I'm using gcc 4.3.2 x86_64
I'm trying to work around consequences of a reinterpret_cast. I
verified that compiling with -fno-strict-aliasing fixes the problem, but
I don't want to do that to the whole compilation unit. I want to limit
the change to the one function with the problem.
Sorry I haven't reduced this to an example small enough to post. I'm
hoping someone know how to use __attribute__((__may_alias__)) well
enough to answer this from just the two lines of code exhibiting the
problem.
DAT* pp=foo();
reinterpret_cast<DAT_PTR*>(&pp)->bar(x);
DAT is a class.
DAT_PTR is the class of an object that is the size of a DAT* and the
contents of a DAT* and acts almost like a DAT* but not exactly.
Constructing or destroying an actual DAT_PTR is impractical at that
point in the code. Making a DAT* and pretending it is a DAT_PTR fits
the situation perfectly.
But the compiler optimizes away the pp=foo() because it sees pp is never
used as itself.
I tried
DAT __attribute__((__may_alias__))* pp=foo();
and
DAT* __attribute__((__may_alias__)) pp=foo();
but neither of those make any difference.
I've solved problems like this in the past with unions, but I'd prefer
not to in this case. I though __attribute__((__may_alias__)) was
supposed to deal with this issue.
I'd like to know how to use that in general for a few other cases (in a
very big collection of old code) where I expect to need it.
