The arc instrumentation code generated by -fprofile-arcs increments
gcov counters in a manner which is not guaranteed to be atomic. As
a result, if two or more cpus enter the same basic block at nearly
the same time they can leave an incorrect arc count, which makes it
impossible for gcov to solve that function's arc count graph later.
Within SGI, this is a significant limitation when doing kernel test
coverage studies, as we have code that needs to be exercised over
multiple processors. The same problem presumably affects people
trying to do coverage studies on multithreaded programs in userspace.
I've written a patch to solve this problem. It adds a new option,
-fprofile-multithread, which changes the tree profiler code to
use the recently-added __sync_add_and_fetch() builtin to update
the counters. I've tested it on ia64 and i386; it generates
the appropriate object code which runs and produces coverage
results.
Here's a diffstat:
common.opt | 4 +
doc/invoke.texi | 14 ++++-
testsuite/gcc.misc-tests/gcov-12.c | 28 ++++++++++
tree-profile.c | 75 +++++++++++++++++++++++++++++
4 files changed, 119 insertions(+), 2 deletions(-)
Reading the documention at gcc.gnu.org, this looks "legally significant"
so I'll need to have some paperwork done right? I assume a copyright
assignment form and an employer disclaimer are the correct documents.
Can someone start that process? Should I just attach the patch now?
--
Summary: Need atomic increment of gcov counters for MP programs
Product: gcc
Version: 4.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: gcov/profile
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: gnb at sgi dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28441
