PR 14042: C++ abstraction penalty (tree-ssa regression/opportunity)

[Timothy J. Wood]

   Since people are working on getting tree-ssa performing as well as or 
better than 3.4, I thought I'd mention this PR.  This is an 
optimization regression/opportunity in that tree-ssa currently 
generates worse code than 3.4 and even what 3.4 generates is far from 
optimal.  The problem presents itself when you have two stack allocated 
instances of a C++ class -- with only one, 3.4 will inline all the 
operations on the class, but with two instances, it starts writing them 
to the stack.  In the given test case, this causes an inner loop to 
have tons of load/store operations.

-tim