[Bug target/58945] New: Improve atomic_compare_and_swap*_doubleword pattern

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58945

            Bug ID: 58945
           Summary: Improve atomic_compare_and_swap*_doubleword pattern
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ubizjak at gmail dot com

Created attachment 31126
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=31126&action=edit
Patch to simplify atomic_compare_and_swap<dwi>_doubleword pattern

Currently, gcc generates lots of unnecessary moves for following testcase
(actually gcc.dg/atomic-store-6.c):

--cut here--
__int128_t i;

int main()
{
  __atomic_store_16(&i, -1, 0);
  if (i != -1)
    __builtin_abort();
  return 0;
}
--cut here--

gcc -O2 -mcx16:

        movq    i(%rip), %rsi
        movq    $-1, %rcx
        movq    i+8(%rip), %rdi
.L2:
        movq    %rsi, %rax
        movq    %rdi, %rdx
        movq    %rcx, %rbx
        lock cmpxchg16b i(%rip)
        movq    %rdx, %rdi
        movq    %rax, %rsi
        jne     .L2

I have experimented with sync x86_64 pattern, to see if we can simplify
atomic_compare_and_swap<dwi>_doubleword pattern, so it would work with real
doublemode values and this way remove unnecessary register moves. However, I
have hit an ICE in assing_by_spills that prevents me from using these
simplifications:

gcc -O2 -mcx16

atomic-store-6.c: In function ‘main’:
atomic-store-6.c:13:1: internal compiler error: in assign_by_spills,
at lra-assigns.c:1281
 }
 ^
0x883e7f assign_by_spills
        ../../gcc-svn/trunk/gcc/lra-assigns.c:1281
0x884320 lra_assign()
        ../../gcc-svn/trunk/gcc/lra-assigns.c:1439
Please submit a full bug report,

I believe that the pattern is correct, and LRA is tripping on some
unknown issue.