From: SourceForge.net <no...@so...> - 2006-09-15 20:29:11
|
Bugs item #1256978, was opened at 2005-08-11 14:51 Message generated for change (Settings changed) made by wcohen You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=116191&aid=1256978&group_id=16191 Please note that this message will contain a full copy of the comment thread, including the initial issue submission, for this request, not just the latest update. Category: None Group: None >Status: Closed Resolution: Fixed Priority: 7 Submitted By: Ravi Chintakunta (ravichintakunta) Assigned to: Philippe Elie (phil_e) Summary: opreport shows % values more than 100 Initial Comment: This is a snapshot of opreport run on an pcc/e500 target. The % for BRANCHES_FINISHED is 1943.61% which seems to be wrong. This can be observed if the profiling is kept run for long hours and allowing to collect huge samples. It seems the sample count variable (possibly long) is overflowing and gives an erratic value. -bash-2.05b# opreport CPU: e500, speed 825 MHz (estimated) Counted BRANCHES_FINISHED events (Branches finished) with a unit mask of 0x00 (No unit mask) count 7500 Counted BRANCHES_MISPREDICTED events (Branch instructions mispredicted due to direction, target, or IAB prediction) with a unit mask of 0x00 (No unit mask) count 7500 Counted COMPLETED_INSNS events (Completed Instructions (0, 1, or 2 per cycle)) with a unit mask of 0x00 (No unit mask) count 45000 Counted CPU_CLK events (Cycles) with a unit mask of 0x00 (No unit mask) count 45000 BRANCHES_FINIS...|BRANCHES_MISPR...|COMPLETED_INSN...| CPU_CLK:45000| samples| %| samples| %| samples| %| samples| %| ------------------------------------------------------------------------ 3889486282 1943.6147 14614764 29.9104 3494025925 85.5681 3103418048 85.5016 no-vmlinux 563256558 281.4649 29294797 59.9544 548219347 13.4258 485258857 13.3693 oprofiled 41498878 20.7374 4917292 10.0637 40129254 0.9828 39420126 1.0861 libc-2.3.3.so ---------------------------------------------------------------------- Comment By: Philippe Elie (phil_e) Date: 2005-08-15 16:34 Message: Logged In: YES user_id=318973 Fixed in cvs, all samples count sum are now done with a 64 unsigned bit type. It remains a posssible overflow over 32 bits integer in our sample file format but this one will be much difficult to reach as it mean 2**32-1 sample to the same eip/binary. regards, Philippe Elie ---------------------------------------------------------------------- Comment By: Philippe Elie (phil_e) Date: 2005-08-14 07:36 Message: Logged In: YES user_id=318973 Overflow when we sum up all counts, we need to do all sums with 64 bits integer. regards, Philippe Elie ---------------------------------------------------------------------- You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=116191&aid=1256978&group_id=16191 |