Hi,
As a Valgrind user, I see few ways to get different results with valgrind :
- your code can use introspection: depending on memory footprint or
computation time, you may choose an algorithm or an other. Usually you
can configure your code to avoid it temporary.
- your binary may contain x87 floating point instructions (with 80bits).
Valgrind is not able to guarantee the reproducibility (as well you are
not able to guarantee the portability of your code upon different
processors). For more details you can search x87 on page
http://valgrind.org/docs/manual/manual-core.html#manual-core.limits . By
default gcc should avoid them except if you use options like
--mfpmath=387 , if you use library with x87 hard-coded with assembly
(like openblas) or if you use long double.
- your binary may modify the floating-point register to change the
rounding-mode and Valgrind will ignore it. The valgrind option
--show-emwarns=yes is for me really useful (whereas described as
"usually not interesting" in the man page).
To pinpoint branch divergences, you can compile your code with coverage
options (-fprofile-arcs -ftest-coverage), and then compare the covers
produced with gcov (with and without valgrind) with a graphical diff. If
(and only if) you are familiar with code coverage it is really easy.
This trick is often used successfully (but not always) with Verrou
(Warning advertising) a valgrind tool to detect floating point rounding
error.
I hope this helps
Bruno Lathuilière
On 10/14/18 11:14 PM, olologin wrote:
Hi everyone,
First of all I want to thank everyone involved valgrind development,
this set of tools, especially memcheck and helgrind saved me a lot of
time. There were so many hard-to-find bugs in our code that were found
only thanks to these tools.
Now to the question: I've been using memcheck and helgrind for almost
a year now, and recently I noticed small instability in results
provided by our code. For example if I run my program 10 times with
the same output, 5/10 times I get wrong output, and I cannot
understand why. So I tried memcheck, but when I run my software in it
I always get right output, no matter how often I rerun it - output is
always stable, and of course memcheck doesn't report any issues. Same
issue happens with helgrind. I can suspect one of dynamic libraries
for this instability, but I cannot find exact place that causes this
issue. I have sources of that library, but it's quite massive, and I'm
not original developer of it and unfortunately I cannot show it since
it's closed-source. Most probably this issue is not caused by
race-conditions, because it happens even when I run everything in
single thread.
So far I tried to play with:
--smc-check=all
--redzone-size=*
--fair-sched=yes
--expensive-definedness-check=yes
Is there anything else I can try? I feel like there is some legitimate
way of executing random branch of code, and somehow when I launch my
program in valgrind - it always runs only good parts of code, this is
why I don't see my program failing, and I don't see any issues
reported in valgrind. But is there a way to make it run in same way as
without valgrind? I mean to make it produce bad output in valgrind?
I'm using valgrind-3.12.0, my program is compiled with gcc6.3
Thanks in advance.
_______________________________________________
Valgrind-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/valgrind-users
_______________________________________________
Valgrind-users mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/valgrind-users
