[flang][cuda] Handle variable with initialization in device global pass

[clementval]

No description provided.

[llvmbot]

@llvm/pr-subscribers-flang-fir-hlfir

Author: Valentin Clement (バレンタイン クレメン) (clementval)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/124307.diff

2 Files Affected:

• (modified) flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp (+5-7)
• (modified) flang/test/Fir/CUDA/cuda-implicit-device-global.f90 (+21)

diff --git a/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp b/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
index 5ce39f99bbb126..7486dde0e281e0 100644
--- a/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
+++ b/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
@@ -36,13 +36,11 @@ static void processAddrOfOp(fir::AddrOfOp addrOfOp,
           addrOfOp.getSymbol().getRootReference().getValue())) {
     // TO DO: limit candidates to non-scalars. Scalars appear to have been
     // folded in already.
-    if (globalOp.getConstant()) {
-      if (recurseInGlobal)
-        globalOp.walk([&](fir::AddrOfOp op) {
-          processAddrOfOp(op, symbolTable, candidates, recurseInGlobal);
-        });
-      candidates.insert(globalOp);
-    }
+    if (recurseInGlobal)
+      globalOp.walk([&](fir::AddrOfOp op) {
+        processAddrOfOp(op, symbolTable, candidates, recurseInGlobal);
+      });
+    candidates.insert(globalOp);
   }
 }
 
diff --git a/flang/test/Fir/CUDA/cuda-implicit-device-global.f90 b/flang/test/Fir/CUDA/cuda-implicit-device-global.f90
index 9b22ed86e419ce..11866d871a607e 100644
--- a/flang/test/Fir/CUDA/cuda-implicit-device-global.f90
+++ b/flang/test/Fir/CUDA/cuda-implicit-device-global.f90
@@ -308,3 +308,24 @@ // Test that global used in device function are flagged with the correct
 // CHECK-DAG: fir.global linkonce_odr @_QM__mod1E.c.__builtin_c_devptr
 // CHECK-DAG: fir.global linkonce_odr @_QM__mod1E.dt.__builtin_c_devptr
 // CHECK-DAG: fir.global linkonce_odr @_QM__mod1E.n.__builtin_c_devptr
+
+// -----
+
+// Variables with initialization are promoted to non constant global.
+// 
+// attributes(global) subroutine kernel4()
+//   integer :: a = 4
+// end subroutine 
+
+func.func @_QPkernel4() attributes {cuf.proc_attr = #cuf.cuda_proc<global>} {
+  %0 = fir.address_of(@_QFkernel4Ea) : !fir.ref<i32>
+  return
+}
+fir.global internal @_QFkernel4Ea : i32 {
+  %c4_i32 = arith.constant 4 : i32
+  fir.has_value %c4_i32 : i32
+}
+
+// CHECK-LABEL: fir.global internal @_QFkernel4Ea : i32
+// CHECK-LABEL: gpu.module @cuda_device_mod
+// CHECK: fir.global internal @_QFkernel4Ea : i32