AMDGPU: Switch barrier intrinsics to using convergent

noduplicate prevents unrolling of small loops that happen to have
barriers in them. If a loop has a barrier in it, it is OK to duplicate
it for the unroll.

llvm-svn: 256075
diff --git a/llvm/test/CodeGen/AMDGPU/flat-address-space.ll b/llvm/test/CodeGen/AMDGPU/flat-address-space.ll
index d65b757..86e0c07 100644
--- a/llvm/test/CodeGen/AMDGPU/flat-address-space.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-address-space.ll
@@ -131,5 +131,5 @@
 declare i32 @llvm.r600.read.tidig.x() #3
 
 attributes #0 = { nounwind }
-attributes #1 = { nounwind noduplicate }
+attributes #1 = { nounwind convergent }
 attributes #3 = { nounwind readnone }