[cxx-mem-model] Change library call for __atomic_op_fetch routines

[Andrew MacLeod]

The external library only provides the 
__atomic_fetch_{add,sub,and,or,xor,nand} routines, which fetch the value 
in memory before the operation is performed.  It does not provide the 
alternate operation which fetches the value after the operation (it is 
trivial to calculate).

If the __atomic_{op}_fetch routine (which returns the value after the 
operation) cannot be inlined, it needs to be transformed into the 
fetch-before form, and have the appropriate calculation done to return 
the correct value.

ie, if

      c = __atomic_add_fetch (&b, a, __ATOMIC_RELAXED)

requires an external library call, it needs to be transformed into

      c = __atomic_fetch_add (&b, a, __ATOMIC_RELAXED) + a;

This patch takes care of that.  It seems to work when I hacked up some 
test cases, but until I implement a flag to disable inlining all 
atomics, its pretty hard to test thoroughly.

I find the bit in  expand_builtin_atomic_fetch_op which changes the 
function decl for the call a bit hacky maybe, but it seems to work.  It 
looks like that is exactly the way a builtin is built up.  (ie, I don't 
think the gcc_assert I added can fail, or previous calls to 
get_callee_fndecl() wouldn't have returned a value which got us to this 
hunk of code in the first place.   Ie, the code there is reverse 
engineered from   get_callee_fndecl() to find the function decl and 
change it.   we know the built-in variations that are being changed are 
identical in parameters and everything else.

thoughts?

Anyway, it bootstraps and passes all tests on x86_64-unknown-linux-gnu.

Andrew

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: op_func.diff
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20111026/796e5ed6/attachment.ksh>