[OpenACC][CIR] Implement 'present' lowering for compute/combined

erichkeane · erichkeane · commit 4eb72312e26b · 2025-06-06T16:54:58.000-07:00
'present' has a data-in operation acc.present, and uses 'delete' for its
data out.  Otherwise it is identical to no_create.
diff --git a/clang/lib/CIR/CodeGen/CIRGenOpenACCClause.cpp b/clang/lib/CIR/CodeGen/CIRGenOpenACCClause.cpp
@@ -871,6 +871,22 @@ class OpenACCClauseCIREmitter final
       return clauseNotImplemented(clause);
     }
   }
+
+  void VisitPresentClause(const OpenACCPresentClause &clause) {
+    if constexpr (isOneOfTypes<OpTy, mlir::acc::ParallelOp, mlir::acc::SerialOp,
+                               mlir::acc::KernelsOp>) {
+      for (auto var : clause.getVarList())
+        addDataOperand<mlir::acc::PresentOp, mlir::acc::DeleteOp>(
+            var, mlir::acc::DataClause::acc_present, /*structured=*/true,
+            /*implicit=*/false);
+    } else if constexpr (isCombinedType<OpTy>) {
+      applyToComputeOp(clause);
+    } else {
+      // TODO: When we've implemented this for everything, switch this to an
+      // unreachable. data & declare remain.
+      return clauseNotImplemented(clause);
+    }
+  }
 };
 
 template <typename OpTy>
diff --git a/clang/test/CIR/CodeGenOpenACC/combined.cpp b/clang/test/CIR/CodeGenOpenACC/combined.cpp
@@ -1114,4 +1114,39 @@ extern "C" void acc_combined_data_clauses(int *arg1, int *arg2) {
   // CHECK-NEXT: } loc
   // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<host>]) {dataClause = #acc<data_clause acc_no_create>, name = "arg2"}
   // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<host>]) {dataClause = #acc<data_clause acc_no_create>, name = "arg1"}
+
+#pragma acc parallel loop present(arg1)
+  for(unsigned I = 0; I < 5; ++I);
+  // CHECK-NEXT: %[[NOCREATE1:.*]] = acc.present varPtr(%[[ARG1]] : !cir.ptr<!cir.ptr<!s32i>>) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg1"}
+  // CHECK-NEXT: acc.parallel combined(loop) dataOperands(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) {
+  // CHECK-NEXT: acc.loop combined(parallel) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) {dataClause = #acc<data_clause acc_present>, name = "arg1"}
+
+#pragma acc serial loop present(arg2)
+  for(unsigned I = 0; I < 5; ++I);
+  // CHECK-NEXT: %[[NOCREATE2:.*]] = acc.present varPtr(%[[ARG2]] : !cir.ptr<!cir.ptr<!s32i>>) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg2"}
+  // CHECK-NEXT: acc.serial combined(loop) dataOperands(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) {
+  // CHECK-NEXT: acc.loop combined(serial) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) {dataClause = #acc<data_clause acc_present>, name = "arg2"}
+
+#pragma acc kernels loop present(arg1, arg2) device_type(host) async
+  for(unsigned I = 0; I < 5; ++I);
+  // CHECK-NEXT: %[[NOCREATE1:.*]] = acc.present varPtr(%[[ARG1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<host>]) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg1"}
+  // CHECK-NEXT: %[[NOCREATE2:.*]] = acc.present varPtr(%[[ARG2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<host>]) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg2"}
+  // CHECK-NEXT: acc.kernels combined(loop) dataOperands(%[[NOCREATE1]], %[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>, !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<host>]) {
+  // CHECK-NEXT: acc.loop combined(kernels) {
+  // CHECK: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.terminator
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<host>]) {dataClause = #acc<data_clause acc_present>, name = "arg2"}
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<host>]) {dataClause = #acc<data_clause acc_present>, name = "arg1"}
 }
diff --git a/clang/test/CIR/CodeGenOpenACC/kernels.c b/clang/test/CIR/CodeGenOpenACC/kernels.c
@@ -469,4 +469,22 @@ void acc_kernels_data_clauses(int *arg1, int *arg2) {
   // CHECK-NEXT: } loc
   // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {dataClause = #acc<data_clause acc_no_create>, name = "arg2"}
   // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {dataClause = #acc<data_clause acc_no_create>, name = "arg1"}
+
+#pragma acc kernels present(arg1)
+  ;
+  // CHECK-NEXT: %[[NOCREATE1:.*]] = acc.present varPtr(%[[ARG1]] : !cir.ptr<!cir.ptr<!s32i>>) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg1"}
+  // CHECK-NEXT: acc.kernels dataOperands(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) {
+  // CHECK-NEXT: acc.terminator
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) {dataClause = #acc<data_clause acc_present>, name = "arg1"}
+
+#pragma acc kernels present(arg1, arg2) device_type(nvidia) async
+  ;
+  // CHECK-NEXT: %[[NOCREATE1:.*]] = acc.present varPtr(%[[ARG1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg1"}
+  // CHECK-NEXT: %[[NOCREATE2:.*]] = acc.present varPtr(%[[ARG2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg2"}
+  // CHECK-NEXT: acc.kernels dataOperands(%[[NOCREATE1]], %[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>, !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {
+  // CHECK-NEXT: acc.terminator
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {dataClause = #acc<data_clause acc_present>, name = "arg2"}
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {dataClause = #acc<data_clause acc_present>, name = "arg1"}
 }
diff --git a/clang/test/CIR/CodeGenOpenACC/parallel.c b/clang/test/CIR/CodeGenOpenACC/parallel.c
@@ -496,5 +496,23 @@ void acc_parallel_data_clauses(int *arg1, int *arg2) {
   // CHECK-NEXT: } loc
   // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<radeon>]) {dataClause = #acc<data_clause acc_no_create>, name = "arg2"}
   // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<radeon>]) {dataClause = #acc<data_clause acc_no_create>, name = "arg1"}
+
+#pragma acc parallel present(arg1)
+  ;
+  // CHECK-NEXT: %[[NOCREATE1:.*]] = acc.present varPtr(%[[ARG1]] : !cir.ptr<!cir.ptr<!s32i>>) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg1"}
+  // CHECK-NEXT: acc.parallel dataOperands(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) {
+  // CHECK-NEXT: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) {dataClause = #acc<data_clause acc_present>, name = "arg1"}
+
+#pragma acc parallel present(arg1, arg2) device_type(radeon) async
+  ;
+  // CHECK-NEXT: %[[NOCREATE1:.*]] = acc.present varPtr(%[[ARG1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<radeon>]) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg1"}
+  // CHECK-NEXT: %[[NOCREATE2:.*]] = acc.present varPtr(%[[ARG2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<radeon>]) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg2"}
+  // CHECK-NEXT: acc.parallel dataOperands(%[[NOCREATE1]], %[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>, !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<radeon>]) {
+  // CHECK-NEXT: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<radeon>]) {dataClause = #acc<data_clause acc_present>, name = "arg2"}
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<radeon>]) {dataClause = #acc<data_clause acc_present>, name = "arg1"}
 }
 
diff --git a/clang/test/CIR/CodeGenOpenACC/serial.c b/clang/test/CIR/CodeGenOpenACC/serial.c
@@ -319,4 +319,21 @@ void acc_serial_data_clauses(int *arg1, int *arg2) {
   // CHECK-NEXT: } loc
   // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {dataClause = #acc<data_clause acc_no_create>, name = "arg2"}
   // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {dataClause = #acc<data_clause acc_no_create>, name = "arg1"}
+
+#pragma acc serial present(arg1)
+  ;
+  // CHECK-NEXT: %[[NOCREATE1:.*]] = acc.present varPtr(%[[ARG1]] : !cir.ptr<!cir.ptr<!s32i>>) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg1"}
+  // CHECK-NEXT: acc.serial dataOperands(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) {
+  // CHECK-NEXT: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) {dataClause = #acc<data_clause acc_present>, name = "arg1"}
+#pragma acc serial present(arg1, arg2) device_type(nvidia) async
+  ;
+  // CHECK-NEXT: %[[NOCREATE1:.*]] = acc.present varPtr(%[[ARG1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg1"}
+  // CHECK-NEXT: %[[NOCREATE2:.*]] = acc.present varPtr(%[[ARG2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) -> !cir.ptr<!cir.ptr<!s32i>> {name = "arg2"}
+  // CHECK-NEXT: acc.serial dataOperands(%[[NOCREATE1]], %[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>, !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {
+  // CHECK-NEXT: acc.yield
+  // CHECK-NEXT: } loc
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE2]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {dataClause = #acc<data_clause acc_present>, name = "arg2"}
+  // CHECK-NEXT: acc.delete accPtr(%[[NOCREATE1]] : !cir.ptr<!cir.ptr<!s32i>>) async([#acc.device_type<nvidia>]) {dataClause = #acc<data_clause acc_present>, name = "arg1"}
 }
