Re: Implementing "restrict", and other alias questions

[Toon Moene <toon at moene dot indiv dot nluug dot nl>]

>  5. Are there any important missing aliasing optimizations
>  in Fortran?

The basic rule in Fortran is:  If you don't *tell* the compiler  
there's overlap then it can assume there isn't.  The only way to  
tell the compiler about overlapping storage is via EQUIVALENCE.

So your question boils down to:

1. Does your alias analysis detect all forms of overlap due to the
   use of the EQUIVALENCE construct, e.g.:

      SUBROUTINE AAP
      DIMENSION WIM(100), ZUS(100)
      EQUIVALENCE (WIM(10), ZUS(20))
      [ initialisation of ZUS ]
      DO I = 1, 91
         WIM(I) = ZUS(I+9) + 10.0
      ENDDO
      END

2. Two different local arrays cannot overlap (unless by EQUIVALENCE, 
   like above).  Two different arrays in COMMON cannot overlap at all.
   Does your alias analysis detect those cases ?

>  (Moving stores out of loops doesn't count -- this is a
>  loop optimizer deficiency.)

For those who don't know the context of this remark - I came across  
this deficiency while looking for code on which John's alias  
analysis would be a win, so naturally I thought that the inability  
of g77 to move stores to loop-constant addresses out of the loop was  
related to alias analysis - however, John pointed out to me that  
the current loop optimisation pass doesn't do this at all.

Toon.