>
> >
> > Please give Honza the chance to chime in.
>
> Does that mean this is approved since Honza has not chimed in over 2 weeks?
> Just want to double check for pushing it.
Sorry for being late. The patch makes sense to me
and thanks for looking into this.
I noticed is that in your original email:
> <bb 4> [local count: 1073741824]:
> # i_4 = PHI <0(2), i_11(3)>
> t_8 = *l_7(D);
> if (t_8 < 0)
> goto <bb 5>; [0.04%]
> else
> goto <bb 6>; [99.96%]
>
> <bb 5> [local count: 429496]:
> __builtin_unreachable ();
The probability of edge bb4->bb5 is wrong to have non-zero probability.
With trunk I see it gets predicted by cold call heuristics:
Predictions for bb 2
first match heuristics: 0.04%
combined heuristics: 0.04%
opcode values positive (on trees) heuristics of edge 2->3 (ignored): 41.00%
cold function call heuristics of edge 2->3 (ignored): 0.04%
noreturn call heuristics of edge 2->3: 0.04%
Predictions for bb 2
1 edges in bb 2 predicted to even probabilities
Predictions for bb 3
0 edges in bb 3 predicted to even probabilities
Predictions for bb 4
1 edges in bb 4 predicted to even probabilities
int size (int * a)
{
int t;
<bb 2> [local count: 1073741824]:
t_3 = *a_2(D);
if (t_3 < 0)
goto <bb 3>; [0.00%]
else
goto <bb 4>; [100.00%]
<bb 3> [count: 0]:
__builtin_unreachable ();
While GCC 15 does:
Predictions for bb 2
1 edges in bb 2 predicted to even probabilities
Predictions for bb 3
0 edges in bb 3 predicted to even probabilities
Predictions for bb 4
1 edges in bb 4 predicted to even probabilities
int size (int * a)
{
int t;
<bb 2> [local count: 1073741824]:
t_3 = *a_2(D);
if (t_3 < 0)
goto <bb 3>; [0.00%]
else
goto <bb 4>; [100.00%]
<bb 3> [count: 0]:
__builtin_unreachable ();
Here edge from bb2->bb3 is predicted with profile_probability::zero() as
part of backwards propagation from __builtin_unreachable and
consequently is not predicted later. Need to work out why this changed.
Honza
