Re: valgrind-tests.m4

[Simon Josefsson]

Bruno Haible <address@hidden> writes:

> The approach I use in GNU gettext is similar: Makefile.am has this:
>
> # For debugging memory leaks and memory allocation bugs.
> # You should build with --disable-shared when using valgrind.
> CHECKER =
> #CHECKER = valgrind --tool=memcheck 
> --suppressions=$(srcdir)/../gnulib-lib/malloca.valgrind --num-callers=20 
> --leak-check=yes --leak-resolution=high --show-reachable=yes
> #CHECKER = valgrind --tool=massif --format=html --depth=10 --alloc-fn=xmalloc 
> --alloc-fn=xrealloc --stacks=no
>
> TESTS_ENVIRONMENT = top_srcdir=$(top_srcdir) \
>                     
> PATH=.:../src:../../gettext-runtime/src:$(top_srcdir)/../gettext-runtime/src:$$PATH
>  \
>                     GETTEXT="$(CHECKER) tstgettext" \
>                     NGETTEXT="$(CHECKER) tstngettext" \
>                     XGETTEXT="$(CHECKER) xgettext" \
>                     MSGATTRIB="$(CHECKER) msgattrib" \
>                     MSGCAT="$(CHECKER) msgcat" \
>                     MSGCMP="$(CHECKER) msgcmp" \
>                     MSGCOMM="$(CHECKER) msgcomm" \
>                     MSGCONV="$(CHECKER) msgconv" \
>                     MSGEN="$(CHECKER) msgen" \
>                     MSGEXEC="$(CHECKER) msgexec" \
>                     MSGFILTER="$(CHECKER) msgfilter" \
>                     MSGFMT="$(CHECKER) msgfmt" \
>                     MSGGREP="$(CHECKER) msggrep" \
>                     MSGINIT="$(CHECKER) msginit" \
>                     MSGMERGE="$(CHECKER) msgmerge" \
>                     MSGUNFMT="$(CHECKER) msgunfmt" \
>                     MSGUNIQ="$(CHECKER) msguniq" \
>                     ...

I like this approach.  I think valgrind-tests.m4 could support this
approach too, it is relatively close to how I'm using it.

> This idiom has the effect that:
>   - Use of valgrind is turned off by default.
>     I think that's the right setting, because
>       1. tests running with valgrind consume much more time that without.
>       2. some tests may fail due to bugs in valgrind. For example, valgrind
>          built for 32-bit failed miserably when run with 64-bit binaries on
>          x86 bi-arch systems.

My preference is to have it on by default (tests that are disabled by
default tends to never get exercised) but this could be a maintainer
option.

>   - The user can enable it by changing 2 characters in the Makefile.
>   - The user does not have to remember the recommended options
>     (--alloc-fn=xmalloc etc.).
>   - The user has the choice among multiple valgrind tools.
>   - Shell invocations are not checked by valgrind, only the invocations
>     of the programs that are meant to be tested.

Sounds good.

/Simon