Re: Loop Optimizer Deficiency revisited

[Zack Weinberg <zack at codesourcery dot com>]

"Nitin Gupta" <ngupta at GlobespanVirata dot com> writes:

> Hi,
> I'm experimenting with sparlet build of 
> GCC 3.2 released. I could reproduce the first
> problem mentioned under section "More pathetic 
> failures of loop optimization" in 
> http://gcc.gnu.org/projects/optimize.html#pathetic-loop
> with the following piece of code. 
>
> void   badopt ( unsigned long ) ;
> void badopt(
>	unsigned long l
> ) {
>
>	long ad ;
>	long  i ;
>
> 	ad = l;
> 
> 	while( ( (*((volatile unsigned long * ) (ad | 0x00007800)))/4) < 1 ) {
> 		i = 1000 ;
> 		while( --i ) ;
> 	}
> }

I don't like the looks of what you're doing with pointers here.  It is
likely that GCC will never be able to optimize this very well.

> This problem is there with the i386 build of 3.2.
> The problem does not reproduce with 2.96 or earlier versions.
>
> I could locate the Zack mail when he reported this problem
> http://gcc.gnu.org/ml/gcc-patches/2000-02/msg00408.html. But could 
> not find any reference of the problem beyond that.
>
> Please help in locating the furthur work done on the problem. 
> If required, I would be like to contribute towards the fix of 
> the problem.

I do not believe any work has been done specifically on that problem.
However, it is likely to be addressed by completing the store-motion
pass, which has never worked properly.  If you would like to help with
that, talk to Daniel Berlin (cc:ed).

zw