[Bug c/78386] New: Powerpc64le: optimization -O2 and higher cause math inconsitency

["brenohl at br dot ibm.com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78386

            Bug ID: 78386
           Summary: Powerpc64le: optimization -O2 and higher cause math
                    inconsitency
           Product: gcc
           Version: 6.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: brenohl at br dot ibm.com
  Target Milestone: ---

When compiling fdlibm with O2 or higher, the math results are different.

I just write a simple example, invoking jcos() and invoking it several times.
If I use -O1 and -O2, the results are different. I dug further and found that
the problem happens when using the optimization '-expensive-optimizations'.

As for example:

 $ gcc -I. -g -static -O1 -fexpensive-optimizations .. -o opt_call
 $ gcc -I. -g -static -O1 .. -o call

 $ ./opt_call 
cosine total sum is: 0.33853897900538427

 $ ./call 
cosine total sum is: 0.33853897900536206

I wrote a testcase to show this problem, please find it at
https://github.com/leitao/fdlibm/tree/master/gcc_bug:

In order to reproduce the problem, just run:

 $ git clone https://github.com/leitao/fdlibm.git
 $ cd fdlibm/gcc_bug
 $ make
 $ ./call
 $ ./opt_call

If any other detail is required, I can dig further:

Machine Information:

 $ gcc --version
gcc (Debian 6.2.0-13) 6.2.0 20161109
Copyright (C) 2016 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

 $ uname -a
Linux testing 4.7.0-1-powerpc64le #1 SMP Debian 4.7.8-1 (2016-10-19) ppc64le
GNU/Linux