[OPENMP][NVPTX] Handle `requires datasharing` flag correctly with

lightweight runtime.

The datasharing flag must be set to `1` when executing SPMD-mode compatible directive with reduction|lastprivate clauses.

llvm-svn: 343492

4 files changed, 30 insertions, 4 deletions

diff --git a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
index 833b145f045..96cf6c1173b 100644
--- a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -1207,6 +1207,10 @@ void CGOpenMPRuntimeNVPTX::emitSPMDKernel(const OMPExecutableDirective &D,
                                    IsOffloadEntry, CodeGen);
 }
 
+static void
+getDistributeLastprivateVars(const OMPExecutableDirective &D,
+                             llvm::SmallVectorImpl<const ValueDecl *> &Vars);
+
 void CGOpenMPRuntimeNVPTX::emitSPMDEntryHeader(
     CodeGenFunction &CGF, EntryFunctionState &EST,
     const OMPExecutableDirective &D) {
@@ -1219,11 +1223,33 @@ void CGOpenMPRuntimeNVPTX::emitSPMDEntryHeader(
   // Initialize the OMP state in the runtime; called by all active threads.
   bool RequiresFullRuntime = CGM.getLangOpts().OpenMPCUDAForceFullRuntime ||
                              !supportsLightweightRuntime(CGF.getContext(), D);
+  // Check if we have inner distribute + lastprivate|reduction clauses.
+  bool RequiresDatasharing = RequiresFullRuntime;
+  if (!RequiresDatasharing) {
+    const OMPExecutableDirective *TD = &D;
+    if (!isOpenMPTeamsDirective(TD->getDirectiveKind()) &&
+        !isOpenMPParallelDirective(TD->getDirectiveKind())) {
+      const Stmt *S = getSingleCompoundChild(
+          TD->getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
+              /*IgnoreCaptured=*/true));
+      TD = cast<OMPExecutableDirective>(S);
+    }
+    if (!isOpenMPDistributeDirective(TD->getDirectiveKind()) &&
+        !isOpenMPParallelDirective(TD->getDirectiveKind())) {
+      const Stmt *S = getSingleCompoundChild(
+          TD->getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(
+              /*IgnoreCaptured=*/true));
+      TD = cast<OMPExecutableDirective>(S);
+    }
+    if (isOpenMPDistributeDirective(TD->getDirectiveKind()))
+      RequiresDatasharing = TD->hasClausesOfKind<OMPLastprivateClause>() ||
+                            TD->hasClausesOfKind<OMPReductionClause>();
+  }
   llvm::Value *Args[] = {
       getThreadLimit(CGF, /*IsInSPMDExecutionMode=*/true),
       /*RequiresOMPRuntime=*/
       Bld.getInt16(RequiresFullRuntime ? 1 : 0),
-      /*RequiresDataSharing=*/Bld.getInt16(RequiresFullRuntime ? 1 : 0)};
+      /*RequiresDataSharing=*/Bld.getInt16(RequiresDatasharing ? 1 : 0)};
   CGF.EmitRuntimeCall(
       createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_spmd_kernel_init), Args);
 
diff --git a/clang/test/OpenMP/nvptx_SPMD_codegen.cpp b/clang/test/OpenMP/nvptx_SPMD_codegen.cpp
index c247cc34594..319b2c9d05a 100644
--- a/clang/test/OpenMP/nvptx_SPMD_codegen.cpp
+++ b/clang/test/OpenMP/nvptx_SPMD_codegen.cpp
@@ -40,7 +40,7 @@ void foo() {
   for (int i = 0; i < 10; ++i)
     ;
 int a;
-// CHECK: call void @__kmpc_spmd_kernel_init(i32 {{.+}}, i16 0, i16 0)
+// CHECK: call void @__kmpc_spmd_kernel_init(i32 {{.+}}, i16 0, i16 1)
 // CHECK: call void @__kmpc_spmd_kernel_init(i32 {{.+}}, i16 0, i16 0)
 // CHECK: call void @__kmpc_spmd_kernel_init(i32 {{.+}}, i16 0, i16 0)
 // CHECK: call void @__kmpc_spmd_kernel_init(i32 {{.+}}, i16 1, i16 {{.+}})
diff --git a/clang/test/OpenMP/nvptx_target_teams_distribute_parallel_for_codegen.cpp b/clang/test/OpenMP/nvptx_target_teams_distribute_parallel_for_codegen.cpp
index 6539b0e3c96..01c753355cd 100644
--- a/clang/test/OpenMP/nvptx_target_teams_distribute_parallel_for_codegen.cpp
+++ b/clang/test/OpenMP/nvptx_target_teams_distribute_parallel_for_codegen.cpp
@@ -69,7 +69,7 @@ int bar(int n){
 
 // CHECK-LABEL: define {{.*}}void {{@__omp_offloading_.+}}_l32(
 // CHECK-DAG: [[THREAD_LIMIT:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.ntid.x()
-// CHECK: call void @__kmpc_spmd_kernel_init(i32 [[THREAD_LIMIT]], i16 0, i16 0)
+// CHECK: call void @__kmpc_spmd_kernel_init(i32 [[THREAD_LIMIT]], i16 0, i16 1)
 // CHECK: [[TEAM_ALLOC:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} 4, i16 0)
 // CHECK: [[BC:%.+]] = bitcast i8* [[TEAM_ALLOC]] to [[REC:%.+]]*
 // CHECK: getelementptr inbounds [[REC]], [[REC]]* [[BC]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
diff --git a/clang/test/OpenMP/nvptx_target_teams_distribute_parallel_for_simd_codegen.cpp b/clang/test/OpenMP/nvptx_target_teams_distribute_parallel_for_simd_codegen.cpp
index 5f7c071140c..1555073cd0b 100644
--- a/clang/test/OpenMP/nvptx_target_teams_distribute_parallel_for_simd_codegen.cpp
+++ b/clang/test/OpenMP/nvptx_target_teams_distribute_parallel_for_simd_codegen.cpp
@@ -64,7 +64,7 @@ int bar(int n){
 
 // CHECK-LABEL: define {{.*}}void {{@__omp_offloading_.+}}_l30(
 // CHECK-DAG: [[THREAD_LIMIT:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.ntid.x()
-// CHECK: call void @__kmpc_spmd_kernel_init(i32 [[THREAD_LIMIT]], i16 0, i16 0)
+// CHECK: call void @__kmpc_spmd_kernel_init(i32 [[THREAD_LIMIT]], i16 0, i16 1)
 // CHECK: [[TEAM_ALLOC:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} 4, i16 0)
 // CHECK: [[BC:%.+]] = bitcast i8* [[TEAM_ALLOC]] to [[REC:%.+]]*
 // CHECK: getelementptr inbounds [[REC]], [[REC]]* [[BC]], i{{[0-9]+}} 0, i{{[0-9]+}} 0