acceptable on a commercial point
of view.

So, yes, the performance of the generated code is not
important for hobbyists, educational institutions, GUI
software, desktop tools, etc. But, for commercial 
industrial software products, it is essential.

> I know few developpers linking just once a day.
> For us, the difference between 20s and 4mns for linking is more than
> important (we actually saw that difference between some compilers for
> same code and executable)

I agree. But it is also true that many Ada tools (including
compilers) are written in Ada. So, improving the compile time
is simply a nice side effect of improving the generated code
performance.

BTW, I also noticed the same magnitude of difference
between compilers, maybe we used the same two...

-Thierry
________________________________________________________
Thierry Lelegard, Paris, France
E-mail: lelegard@club-internet.fr