[Bug rtl-optimization/48986] Missed optimization in atomic decrement on x86/x64

["jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48986

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |NEW
                 CC|                            |uros at gcc dot gnu.org
         AssignedTo|jakub at gcc dot gnu.org    |unassigned at gcc dot
                   |                            |gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-05-16 11:26:51 UTC ---
On:
int
foo (int *p)
{
  return __sync_fetch_and_add (p, -1) == 1;
}

int
bar (int *p)
{
  return __sync_add_and_fetch (p, -1) == 0;
}

I get better generated code for the second routine if I do:
--- gcc/config/i386/sync.md.jj 72010-05-21 11:46:29.000000000 +0200
+++ gcc/config/i386/sync.md 2011-05-16 13:06:13.000000000 +0200
@@ -170,7 +170,7 @@
   [(match_operand:SWI 1 "memory_operand" "+m")] UNSPECV_XCHG))
    (set (match_dup 1)
         (plus:SWI (match_dup 1)
-                  (match_operand:SWI 2 "register_operand" "0")))
+                  (match_operand:SWI 2 "nonmemory_operand" "0")))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_XADD"
   "lock{%;} xadd{<imodesuffix>}\t{%0, %1|%1, %0}")

and for foo identical code, so maybe that change is always beneficial, allowing
combiner and other early RTL passes to see there a constant instead of a REG.
Unfortunately, even with this change the combiner doesn't attempt to combine
this pattern with the following cmpsi_1 pattern, supposedly because
sync_old_addsi pattern isn't single_set.  I guess we could handle this during
expansion, but it would be a mess, or some other pass (e.g. peephole2 or
something similar).  peephole2 might kind of too late though, by that time the
constant must be loaded already into some register, so we'd need to peephole2 3
insns, where the load of the constant might often not be the first one.