[flang-commits] [flang] [flang][cuda] Import type descriptor in the gpu module when needed (PR #149157)

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

https://github.com/clementval created https://github.com/llvm/llvm-project/pull/149157

None

>From 557fee5038c1620c73601a57a41ccb92d5810b80 Mon Sep 17 00:00:00 2001
From: Valentin Clement <clementval at gmail.com>
Date: Wed, 16 Jul 2025 10:19:45 -0700
Subject: [PATCH] [flang][cuda] Import type descriptor in the gpu module when
 needed

---
 .../Optimizer/Transforms/CUFDeviceGlobal.cpp  | 42 +++++++++++++------
 flang/test/Fir/CUDA/cuda-device-global.f90    | 22 ++++++++++
 2 files changed, 52 insertions(+), 12 deletions(-)

diff --git a/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp b/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
index bfb0daeacb8c3..35badb6eadb1c 100644
--- a/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
+++ b/flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp
@@ -53,21 +53,26 @@ static void processAddrOfOp(fir::AddrOfOp addrOfOp,
   }
 }
 
+static void processTypeDescriptor(fir::RecordType recTy,
+                                  mlir::SymbolTable &symbolTable,
+                                  llvm::DenseSet<fir::GlobalOp> &candidates) {
+  if (auto globalOp = symbolTable.lookup<fir::GlobalOp>(
+          fir::NameUniquer::getTypeDescriptorName(recTy.getName()))) {
+    if (!candidates.contains(globalOp)) {
+      globalOp.walk([&](fir::AddrOfOp op) {
+        processAddrOfOp(op, symbolTable, candidates,
+                        /*recurseInGlobal=*/true);
+      });
+      candidates.insert(globalOp);
+    }
+  }
+}
+
 static void processEmboxOp(fir::EmboxOp emboxOp, mlir::SymbolTable &symbolTable,
                            llvm::DenseSet<fir::GlobalOp> &candidates) {
   if (auto recTy = mlir::dyn_cast<fir::RecordType>(
-          fir::unwrapRefType(emboxOp.getMemref().getType()))) {
-    if (auto globalOp = symbolTable.lookup<fir::GlobalOp>(
-            fir::NameUniquer::getTypeDescriptorName(recTy.getName()))) {
-      if (!candidates.contains(globalOp)) {
-        globalOp.walk([&](fir::AddrOfOp op) {
-          processAddrOfOp(op, symbolTable, candidates,
-                          /*recurseInGlobal=*/true);
-        });
-        candidates.insert(globalOp);
-      }
-    }
-  }
+          fir::unwrapRefType(emboxOp.getMemref().getType())))
+    processTypeDescriptor(recTy, symbolTable, candidates);
 }
 
 static void
@@ -85,6 +90,17 @@ prepareImplicitDeviceGlobals(mlir::func::FuncOp funcOp,
   }
 }
 
+static void
+processPotentialTypeDescriptor(mlir::Type candidateType,
+                               mlir::SymbolTable &symbolTable,
+                               llvm::DenseSet<fir::GlobalOp> &candidates) {
+  if (auto boxTy = mlir::dyn_cast<fir::BaseBoxType>(candidateType))
+    candidateType = boxTy.getEleTy();
+  candidateType = fir::unwrapSequenceType(fir::unwrapRefType(candidateType));
+  if (auto recTy = mlir::dyn_cast<fir::RecordType>(candidateType))
+    processTypeDescriptor(recTy, symbolTable, candidates);
+}
+
 class CUFDeviceGlobal : public fir::impl::CUFDeviceGlobalBase<CUFDeviceGlobal> {
 public:
   void runOnOperation() override {
@@ -115,6 +131,8 @@ class CUFDeviceGlobal : public fir::impl::CUFDeviceGlobalBase<CUFDeviceGlobal> {
     for (auto globalOp : mod.getOps<fir::GlobalOp>()) {
       if (cuf::isRegisteredDeviceGlobal(globalOp)) {
         candidates.insert(globalOp);
+        processPotentialTypeDescriptor(globalOp.getType(), parentSymTable,
+                                       candidates);
       } else if (globalOp.getConstant() &&
                  mlir::isa<fir::SequenceType>(
                      fir::unwrapRefType(globalOp.resultType()))) {
diff --git a/flang/test/Fir/CUDA/cuda-device-global.f90 b/flang/test/Fir/CUDA/cuda-device-global.f90
index 4c634513745fd..a2875a2dd241d 100644
--- a/flang/test/Fir/CUDA/cuda-device-global.f90
+++ b/flang/test/Fir/CUDA/cuda-device-global.f90
@@ -24,3 +24,25 @@ module attributes {fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.conta
 // CHECK: gpu.module @cuda_device_mod
 // CHECK-DAG: fir.global @_QMm2ECc
 // CHECK-DAG: fir.global @_QMm1ECb
+
+// -----
+
+module attributes {fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.container_module} {
+  fir.global @_QMmEddarrays {data_attr = #cuf.cuda<managed>} : !fir.box<!fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>> {
+    %c0 = arith.constant 0 : index
+    %0 = fir.zero_bits !fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>
+    %1 = fir.shape %c0 : (index) -> !fir.shape<1>
+    %2 = fir.embox %0(%1) {allocator_idx = 3 : i32} : (!fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>, !fir.shape<1>) -> !fir.box<!fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>>
+    fir.has_value %2 : !fir.box<!fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>>
+  }
+  fir.global linkonce_odr @_QMmE.dt.devicearrays constant target : !fir.type<_QM__fortran_type_infoTderivedtype{binding:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTbinding{proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>}>>>>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>,sizeinbytes:i64,uninstantiated:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,kindparameter:!fir.box<!fir.ptr<!fir.array<?xi64>>>,lenparameterkind:!fir.box<!fir.ptr<!fir.array<?xi8>>>,component:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,genre:i8,category:i8,kind:i8,rank:i8,__padding0:!fir.array<4xi8>,offset:i64,characterlen:!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>,derived:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,lenvalue:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,bounds:!fir.box<!fir.ptr<!fir.array<?x?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_ptr{__address:i64}>}>>>>,procptr:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTprocptrcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,offset:i64,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,special:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTspecialbinding{which:i8,isargdescriptorset:i8,istypebound:i8,isargcontiguousset:i8,__padding0:!fir.array<4xi8>,proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,specialbitset:i32,hasparent:i8,noinitializationneeded:i8,nodestructionneeded:i8,nofinalizationneeded:i8,nodefinedassignment:i8,__padding0:!fir.array<3xi8>}> {
+    %0 = fir.undefined !fir.type<_QM__fortran_type_infoTderivedtype{binding:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTbinding{proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>}>>>>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>,sizeinbytes:i64,uninstantiated:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,kindparameter:!fir.box<!fir.ptr<!fir.array<?xi64>>>,lenparameterkind:!fir.box<!fir.ptr<!fir.array<?xi8>>>,component:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,genre:i8,category:i8,kind:i8,rank:i8,__padding0:!fir.array<4xi8>,offset:i64,characterlen:!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>,derived:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,lenvalue:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,bounds:!fir.box<!fir.ptr<!fir.array<?x?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_ptr{__address:i64}>}>>>>,procptr:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTprocptrcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,offset:i64,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,special:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTspecialbinding{which:i8,isargdescriptorset:i8,istypebound:i8,isargcontiguousset:i8,__padding0:!fir.array<4xi8>,proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,specialbitset:i32,hasparent:i8,noinitializationneeded:i8,nodestructionneeded:i8,nofinalizationneeded:i8,nodefinedassignment:i8,__padding0:!fir.array<3xi8>}>
+    fir.has_value %0 : !fir.type<_QM__fortran_type_infoTderivedtype{binding:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTbinding{proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>}>>>>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>,sizeinbytes:i64,uninstantiated:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,kindparameter:!fir.box<!fir.ptr<!fir.array<?xi64>>>,lenparameterkind:!fir.box<!fir.ptr<!fir.array<?xi8>>>,component:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,genre:i8,category:i8,kind:i8,rank:i8,__padding0:!fir.array<4xi8>,offset:i64,characterlen:!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>,derived:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,lenvalue:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,bounds:!fir.box<!fir.ptr<!fir.array<?x?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_ptr{__address:i64}>}>>>>,procptr:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTprocptrcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,offset:i64,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,special:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTspecialbinding{which:i8,isargdescriptorset:i8,istypebound:i8,isargcontiguousset:i8,__padding0:!fir.array<4xi8>,proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,specialbitset:i32,hasparent:i8,noinitializationneeded:i8,nodestructionneeded:i8,nofinalizationneeded:i8,nodefinedassignment:i8,__padding0:!fir.array<3xi8>}>
+  }
+}
+
+// CHECK: fir.global @_QMmEddarrays
+// CHECK: fir.global linkonce_odr @_QMmE.dt.devicearrays
+// CHECK: gpu.module @cuda_device_mod
+// CHECK: fir.global @_QMmEddarrays
+// CHECK: fir.global linkonce_odr @_QMmE.dt.devicearrays