Update wrapper function for atomic add intrinsic.

Until recently the atomic-fadd-no-rtn-insts attribute was needed to bypass feature checks when compiling device library code. Now that we have v_illegal the compiler can handle this case (https://reviews.llvm.org/D123693). However, the code that lowers to either global_atomic_fadd or v_illegal (SIISelLowering.cpp:7825) must rely on atomic-fadd-no-rtn-insts to determine if the target supports it or not (https://reviews.llvm.org/differential/changeset/?ref=3507251).

Change-Id: I99fa0a669abcc9c1c1738d1a4c694455b7bee856
(cherry picked from commit 5bc72d0)

ockl/src/gaaf.cl

@@ -17,7 +17,7 @@
 
 extern float __llvm_amdgcn_global_atomic_fadd_f32_p1f32_f32(__global float *, float) __asm("llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32");
 
-__attribute__((target("atomic-fadd-no-rtn-insts"))) static void
+static void
 global_atomic_fadd(__global float *p, float v)
 {
     __llvm_amdgcn_global_atomic_fadd_f32_p1f32_f32(p, v);