[Bug gcov-profile/67937] New: gcov gives wrong results when negative counts are involved

[Pidgeot18 at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67937

            Bug ID: 67937
           Summary: gcov gives wrong results when negative counts are
                    involved
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: gcov-profile
          Assignee: unassigned at gcc dot gnu.org
          Reporter: Pidgeot18 at gmail dot com
  Target Milestone: ---

This is hard to name correctly, since there's really a bigger bug here, and I'm
focusing on the lesser bug because it's easier for me to build a minimized test
case.

First the explanation: for whatever reason, sometimes gcov produces counts with
negative edges (that's a bug by itself). An minimized example of such a case
will be attached to the bug. In this scenario, the graph looks like this:

     1147 +-+ -42
   /----->|B|------\
  +-+     +-+     +-+
/>|A|      V 1189 |D| \
| +-+     +-+     +-+ |
|  \----->|C|------/  |
\     0   +-+ 1189    /
 \-------------------/
          23

(This is a result of an optimized loop, I don't have the corresponding code on
hand, sorry :-( ).

In this scenario, the loop detection code finds the following loops:
D -> A -> B -> C -> D: loop count = 23
D -> A -> B -> D: loop count = -42
D -> A -> C -> D: loop count = 0
A -> B -> C -> D -> A: loop count = 42
A -> B -> D -> A: loop count = 0
B -> C -> D -> A -> B: loop count = 0
C -> D -> A -> B -> C: loop count = 0

The last four loops are finding permutations of the three simple loops, which
is wrong (of course, the fact that one edge has a count of -42 in the first
place is wrong in the first place.