Re: results of gnulib tests with -fsanitize=address

[Paul Eggert]

On 05/19/2017 08:02 AM, Bruno Haible wrote:
> the problem comes from the 'closeout' module: When main() is finished,
> it closes fd 1 and 2, and the libasan then cannot perform the output it wants 
> to do.

Yes, closeout should not close stderr when addresses are being 
sanitized. I installed the attached patch to work around the problem. 
Perhaps a better fix could be made (e.g., close_stdout could 
conditionally call AddressSanitizer's leak_check_at_exit before closing 
stderr), but this workaround should suffice for now.

> 2) test-u8-casecoll
> It fails because HAVE_ICONV is not defined - because the test program in 
> m4/iconv.m4
> fails due to an iconv_t leak / memory leak.
>
> What is the best practice here? Should -fsanitize=address only be activated
> after configure has run? Or should all configure run tests be changed so that 
> they
> free all allocated memory?

The latter is surely too much work for not enough benefit. In general, a 
'configure'-time test should focus on one individual feature, and not 
attempt to test a combination of features, as there are too many 
combinations. This is partly why we advise people not to run 'configure' 
with -Werror, as that tests not only the desired features but also 
whether the 'configure'-time test pacifies picky compilers. Similarly, 
we should advise people not to run 'configure' with sanitization options 
like memory-leak detection that affect valid (albeit inefficient) test 
programs.

However, it should be OK to configure with -fsanitize=address if the 
ASAN_OPTIONS environment variable suppresses checks for 
valid-albeit-inefficent behavior. For example, 
ASAN_OPTIONS='detect_leaks=0' may suffice. (I haven't checked this.)

0001-closeout-don-t-close-stderr-when-sanitizing.patch
Description: Text Data