Re: atf(7) results on NetBSD/m68k with gcc48

On Mon, Dec 01, 2014 at 11:51:45PM +0900, Izumi Tsutsui wrote:
> I wrote:
> 
> > Now atf-run(1) on TT030 is complete.
> 
> I've also tested atf-run(1) on XM6i (X68030 emulator) with
> both stock gcc 4.5.4 and gcc 4.8.4 on NetBSD/amd64 6.1.5 (Core i7 3770).
> It's much faster (~9 hours with 50MHz/1008MB RAM and "no-wait" mode)
> than real 32MHz/68MB TT030..
> 
> With gcc 4.5.4 (201411272240Z snapshot):
>  https://gist.github.com/tsutsui/6a0340297e57e538a920
> ---
> Summary for 592 test programs:
> 3709 passed test cases.
> 112 failed test cases.
> 28 expected failed test cases.
> 94 skipped test cases. 
> ---
> 
> With gcc 4.8.4 (built locally, a bit different date)
>  https://gist.github.com/tsutsui/ea3e018eb8257edf0da3
> ---
> Summary for 592 test programs:
> 3701 passed test cases.
> 119 failed test cases.
> 28 expected failed test cases.
> 95 skipped test cases. 
> ---
> 
> I have not checked which tests actually fail, but looks not so bad.


I did a similar comparison on mac68k and the only failure I saw
that was new with gcc48 was in lib/libc/t_wcstod.
in particular, this line of the table is the one that fails:

{ L"0.12345678",        10,     0.12345678,             0 },


the problem is in the test program rather than libc.
but I didn't see anything wrong with the generated code after
staring at it for a while.

I found two workarounds: compile t_wcstod.c with -O0,
or add "volatile" to the "double d" in the loop.

-Chuck