Re: Wrong RTL instruction deleted

[Robert Dewar <dewar at adacore dot com>]

Richard Kenner wrote:

> (1) Remaining bugs in setting alias sets.
> (2) Amount of optimization lost if everything is alias set 0.
> (3) Detecting and warning about dubious aliasing cases.
>
> Personally, I think (3) is key. 

I disagree, (2) is the starting point. If real applications do not
show a significant loss in performance, then I would favor correct
functioning of existing code over minor performance gains.

For years I have followed an important rule on compiler optimization
which is "all optimizations are disappointing". With few exceptions,
implementors are often convinced that the wonderful optimizations
they do will make a real big difference, and at best they only make
a small difference.

The lack of data is always frustrating. So far I am comparing real
data from a real large scale user vs Richard's intuitive guesses.
I prefer the former to base my judgmnents on.

I would welcome some data that showed that Tree SSA produced better
code in general, and in particular better code resulting from
better aliasing analysis, because in theory Ada should indeed be
able to make much better use of this. But lacking such data, I
remain skeptical.