Re: [tree-ssa]: Work on making SSAPRE bootstrap

[Daniel Berlin <dberlin at dberlin dot org>]

On Monday, March 3, 2003, at 06:34  PM, Pop Sébastian wrote:

> On Mon, Mar 03, 2003 at 03:21:43PM -0500, Daniel Berlin wrote:
> > Sigh.
> > SSAPRE can be aggravating sometimes. The papers are buggy and missing
> > bits of info in some cases, and Open64 (a reference implementation) is
> > completely different than the papers (though i've implemented pieces 
> > of
> > it for comparison while bugfixing, as it's sometimes the only way to 
> > find
> > a bug).  Did I mention their SSAPRE implementation is also roughly 
> > 14000
> > lines of incredibly object oriented C++?
>
> I have browsed Open64's SSAPRE last summer and indeed it is a complex 
> piece
> of code.
>  I wonder how they verified the rightness of this code...

They have a testsuite that isn't public, i would assume it tests 
various cases.
I actually have a testsuite of 10 pre (in increasing complexity), 10 
strength reduction, and 10 load pre cases that i use.  But I hand 
verify results, since the best i could do reasonably (for example, the 
temp ids/ssa version numbers and labels and whatnot can change between 
compilations, so i can't rely on resulting code compares) is compare 
the results of the dataflow computations, which almost never are broken 
(Most bugs occur in either rename or finalizing).
> A general question is do they used a good testsuite, or do they had
> another technique for verifying complex optimizations?
No clue if they use anything else, however.