[Bug tree-optimization/28778] [4.0/4.1/4.2 Regression] alias bug with cast and call clobbered

["dberlin at dberlin dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #28 from dberlin at gcc dot gnu dot org  2006-08-27 21:09 -------
Subject: Re:  [4.0/4.1/4.2 Regression] alias bug with cast and call clobbered

> Since we have never done that before, it does require new code.
>

And to answer what may be a followup, which is "why hasn't this broken
anything else before", it iss because even though the clobbering
question and the alias question are subtly different, the answers are
almost always the same.  It is only because we prune the local answers
using TBAAthat we get different answers depending on context.

If you exclude strict type-based rules, the answer to "what can foo
clobber" in the example is the same as asking "what can the first
argument of foo access in foo and its callees".  Because of the
type-based rules, we end up with a disconnect.  It would be undefined
to dereference the first argument of foo without casting it first.  In
effect, the strict-type rules allow you can take an unaliased pointer
in a caller, and validly turn it back into an aliased pointer in the
callee.

In order to account for this, you need to build the sets without
accounting for the type-based rules,  and if you want to apply the
type-based rules to clobbering, use them on a context-sensitive basis
(IE at the point of dereference, determine what could be accessed, and
propagate that back up through other functions to get the a better
clobber answer).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28778