Re: [tree-ssa][mainline] Fix counting of line executions in gcov

[law at redhat dot com]

In message <20030826174533.GA1127@atrey.karlin.mff.cuni.cz>, Zdenek Dvorak writ
es:
 >Hello,
 >
 >the following problem showed up with tree-ssa-branch + LOOP_EXPR
 >lowering patch; since gcov is unchanged there, it also applies for
 >mainline.
 >
 >As far as I understand the code in accumulate_line_counts to determine
 >number of executions of a line that contains several basic blocks, it
 >tries to split the graph into elementary cycles.  The problem is that
 >there are a few bugs in the algorithm; when the cycle is found, it
 >should be removed by decreasing counts of all edges it belong to it,
 >this is not done.  Instead the markers of the path are left in the
 >graph, which does not work -- consider
 >
 >         - 10 <---------
 >       /                 \
 >     /                     \
 >A   C -5-> D -5-> E --10--> B
 >  \   \         /          /
 >    \    -5->--          /
 >      ->-- 1 -----------
 >
 >With current algorithm, a cycle B - C - D - E is found, contributing 5.
 >E is left marked as a part of the path, so it does not find any
 >other cycles, giving total result of 6 (with 1 from entry edge),
 >instead of 11 we want.
 >
 >Zdenek
 >
 >	* gcov.c (typedef struct arc_info): New field cs_count.
 >	(accumulate_line_counts): Find cycles correctly.
This is fine if you install a testcase with the patch.  It would seem
to me the easiest way to get a testcase is to reformulate the original
failing test's loop using gotos and use the result as your new testcase.

jeff