Re: Regarding chi2_quality.cc test in libstdc++-v3 testsuite for gcc

[Renlin Li <renlin dot li at arm dot com>]

Thank you for your reply, Jonathan.

Sorry, it dose pass for x86_64_-unknown-linux-gnu and i686-*-linux-gnu. 
I just mistakenly use "-std=gnu++0x -DSAMPLES=10000" compilation option 
for x86 target. This dg-options is set when running the test on 
simulators, so it takes less time. If not overwritten, it should be 
300000 by default.

But test for x86_64_-unknown-linux-gnu and aarch64-none-linux-gnu target 
still pass even SAMPLES is redefined as 10000. Test for 
x86_64_-unknown-linux-gnu using -m32 and arm-none-linux-gnu fail with 
SAMPLES=10000.

So, there are two factors that make the differences.
1, the size of SAMPLES. The bigger the better.
2, different hash() results for 64-bit and 32-bit target.

I have tested that SAMPLES=30000 is a good value that make it pass for 
32bit target running on simulator. And it doesn't take that much time 
either.

Any opinions about this?

Kind regards,
Renlin Li

On 30/10/13 11:44, Jonathan Wakely wrote:
> On 30 October 2013 10:53, Renlin Li wrote:
> > Has anyone noticed that, this test fails for 32bit architecture linking with
> > glibc? The VERIFY statement ( VERIFY( chi2 < k*1.1 );) in
> > test_uniform_random() and test_bit_flip_set() fail.
> >
> > I have tested it on x86_32 and arm-none-linux-eabi, they both fail with the
> > same result. However it passes for x86_64 and aarch64-none-linux-gnu.
> It passes for me on x86_64-unknown-linux-gnu when using -m32, and it
> passes for i686-*-linux-gnu according to reports on the
> gcc-testresults mailing list.