[Bug tree-optimization/48702] [4.6/4.7 Regression] optimization regression with gcc-4.6 on x86_64-unknown-linux-gnu

[rakdver at kam dot mff.cuni.cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48702

--- Comment #17 from rakdver at kam dot mff.cuni.cz <rakdver at kam dot mff.cuni.cz> 2011-05-17 20:18:39 UTC ---
> This is not really heuristic -- it prevents compiler from generating code in
> ivopt that violates the aliasing assumption. 

no, it does not.  For example, you only consider the case whent the base of the
memory
reference is addr_expr, and when the step is negative.  However, ivopts may
generate
"pointers" outside of the memory objects in other cases as well; and if those
get used in
mem_refs (where the computations are currently interpreted according to the
pointer
arithmetics rules, for the purpose of alias analysis), this leads to
misscompilations.