https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82527
Bug ID: 82527
Summary: Branch probabilities does not match on optimized
switch cases
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: gcov-profile
Assignee: unassigned at gcc dot gnu.org
Reporter: eDeviser at mailbox dot org
CC: marxin at gcc dot gnu.org
Target Milestone: ---
Having a switch case with an empty case as shown below, causes a wrong branch
report:
#include <stdio.h>
int main() {
for (int i = 0; i < 100; i++) {
switch (i) {
case 0:
printf("0");
break;
case 9999:
break;
default:
printf("d");
break;
}
}
return 0;
}
Compile it, execute it and generate the *.gcov file:
gcc main.c -O3 -fprofile-arcs -ftest-coverage && ./a.out && gcov -bc main.c &&
cat main.c.gcov
Take a look to the report:
-: 0:Source:main.c
-: 0:Graph:main.gcno
-: 0:Data:main.gcda
-: 0:Runs:1
-: 0:Programs:1
-: 1:#include <stdio.h>
function main called 1 returned 100% blocks executed 100%
1: 2: int main() {
101: 3: for (int i = 0; i < 100; i++) {
branch 0 taken 100
branch 1 taken 1 (fallthrough)
100: 4: switch (i) {
branch 0 taken 1
branch 1 taken 99
branch 2 taken 0
-: 5: case 0:
-: 6: printf("0");
-: 7: break;
-: 8: case 9999:
-: 9: break;
-: 10: default:
-: 11: printf("d");
-: 12: break;
-: 13: }
-: 14: }
-: 15: return 0;
-: 16: }
Doing the same thing again but with `-O0` instead of `-O3` shows this report:
-: 0:Source:main.c
-: 0:Graph:main.gcno
-: 0:Data:main.gcda
-: 0:Runs:1
-: 0:Programs:1
-: 1:#include <stdio.h>
function main called 1 returned 100% blocks executed 91%
1: 2: int main() {
101: 3: for (int i = 0; i < 100; i++) {
branch 0 taken 100
branch 1 taken 1 (fallthrough)
100: 4: switch (i) {
branch 0 taken 1
branch 1 taken 0
branch 2 taken 99
-: 5: case 0:
1: 6: printf("0");
call 0 returned 1
1: 7: break;
-: 8: case 9999:
#####: 9: break;
-: 10: default:
99: 11: printf("d");
call 0 returned 99
99: 12: break;
-: 13: }
-: 14: }
1: 15: return 0;
-: 16: }
The coverage report says, that the switch case has three branches. Thats right.
The first branch is the `case 0`, the second one is the `case 999` and the last
one is the `default`. But the optimized report looks really wrong.
Lets put both branch reports next to each other to see the problem:
CASE | WITH -O0 | WITH -O3
==========|======================|=====================
case 0 | branch 0 taken 1 | branch 0 taken 1
case 9999 | branch 1 taken 0 | branch 1 taken 99
default | branch 2 taken 99 | branch 2 taken 0
This is an idea what the bug is about: For me it the optimized version seems to
remove the 9999 case because it is empty. So Branch 1 with -O3 is the case
default and Branch 2 with -O3 is not existent and must not be shown in the
report.
I hope the thing is clear. If not feel free to ask! :-)
