On Sun, Oct 18, 2020 at 10:31:55PM +0100, Ramsay Jones wrote:
> On 17/10/2020 23:56, Luc Van Oostenryck wrote:
> > Most __sync_* or __atomic_* builtin functions have one or
> > more arguments having its real type determined by the first
> > argument: either the same type (a pointer to an integral type)
> > or the type of the object it points to.
> >
> > Currently, only __sync_{bool,val}_compare_and_swap() are handled
> > but lots of very similar variants would be needed for the others.
> > So, make it a generic function, able to handle all these builtins.
> >
> > Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@xxxxxxxxx>
> > ---
> > builtin.c | 47 +++++++++++++++++++++++++++++++----------------
> > 1 file changed, 31 insertions(+), 16 deletions(-)
> >
> > diff --git a/builtin.c b/builtin.c
> > index 0d4cb12cca22..880dd54f647e 100644
> > --- a/builtin.c
> > +++ b/builtin.c
> > @@ -31,6 +31,14 @@
> > #include "compat/bswap.h"
> > #include <stdarg.h>
> >
> > +#define dyntype incomplete_ctype
> > +static bool is_dynamic_type(struct symbol *t)
> > +{
> > + if (t->type == SYM_NODE)
> > + t = t->ctype.base_type;
> > + return t == &dyntype;
> > +}
> > +
> > static int evaluate_to_int_const_expr(struct expression *expr)
> > {
> > expr->ctype = &int_ctype;
> > @@ -362,29 +370,32 @@ static struct symbol_op overflow_p_op = {
> > };
> >
> >
> > -static int eval_sync_compare_and_swap(struct expression *expr)
> > +static int eval_atomic_common(struct expression *expr)
> > {
> > + struct symbol *fntype = expr->fn->ctype->ctype.base_type;
> > struct symbol_list *types = NULL;
> > struct symbol *ctype = NULL;
> > + struct symbol *t;
> > struct expression *arg;
> > int n = 0;
> >
> > - /* the first arg is a pointer type; we'd already verified that */
> > + // The number of arguments have already be verified.
> > + // The first arg must be a pointer type to an integral type.
>
> s/pointer type to/pointer to/
> (it just sounds odd, otherwise)
> or maybe ... must be a 'pointer to an integral' type. ?
Well, yes, it's a bit hard to express clearly. The real infos are:
* it must be a 'struct symbol *', the are used for any type, pointer or not
* the type represented by this 'struct symbol *' must indeed be a pointer :
(first_args->type == SYM_PTR) to an integral type:
(first_args->ctype.base_type = &int_ctype).
So yes, "must be a 'pointer to an integral' type. " seems to do the job.
Thanks
-- Luc
