Hello,
I've proposed a patch [1] for condition coverage profiling in gcc,
implemented in the middle-end alongside the branch coverage. I've
written most of the tests for C and a few for C++ and finally got around
to try it with a toy example for D and go and noticed something odd
about Go's CFG construction.
abc.c:
int fn (int a, int b, int c) {
if (a && (b || c))
return a;
else
return b * c;
}
abc.d:
int fn (int a, int b, int c) {
if (a && (b || c))
return a;
else
return b * c;
}
abc.go:
func fn (a int, b int, c int) int {
a_ := a != 0;
b_ := b != 0;
c_ := c != 0;
if a_ && (b_ || c_) {
return 1;
} else {
return 0;
}
}
All were built with gcc --coverage -fprofile-conditions (my patch, but
it does not affect this) and no optimization. The C and D programs
behaved as expected:
gcov --conditions abc.d:
3: 3:int fn (int a, int b, int c) {
3*: 4: if (a && (b || c))
conditions outcomes covered 3/6
condition 1 not covered (false)
condition 2 not covered (true)
condition 2 not covered (false)
1: 5: return a;
-: 6: else
2: 7: return b * c;
gcov --conditions abc.go:
3: 5:func fn (a int, b int, c int) int {
3: 6: a_ := a != 0;
3: 7: b_ := b != 0;
3: 8: c_ := c != 0;
-: 9:
3*: 10: if a_ && (b_ || c_) {
condition outcomes covered 2/2
condition outcomes covered 1/2
condition 0 not covered (true)
condition outcomes covered 2/2
1: 11: return 1;
-: 12: } else {
2: 13: return 0;
-: 14: }
-: 15:}
So I dumped the gimple gcc -fdump-tree-gimple abc.go:
int main.fn (int a, int b, int c)
{
int D.2725;
int $ret0;
$ret0 = 0;
{
bool a_;
bool b_;
bool c_;
a_ = a != 0;
b_ = b != 0;
c_ = c != 0;
{
{
GOTMP.0 = a_;
if (GOTMP.0 != 0) goto <D.2719>; else goto <D.2720>;
<D.2719>:
{
{
GOTMP.1 = b_;
_1 = ~GOTMP.1;
if (_1 != 0) goto <D.2721>; else goto <D.2722>;
<D.2721>:
{
GOTMP.1 = c_;
}
<D.2722>:
}
GOTMP.2 = GOTMP.1;
GOTMP.0 = GOTMP.2;
}
<D.2720>:
}
if (GOTMP.0 != 0) goto <D.2723>; else goto <D.2724>;
<D.2723>:
{
{
$ret0 = 1;
D.2725 = $ret0;
// predicted unlikely by early return (on trees) predictor.
return D.2725;
}
}
<D.2724>:
{
{
$ret0 = 0;
D.2725 = $ret0;
// predicted unlikely by early return (on trees) predictor.
return D.2725;
}
}
}
}
}
Where as D (and C) is more-or-less as you would expect:
int fn (int a, int b, int c)
{
int D.7895;
if (a != 0) goto <D.7893>; else goto <D.7891>;
<D.7893>:
if (b != 0) goto <D.7892>; else goto <D.7894>;
<D.7894>:
if (c != 0) goto <D.7892>; else goto <D.7891>;
<D.7892>:
D.7895 = a;
// predicted unlikely by early return (on trees) predictor.
return D.7895;
<D.7891>:
D.7895 = b * c;
// predicted unlikely by early return (on trees) predictor.
return D.7895;
}
Clearly the decision inference algorithm is unable to properly
instrument to Go program for condition coverage because of the use of
temporaries in the emitted GIMPLE. The question is: is this intentional
and/or required from Go's semantics or could it be considered a defect?
Is emitting the GIMPLE without the use of temporaries feasible at all?
Thanks,
Jørgen
[1] https://gcc.gnu.org/pipermail/gcc-patches/2022-July/598165.html
