[flang][cuda] Use cuf.register_variable_static instead of CUFRegisterExternalVariable (#211886)

This gives more flexibility to the backend to select a different
registration mechanism

flang-rt/lib/cuda/registration.cpp

@@ -48,14 +48,6 @@ void RTDEF(CUFRegisterVariable)(
   __cudaRegisterVar(module, varSym, varName, varName, 0, size, 0, 0);
 }
 
-void RTDEF(CUFRegisterExternalVariable)(
-    void **module, char *varSym, const char *varName, int64_t size) {
-  // Tell the CUDA driver to bind the device-side global <varName> to the
-  // host-resident storage at <varSym>. Kernel accesses to <varName> then go
-  // through the host address; HMM/ATS handles migration.
-  __cudaRegisterHostVar(module, varName, varSym, size);
-}
-
 void RTDEF(CUFRegisterManagedVariable)(
     void **module, void **varSym, char *varName, int64_t size) {
   __cudaRegisterManagedVar(module, varSym, varName, varName, 0, size, 0, 0);

flang/include/flang/Optimizer/Transforms/Passes.td

@@ -556,7 +556,7 @@ def CUFAddConstructor : Pass<"cuf-add-constructor", "mlir::ModuleOp"> {
     Option<"cudaUnified", "cuda-unified", "bool", /*default=*/"false",
            "Treat host module globals as unified memory (-gpu=mem:unified): "
            "register every plain host module global that is mirrored in the "
-           "GPU module via CUFRegisterExternalVariable, so the CUDA driver "
+           "GPU module via cuf.register_variable_static, so the CUDA driver "
            "maps the device-side symbol to the host pointer at module-load "
            "time and HMM/ATS handles migration.">
   ];

flang/include/flang/Runtime/CUDA/registration.h

@@ -28,13 +28,6 @@ void RTDECL(CUFRegisterFunction)(
 void RTDECL(CUFRegisterVariable)(
     void **module, char *varSym, const char *varName, int64_t size);
 
-/// Register a module-scope variable as host-resident under -gpu=mem:unified,
-/// so that the device-side symbol of the same name is mapped to the host
-/// pointer at module-load time. Wraps __cudaRegisterHostVar. Kernel accesses
-/// to the variable then reach the host storage directly via HMM/ATS.
-void RTDECL(CUFRegisterExternalVariable)(
-    void **module, char *varSym, const char *varName, int64_t size);
-
 /// Register a managed variable.
 void RTDECL(CUFRegisterManagedVariable)(
     void **module, void **varSym, char *varName, int64_t size);

flang/lib/Optimizer/Transforms/CUDA/CUFAddConstructor.cpp

@@ -77,7 +77,7 @@ static fir::GlobalOp createManagedPointerGlobal(fir::FirOpBuilder &builder,
 /// Return true if \p hostGlobal is a host module-scope global that has been
 /// mirrored in the GPU module as an external (no-body) declaration by the
 /// CUFDeviceGlobal pass under -gpu=mem:unified. Such globals must be
-/// registered with the CUDA driver via CUFRegisterExternalVariable so the
+/// registered with the CUDA driver via cuf.register_variable_static so the
 /// device-side `.extern` symbol resolves to the host pointer at module-load
 /// time and HMM/ATS handles migration.
 static bool isCudaUnifiedExternalGlobal(fir::GlobalOp hostGlobal,
@@ -173,8 +173,7 @@ static uint64_t getGlobalSizeInBytes(mlir::Location loc,
 /// Emit a call to a CUF registration runtime function with the canonical
 /// (module, addr, name, size) signature, where addr is the address of \p
 /// addrGlobal taken via fir.address_of and name/size describe \p nameGlobal.
-/// Used both for CUFRegisterVariable / CUFRegisterManagedVariable / and
-/// CUFRegisterExternalVariable.
+/// Used both for CUFRegisterVariable / CUFRegisterManagedVariable.
 static void
 emitCUFRegistrationCall(fir::FirOpBuilder &builder, mlir::Location loc,
                         mlir::Type idxTy, const mlir::DataLayout &dl,
@@ -410,12 +409,13 @@ struct CUFAddConstructor
           for (fir::GlobalOp globalOp : mod.getOps<fir::GlobalOp>()) {
             if (!isCudaUnifiedExternalGlobal(globalOp, gpuSymTable))
               continue;
-            auto func = fir::runtime::getRuntimeFunc<mkRTKey(
-                CUFRegisterExternalVariable)>(loc, builder);
-            emitCUFRegistrationCall(builder, loc, idxTy, *dl, kindMap,
-                                    typeConverter, registeredMod, func,
-                                    /*addrGlobal=*/globalOp,
-                                    /*nameGlobal=*/globalOp);
+            uint64_t szBytes = getGlobalSizeInBytes(loc, *dl, kindMap,
+                                                    typeConverter, globalOp);
+            cuf::RegisterVariableStaticOp::create(
+                builder, loc,
+                mlir::SymbolRefAttr::get(ctx, globalOp.getSymName()),
+                builder.getStringAttr(globalOp.getSymName()),
+                builder.getI64IntegerAttr(szBytes));
           }
         }
 

flang/lib/Optimizer/Transforms/CUDA/CUFDeviceGlobal.cpp

@@ -249,7 +249,7 @@ public:
       // External linkage (see convertLinkage in CodeGen.cpp), so an
       // initializer-less global emits as `.extern .global ...` in PTX.
       // The host-side definition stays. CUFAddConstructor will emit
-      // CUFRegisterExternalVariable (= __cudaRegisterHostVar) so the CUDA
+      // cuf.register_variable_static so the CUDA
       // runtime maps the device extern to the host pointer at module-load
       // time, and HMM/ATS handles migration.
       if (cudaUnified && !globalOp.getConstant() &&

flang/test/Fir/CUDA/cuda-constructor-2.f90

@@ -211,7 +211,7 @@ module attributes {dlti.dl_spec = #dlti.dl_spec<#dlti.dl_entry<!llvm.ptr, dense<
 // Under -gpu=mem:unified, a non-allocatable host module global referenced from
 // device code (mirrored as an external-linkage clone in the GPU module by the
 // CUFDeviceGlobal pass) must be registered with
-// _FortranACUFRegisterExternalVariable so the device-side `.extern` symbol
+// cuf.register_variable_static so the device-side `.extern` symbol
 // resolves to the host pointer at module-load time.
 
 module attributes {dlti.dl_spec = #dlti.dl_spec<i8 = dense<8> : vector<2xi64>, i16 = dense<16> : vector<2xi64>, i1 = dense<8> : vector<2xi64>, !llvm.ptr = dense<64> : vector<4xi64>, f80 = dense<128> : vector<2xi64>, i128 = dense<128> : vector<2xi64>, i64 = dense<64> : vector<2xi64>, !llvm.ptr<271> = dense<32> : vector<4xi64>, !llvm.ptr<272> = dense<64> : vector<4xi64>, f128 = dense<128> : vector<2xi64>, !llvm.ptr<270> = dense<32> : vector<4xi64>, f16 = dense<16> : vector<2xi64>, f64 = dense<64> : vector<2xi64>, i32 = dense<32> : vector<2xi64>, "dlti.stack_alignment" = 128 : i64, "dlti.endianness" = "little">, fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.container_module, llvm.data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128", llvm.target_triple = "x86_64-unknown-linux-gnu"} {
@@ -229,14 +229,13 @@ module attributes {dlti.dl_spec = #dlti.dl_spec<i8 = dense<8> : vector<2xi64>, i
 
 // UNIFIED: llvm.func internal @__cudaFortranConstructor()
 // UNIFIED: cuf.register_module @cuda_device_mod -> !llvm.ptr
-// UNIFIED: fir.address_of(@_QMmtestsEm) : !fir.ref<!fir.array<5xi32>>
-// UNIFIED: fir.call @_FortranACUFRegisterExternalVariable
+// UNIFIED: cuf.register_variable_static @_QMmtestsEm
 // UNIFIED-NOT: fir.call @_FortranACUFInitModule
 
 // -----
 
 // Under -gpu=mem:unified, an allocatable host module global also gets
-// registered via _FortranACUFRegisterExternalVariable; the registered symbol
+// registered via cuf.register_variable_static; the registered symbol
 // is the descriptor (fir.box<fir.heap<...>>). The host runtime allocates the
 // data buffer in HMM/ATS-accessible memory.
 
@@ -257,12 +256,12 @@ module attributes {dlti.dl_spec = #dlti.dl_spec<i8 = dense<8> : vector<2xi64>, i
 
 // UNIFIED: llvm.func internal @__cudaFortranConstructor()
 // UNIFIED: cuf.register_module @cuda_device_mod -> !llvm.ptr
-// UNIFIED: fir.address_of(@_QMmtestsEma) : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
-// UNIFIED: fir.call @_FortranACUFRegisterExternalVariable
+// UNIFIED: cuf.register_variable_static @_QMmtestsEma
 // UNIFIED-NOT: fir.call @_FortranACUFInitModule
 
 // -----
 
+module attributes {dlti.dl_spec = #dlti.dl_spec<i8 = dense<8> : vector<2xi64>, i16 = dense<16> : vector<2xi64>, i1 = dense<8> : vector<2xi64>, !llvm.ptr = dense<64> : vector<4xi64>, f80 = dense<128> : vector<2xi64>, i128 = dense<128> : vector<2xi64>, i64 = dense<64> : vector<2xi64>, !llvm.ptr<271> = dense<32> : vector<4xi64>, !llvm.ptr<272> = dense<64> : vector<4xi64>, f128 = dense<128> : vector<2xi64>, !llvm.ptr<270> = dense<32> : vector<4xi64>, f16 = dense<16> : vector<2xi64>, f64 = dense<64> : vector<2xi64>, i32 = dense<32> : vector<2xi64>, "dlti.stack_alignment" = 128 : i64, "dlti.endianness" = "little">, fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.container_module, llvm.data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128", llvm.target_triple = "x86_64-unknown-linux-gnu"} {
 func.func @_QPsub1() {
   %0 = cuf.alloc !fir.box<!fir.heap<!fir.array<?xf32>>> {bindc_name = "a", data_attr = #cuf.cuda<device>, uniq_name = "_QFsub1Ea"} -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>
   %4:2 = hlfir.declare %0 {data_attr = #cuf.cuda<device>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QFsub1Ea"} : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>) -> (!fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>>)
@@ -274,6 +273,7 @@ func.func @_QPsub1() {
   cuf.free %4#1 : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf32>>>> {data_attr = #cuf.cuda<device>}
   return
 }
+}
 
 // CHECK: llvm.func internal @__cudaFortranConstructor()
 // CHECK: llvm.call @_FortranACUFRegisterAllocator()

flang/test/Fir/CUDA/cuda-unified-module-global.f90

@@ -3,7 +3,7 @@
 //   1. is mirrored into the GPU module by CUFDeviceGlobal as a no-body
 //      external declaration (so PTX gets `.extern .global ...`); and
 //   2. is registered with the CUDA driver via
-//      _FortranACUFRegisterExternalVariable (= __cudaRegisterHostVar) from
+//      cuf.register_variable_static from
 //      __cudaFortranConstructor, so the device-side symbol is mapped to
 //      the host-resident storage at module-load time and HMM/ATS handles
 //      migration.
@@ -44,6 +44,5 @@ module attributes {dlti.dl_spec = #dlti.dl_spec<i8 = dense<8> : vector<2xi64>, i
 // Constructor registers the host pointer.
 // CHECK: llvm.func internal @__cudaFortranConstructor()
 // CHECK: cuf.register_module @cuda_device_mod -> !llvm.ptr
-// CHECK: fir.address_of(@_QMmtestsEm) : !fir.ref<!fir.array<5xi32>>
-// CHECK: fir.call @_FortranACUFRegisterExternalVariable
+// CHECK: cuf.register_variable_static @_QMmtestsEm
 // CHECK-NOT: fir.call @_FortranACUFInitModule