[flang][cuda] Bring PARAMETER arrays into the GPU module

[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/146416.diff

2 Files Affected:

• (modified) flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp (+9-1)
• (modified) flang/test/Fir/CUDA/cuda-device-global.f90 (+13)

diff --git a/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp b/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
index 328e2374115b0..bfb0daeacb8c3 100644
--- a/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
+++ b/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
@@ -113,8 +113,16 @@ class CUFDeviceGlobal : public fir::impl::CUFDeviceGlobalBase<CUFDeviceGlobal> {
       return signalPassFailure();
     mlir::SymbolTable gpuSymTable(gpuMod);
     for (auto globalOp : mod.getOps<fir::GlobalOp>()) {
-      if (cuf::isRegisteredDeviceGlobal(globalOp))
+      if (cuf::isRegisteredDeviceGlobal(globalOp)) {
         candidates.insert(globalOp);
+      } else if (globalOp.getConstant() &&
+                 mlir::isa<fir::SequenceType>(
+                     fir::unwrapRefType(globalOp.resultType()))) {
+        mlir::Attribute initAttr =
+            globalOp.getInitVal().value_or(mlir::Attribute());
+        if (initAttr && mlir::dyn_cast<mlir::DenseElementsAttr>(initAttr))
+          candidates.insert(globalOp);
+      }
     }
     for (auto globalOp : candidates) {
       auto globalName{globalOp.getSymbol().getValue()};
diff --git a/flang/test/Fir/CUDA/cuda-device-global.f90 b/flang/test/Fir/CUDA/cuda-device-global.f90
index 8cac643b27c34..4c634513745fd 100644
--- a/flang/test/Fir/CUDA/cuda-device-global.f90
+++ b/flang/test/Fir/CUDA/cuda-device-global.f90
@@ -11,3 +11,16 @@ module attributes {fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.conta
 
 // CHECK: gpu.module @cuda_device_mo
 // CHECK-NEXT: fir.global @_QMmtestsEn(dense<[3, 4, 5, 6, 7]> : tensor<5xi32>) {data_attr = #cuf.cuda<device>} : !fir.array<5xi32>
+
+// -----
+
+module attributes {fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.container_module} {
+  fir.global @_QMm1ECb(dense<[90, 100, 110]> : tensor<3xi32>) constant : !fir.array<3xi32>
+  fir.global @_QMm2ECc(dense<[100, 200, 300]> : tensor<3xi32>) constant : !fir.array<3xi32>
+}
+
+// CHECK: fir.global @_QMm1ECb
+// CHECK: fir.global @_QMm2ECc
+// CHECK: gpu.module @cuda_device_mod
+// CHECK-DAG: fir.global @_QMm2ECc
+// CHECK-DAG: fir.global @_QMm1ECb