Re: [Valgrind-developers] Shell fails when under valgrind-3.7.0'scontrol with "--tool=none" ? ... / was: Fwd:ast-ksh.2012-05-18 failures in "arith.sh" when under valgrindcontrol...

[Philippe W. <phi...@sk...>]

On Thu, 2012-06-28 at 05:28 +0200, Roland Mainz wrote:
> Erm... ksh93 doesn't use |fork()| for command substitutions in a
> subshell (e.g. x=$(...)) or in subshells itself (mainly for
> performance reasons... which gives a *major* performance boost (and
> causes less scalabilty issues with very large SMP/NUMA machines))
> unless you touch a system resource which can't be reversed by the
> shell itself (e.g. changing ulimit will force a |fork()| ... which is
> used in some tests of the AST/ksh93 test suite to test whether
> |fork()| causes trouble... see below...). Shell variables (and other
> resources) are handled in a copy-on-write manner if changed in the
> subshell.
> The interesting part is... if we force a |fork()| in the tests (e.g.
> do a x=$( ulimit -c 0 ; ...) instead of x=$(...)) the problem goes
> away.
You might try other tools to search for possible bugs causing
this (e.g. --tool=helgrind, --tool=drd, --tool=exp-sgcheck).
(is the ksh93 using threads ? if not, helgrind/drd cannot help)

Alternatively, if you are doing tricky things with signals,
maybe --vex-iropt-precise-memory-exns=yes might help.

You can also debug in parallel a native run of ksh93 and a run
under Valgrind, and see when they start to diverge.


Philippe