long double test failures due to underflow exception getting set?

[Steve Ellcey]

On Thu, 2017-07-20 at 17:32 +0000, Wilco Dijkstra wrote:
> Carlos wrote:
> > 
> > Neither DJ nor Szabolcs's glibc 2.26 builds showed those failures:
> Actually Szabolcs mentioned trunk GCC has some issues with long double.
> I don't know the details but I imagine GCC lifts some FP operations across
> fegetenv/fesetenv calls (or similar) which may be a bug depending on the
> optimization setting (I guess you need -ftrapping-math to block such
> lifting).
> 
> Wilco

Yes, this does appear to be a GCC bug. Â I have created a test case that
I can submit to GCC where the code does the right thing at -O0 and the
wrong thing with -O2. Â I tried using -ftrapping-math but that didn't
help. Â I will submit a GCC bug shortly, here is the test case if you
are interested, it is based on e_expl.c but can be compiled as a
standalone program.

Steve Ellcey
sellcey@cavium.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: x.c
Type: text/x-csrc
Size: 183157 bytes
Desc: not available
URL: <http://sourceware.org/pipermail/libc-alpha/attachments/20170720/ab8e1e62/attachment.bin>