[patch] Predictive commoning, updated

[Zdenek Dvorak]

Hello,

> >You can also use the omega test: -fcheck-data-deps that should compute
> >the distance vectors using two methods, but that will stress only a part
> >of the data dependence analysis, i.e. not the extraction of the dependence
> >test, but only its resolution.  Last time I checked, it bootstrapped gcc, 
> >but
> >I haven't tried it yet on the cpu2000.
> >
> 
> Bootstrap (c only) without your patch is passing with BOOT_CFLAGS= "-O2
> -fcheck-data-deps", but including your patch it stops with the error:
> 
> (Number of distance vectors differ: Banerjee has 1, Omega has 2.
> Banerjee dist vectors:
>  0
> Omega dist vectors:
>  0
>  1
> data dependence relation:
> (Data Dep:
>  inner loop index: 0
>  loop nest: (49 )
>  distance_vector:   0
>  distance_vector:   1
>  direction_vector:     =
>  direction_vector:     +
> )
> )
> 
> For an array accessed twice using a constant access function, i.e.
> loop_0
>  A[5] =       (1)
>  ... = A[5]   (2)
> endloop_0
> 
> you will have both a distance of 0, as there is a dependence between (1) and
> (2) within the same iteration, and a distance of 1, as there is also a
> dependence
> from (2) to (1) but in the next iteration, so this is at a distance of 1.

that's sort of weird, as the patch cannot affect anything in this case.
I will check.

Zdenek