[Bug middle-end/66867] Suboptimal code generation for atomic_compare_exchange

[dhowells at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66867

--- Comment #11 from dhowells at redhat dot com <dhowells at redhat dot com> ---
I applied the patch to the Fedora cross-gcc-6.1.1 rpm with one minor fixup. 
Using the example code I added in bug 70825 I now see:

0000000000000000 <test_atomic_cmpxchg>:
   0:   ba 2a 00 00 00          mov    $0x2a,%edx
   5:   b8 17 00 00 00          mov    $0x17,%eax
   a:   f0 0f b1 17             lock cmpxchg %edx,(%rdi)
   e:   c3                      retq   

there's now no extraneous store before the locked instruction.  And:

000000000000000f <test_atomic_cmpxchg_2>:
   f:   ba 2a 00 00 00          mov    $0x2a,%edx
  14:   b8 17 00 00 00          mov    $0x17,%eax
  19:   f0 0f b1 17             lock cmpxchg %edx,(%rdi)
  1d:   c3                      retq   

it now just passes the return value of cmpxchg back directly without
potentially putting on and off the stack and maybe jumping round that bit. 
And:

0000000000000043 <test_atomic_cmpxchg_B>:
  43:   ba 2a 00 00 00          mov    $0x2a,%edx
  48:   b8 17 00 00 00          mov    $0x17,%eax
  4d:   f0 0f b1 17             lock cmpxchg %edx,(%rdi)
  51:   c3                      retq   

where it makes no difference changing how the the return-statements are
contructed in C.

I've also booted a kernel built with the patched compiler.