On Thu, Sep 18, 2025 at 06:54:06PM +0200, Jason Merrill wrote:
> > There are some regressions caused by the removal of {CLOBBER(bob)}
> > clobbers from the start of certain constructors, e.g. one testcase has
> > struct A
> > {
> > int f,g;
> > A() { f = g; // { dg-warning "g. is used uninitialized" } }
> > } a;
> > and this doesn't warn anymore for any C++ mode. Wonder what we could
> > do for the warnings, for code generation we have to assume the code
> > can read members in *this (for C++26 read there .DEFERRED_INIT, for
> > older C++ versions with -flifetime-dse=2 {CLOBBER(bob)}).
> > Perhaps we should add some artificial attribute to the functions we'd
> > add the -flifetime-dse=2 clobber to and treat *this reaching the start
> > of the function for those specially in the uninit pass?
>
> Or some way to indicate in GIMPLE that whatever value is there, it's
> poisoned?
Some new ifn which is say:
*this = .POISON (this);
or
*this = .POISON (*this);
would be an option (which would expand to nothing if the pointers are the
same), dunno what would be easier to implement though.
I'll try artificial attribute on the this PARM_DECL.
Richi, what is your preference from the middle-end POV?
Just a way to force -Wuninitialized/-Wmaybe-uninitialized warnings on
reads but similarly to .DEFERRED_INIT which sets to specific bytes
use whatever was in memory before.
Jakub
