Svadev mailing list

["Ryan M. Lefever" <lefever AT illinois.edu>]

In the 2007 PLDI DSA paper, one of the cases for the local analysis 
(that constructs initial DSA graphs) is a value preserving cast.  Does a 
value preserving cast include int2ptr and ptr2int casts?  I am curious 
about the definition as far as the DSA algorithm is concerned and as far 
as the implementation is concerned.

It seams to me that DSA should not necessitate an implementation in 
which DS nodes associated with int2ptr and ptr2int casts have to be 
marked as unknown.  Int2ptr and ptr2int casts can just be considered 
part of the definition of a value preserving casts.  As long as the 
integer is big enough to hold a pointer, why does it matter that its 
typed as an integer?  DSA only uses type information during loads, 
stores, and pointer arithmetic.  If DSA can not preserve 
field-sensitivity after a ptr2int cast because of integer arithmetic, 
its no different than a program casting a value of T* to a char* and 
then performing pointer arithmetic.  If the integer arithmetic can not 
be deciphered, the associated node simply collapses and loses 
field-sensitivity.

In the end, by treating int2ptr and ptr2int casts as value-preserving 
casts, you'll be able to eliminate unknown nodes.  Sure some nodes might 
get merged together that would not have otherwise been merged, but if 
you hadn't treated int2ptr and ptr2int as value-preserving casts, those 
nodes would have been marked unknown and would't be reliable anyhow. 
You will probably end up with larger DS graphs too because you'll have 
more virtual registers to track, i.e., any virtual register big enough 
to hold an pointer.  However, eliminating unknown nodes seams like it 
should be of utmost priority.  After all, if you have one unknown node, 
since the unknown node can alias with any other unknown node, then by 
any node can alias any other node due to transitivity.

Does what I'm arguing seam correct?

Thanks,
Ryan