On Mon, 2 Sept 2024 at 18:40, Jonathan Wakely <jwakely.gcc@xxxxxxxxx> wrote:
>
> On Mon, 2 Sept 2024 at 18:03, James K. Lowden <jklowden@xxxxxxxxxxxxxxx> wrote:
> >
> > I am experiencing problems that appear to be in std::variant. I wonder
> > if this has been reported before, and I just need to upgrade.
> >
> > The problem appears to be in this switch:
> >
> > class number_t {
> > enum type_t { int_e, float_e };
> > std::variant<ssize_t, double> value;
> > public:
> > ...
> > number_t& operator=( const number_t& that ) {
> > assert( that.value.index() != std::variant_npos );
> > switch( type_t(that.value.index()) ) {
> > case int_e: value = std::get<ssize_t>(that.value); break;
> > case float_e: value = std::get<double>(that.value); break;
> > }
> > return *this;
> > }
> >
> > That's not how I wrote the code originally. That's a re-write to try to isolate the problem.
> >
> > valgrind reports:
> >
> > ==999106== Conditional jump or move depends on uninitialised value(s)
> > ==999106== at 0x1130B8: number_t::operator=(number_t const&) (expression.h:80)
> > ==999106== by 0x123C23: env_t::numop(number_t& (*)(number_t&, exp_t const&), exp_t const&) (core.cc:429)
> > ==999106== by 0x123FA3: env_t::mul(exp_t const&, env_t&) (core.cc:457)
> > ==999106== by 0x12286F: env_t::eval(exp_t const&, env_t&) (core.cc:148)
> > ==999106== by 0x122767: env_t::eval(exp_t const&, env_t&) (core.cc:136)
> > ==999106== by 0x1104EB: yyparse() (parse.y:319)
> > ==999106== by 0x12DF3F: repl() (main.cc:30)
> > ==999106== by 0x12E20F: main (main.cc:79)
> >
> > It seems to me that if the assert doesn't trip, then the object must be initialized, in which case the switch cannot depend on an uninitialized value.
>
> I don't think your assumption is valid. If the variant's _M_index
> value is uninitialized, then the assert could pass because the index
> could be any garbage value. Is the RHS of the assignment (the object
> that the 'that' reference is bound to) correctly initialized? Are you
> sure?
Also, is the assert actually being checked, i.e. no NDEBUG?
