enhanced structure offset tracking

[Joseph S. Myers]

On Mon, 3 Dec 2001, Richard Henderson wrote:

> > (through the alias information allowing a call to a nonexistent function
> > to be optimised away).
> 
> Maybe, but that's getting fairly tricky.

As long as we don't have a system for unit testing that specific
optimisations have specific effects on specific tree or RTL fragments, we
need to include testcases that do test the optimisations as best as we can
from C code.  Otherwise the optimisations will get broken and no-one will
notice immediately.  Tests for link failure work fine in
gcc.c-torture/execute/builtin-{abs*,complex-1,noret-1}.c; likewise you can
test that the compiler proves that a load and a store do not conflict (and
of course make sure that the test does indeed fail with the compiler
before your patch).

-- 
Joseph S. Myers
jsm28@cam.ac.uk