https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104106
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think DSE doesn't quite understand free() as must-def, likewise for
__builtin_stack_restore though that is _much_ more difficult to tie to
a specific alloca allocation.
It would be nice if the gimplifier would emit CLOBBERs for the VLAs
that go out of scope at the __builtin_stack_restore point, that would
fix the VLA case I think. For
int f(char *a, unsigned n) {
char *tmp_a = __builtin_malloc (n);
for (unsigned i = 1; i != n; i++) tmp_a[i] = a[i];
__builtin_free (tmp_a);
return a[0];
}
it should already work but DSE is not set up to consider variable indexed
stores in loops that end up as pointer accesses like
*_5 = _6;
or rather stmt_kills_ref_p is too simple minded when looking at free():
case BUILT_IN_FREE:
{
tree ptr = gimple_call_arg (stmt, 0);
tree base = ao_ref_base (ref);
if (base && TREE_CODE (base) == MEM_REF
&& TREE_OPERAND (base, 0) == ptr)
{
++alias_stats.stmt_kills_ref_p_yes;
return true;
}
it might be able to use points-to analysis or track down the base of
_5 via DR analysis but in the end I think it's DSEs job to do better
here.
With
int f(char *a, unsigned n) {
char (*tmp_a)[n] = __builtin_malloc (n);
for (unsigned i = 1; i != n; i++) (*tmp_a)[i] = a[i];
__builtin_free (tmp_a);
return a[0];
}
this issue is avoided but we then run into
/* If we visit this PHI by following a backedge then we have to
make sure ref->ref only refers to SSA names that are invariant
with respect to the loop represented by this PHI node. */
if (dominated_by_p (CDI_DOMINATORS, gimple_bb (stmt),
gimple_bb (temp))
&& !for_each_index (ref->ref ? &ref->ref : &ref->base,
check_name, gimple_bb (temp)))
return DSE_STORE_LIVE;
which explicitely disables DSE of variably indexed accesses with the index
varying in the loop. That's because of cross iteration dependences otherwise
mishandled.
