[flang-commits] [flang] [llvm] [mlir] [flang][cuda] Support non-allocatable module-level managed variables (PR #188526)

[Valentin Clement バレンタイン クレメン via flang-commits]

================
@@ -915,3 +915,13 @@ llvm.func @nanosleep(%duration: i32) {
   nvvm.nanosleep %duration
   llvm.return
 }
+
+// -----
+
+// CHECK: @managed_g = addrspace(1) global i32 0
+// CHECK: !nvvm.annotations = !{![[MANAGED:[0-9]+]]}
+// CHECK: ![[MANAGED]] = !{ptr addrspace(1) @managed_g, !"managed", i32 1}
+llvm.mlir.global external @managed_g() {addr_space = 1 : i32, nvvm.managed} : i32 {
+  %0 = llvm.mlir.constant(0 : i32) : i32
----------------
clementval wrote:

I think we can update the op to imply 0 offset when none is given

https://github.com/llvm/llvm-project/pull/188526