https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119778
--- Comment #14 from Richard Biener <rguenth at gcc dot gnu.org> ---
In this case speculative call expansion is wrecking abnormal edges. Instead
of distributing an abnormal edge to both calls we leave it on the merge block
of the if ().
The actual issue seems to be that when early inlining
void S::foo (struct S * const this)
{
int (*) () * _1;
int (*) () _2;
;; basic block 2, loop depth 0
;; pred: ENTRY
_1 = this_4(D)->_vptr.S;
_2 = *_1;
OBJ_TYPE_REF(_2;(struct S)this_4(D)->0B) (this_4(D));
return;
into qux at
;; basic block 6, loop depth 0
;; pred: 2
;; 4
# a.1_4(ab) = PHI <a.1_9(D)(2), a.1_2(ab)(4)>
S::foo (y_13(D));
;; succ: 7
;; 5
while S::foo (y_13(D)) is marked as ctrl-altering due to the abnormal edge
the inlined OBJ_TYPE_REF(_2;(struct S)this_4(D)->0B) (this_4(D)) call is
not. We're set up to purge edges, but the indirect call expansion done
fails to mark inlined calls it adds abnormal edges from as control-altering.
