https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79347
Bug ID: 79347
Summary: [7 regression] vect_do_peeling is messing up profile
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: hubicka at gcc dot gnu.org
Target Milestone: ---
Hi,
compiling
short *a;
int c;
n()
{
for (int i = 0; i<c;i++)
a[i]++;
}
with -Ofast give now following mismatches in profile:
t.c.159t.vect:;; Invalid sum of incoming frequencies 1148, should be 1275
t.c.159t.vect:;; Invalid sum of incoming frequencies 1116, should be 1275
t.c.159t.vect:;; Invalid sum of incoming frequencies 1434, should be 1275
t.c.159t.vect:;; Invalid sum of incoming frequencies 2178, should be 1063
t.c.159t.vect:;; Invalid sum of incoming frequencies 159, should be 1275
t.c.159t.vect:;; Invalid sum of incoming frequencies 6965, should be 6694
t.c.159t.vect:;; Invalid sum of incoming frequencies 1163, should be 1275
while GCC 6 has only two:
/aux/hubicka/t.c.148t.vect:;; Invalid sum of incoming frequencies 962, should
be 590
/aux/hubicka/t.c.148t.vect:;; Invalid sum of incoming frequencies 53, should
be 425
Unforutnately I got lost in the code, so i am not sure how to fix this up.
This bug is now #1 source of profile mismatches when compiling tramp3d.
This seems to be caused by rewrite the way we generate loop prologue
https://gcc.gnu.org/ml/gcc-patches/2016-09/msg00821.html
and the profiles of prologue are not scaled down when the guards are inserted.
