Hi!
On Wed, Nov 02, 2011 at 09:28:52PM -0600, Jeff Law wrote:
> This was caught by some prototype code to warn for potential NULL
> pointer dereferences.
Clearly we have a set of vec.h operations that require the vector to be
non-NULL.
Seems it is last, index (which already have such check in their VEC_ASSERT),
then the 7 ops you've changed, embedded_init which I think you've missed
and splice which you've changed which requires dst to be non-NULL, but src
may be NULL. For checking builds I'd say it is ok to add those to the
VEC_ASSERT, but don't we have to do anything about the non-ENABLE_CHECKING
case? We will still have code like:
(VEC_rtx_base_replace(((reg_base_value) ? &(reg_base_value)->base :
0),(rhs_regno(reg1)),((rhs_regno(reg2)) <
(VEC_rtx_base_length(((reg_base_value) ? &(reg_base_value)->base : 0))) ?
(VEC_rtx_base_index(((reg_base_value) ? &(reg_base_value)->base :
0),(rhs_regno(reg2)) )) : 0) ));
with
static __inline__ rtx
VEC_rtx_base_replace (VEC_rtx_base * vec_, unsigned ix_, rtx obj_)
{
rtx old_obj_;
(void) (vec_ && ix_ < vec_->prefix.num);
old_obj_ = vec_->vec[ix_];
vec_->vec[ix_] = obj_;
return old_obj_;
}
Shouldn't we for non-ENABLE_CHECKING just call it with
(&(reg_base_value)->base) instead of
((reg_base_value) ? &(reg_base_value)->base : 0)
?
We could do that with:
/* Convert to base type. */
#define VEC_BASE(P) ((P) ? &(P)->base : 0)
#if ENABLE_CHECKING
#define VEC_BASE_NONNULL(P) VEC_BASE (P)
#else
#define VEC_BASE_NUNNULL(P) (&(P)->base)
#endif
and use VEC_BASE_NONNULL in the macros where appropriate.
> * vec.h (splice): Assert DST_ is non-null.
> (quick_push, pop, replace, quick_insert): Similarly for VEC_.
> (ordered_remove, unordered_remove, block_remove): Likewise.
>
Jakub
