[OPENMP, NVPTX] Add check for SPMD mode in orphaned parallel directives.

If the orphaned directive is executed in SPMD mode, we need to emit the
check for the SPMD mode and run the orphaned parallel directive in
sequential mode.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@332467 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp b/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
index 1f5bfee41bd90e3aff6999ebe5616c19a55662d5..559a26f6f357101dbb9602d660d7f2468fc3ae74 100644 (file)
--- a/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ b/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -96,6 +96,8 @@ enum OpenMPRTLFunctionNVPTX {
   /// Call to uint16_t __kmpc_parallel_level(ident_t *loc, kmp_int32
   /// global_tid);
   OMPRTL_NVPTX__kmpc_parallel_level,
+  /// Call to int8_t __kmpc_is_spmd_exec_mode();
+  OMPRTL_NVPTX__kmpc_is_spmd_exec_mode,
 };
 
 /// Pre(post)-action for different OpenMP constructs specialized for NVPTX.
@@ -220,8 +222,7 @@ class CheckVarsEscapingDeclContext final
                "Parameter captured by value with variably modified type");
         EscapedParameters.insert(VD);
       }
-    } else if (VD->getType()->isAnyPointerType() ||
-               VD->getType()->isReferenceType())
+    } else if (VD->getType()->isReferenceType())
       // Do not globalize variables with reference or pointer type.
       return;
     if (VD->getType()->isVariablyModifiedType())
@@ -317,8 +318,18 @@ public:
       return;
     if (D->hasAssociatedStmt()) {
       if (const auto *S =
-              dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt()))
+              dyn_cast_or_null<CapturedStmt>(D->getAssociatedStmt())) {
+        // Do not analyze directives that do not actually require capturing,
+        // like `omp for` or `omp simd` directives.
+        llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions;
+        getOpenMPCaptureRegions(CaptureRegions, D->getDirectiveKind());
+        if (CaptureRegions.size() == 1 &&
+            CaptureRegions.back() == OMPD_unknown) {
+          VisitStmt(S->getCapturedStmt());
+          return;
+        }
         VisitOpenMPCapturedStmt(S);
+      }
     }
   }
   void VisitCapturedStmt(const CapturedStmt *S) {
@@ -1411,6 +1422,12 @@ CGOpenMPRuntimeNVPTX::createNVPTXRuntimeFunction(unsigned Function) {
     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_parallel_level");
     break;
   }
+  case OMPRTL_NVPTX__kmpc_is_spmd_exec_mode: {
+    // Build int8_t __kmpc_is_spmd_exec_mode();
+    auto *FnTy = llvm::FunctionType::get(CGM.Int8Ty, /*isVarArg=*/false);
+    RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_is_spmd_exec_mode");
+    break;
+  }
   }
   return RTLFn;
 }
@@ -1828,7 +1845,9 @@ void CGOpenMPRuntimeNVPTX::emitNonSPMDParallelCall(
       RCG(CGF);
     } else {
       // Check for master and then parallelism:
-      // if (is_master) {
+      // if (__kmpc_is_spmd_exec_mode()) {
+      //  Serialized execution.
+      // } else if (is_master) {
       //   Worker call.
       // } else if (__kmpc_parallel_level(loc, gtid)) {
       //   Serialized execution.
@@ -1837,13 +1856,22 @@ void CGOpenMPRuntimeNVPTX::emitNonSPMDParallelCall(
       // }
       CGBuilderTy &Bld = CGF.Builder;
       llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".exit");
+      llvm::BasicBlock *SPMDCheckBB = CGF.createBasicBlock(".spmdcheck");
       llvm::BasicBlock *MasterCheckBB = CGF.createBasicBlock(".mastercheck");
       llvm::BasicBlock *ParallelCheckBB =
           CGF.createBasicBlock(".parallelcheck");
+      llvm::Value *IsSPMD = Bld.CreateIsNotNull(CGF.EmitNounwindRuntimeCall(
+          createNVPTXRuntimeFunction(OMPRTL_NVPTX__kmpc_is_spmd_exec_mode)));
+      Bld.CreateCondBr(IsSPMD, SPMDCheckBB, MasterCheckBB);
+      CGF.EmitBlock(SPMDCheckBB);
+      SeqGen(CGF, Action);
+      CGF.EmitBranch(ExitBB);
+      CGF.EmitBlock(MasterCheckBB);
+      llvm::BasicBlock *MasterThenBB = CGF.createBasicBlock("master.then");
       llvm::Value *IsMaster =
           Bld.CreateICmpEQ(getNVPTXThreadID(CGF), getMasterThreadID(CGF));
-      Bld.CreateCondBr(IsMaster, MasterCheckBB, ParallelCheckBB);
-      CGF.EmitBlock(MasterCheckBB);
+      Bld.CreateCondBr(IsMaster, MasterThenBB, ParallelCheckBB);
+      CGF.EmitBlock(MasterThenBB);
       L0ParallelGen(CGF, Action);
       CGF.EmitBranch(ExitBB);
       // There is no need to emit line number for unconditional branch.

diff --git a/test/OpenMP/nvptx_target_codegen.cpp b/test/OpenMP/nvptx_target_codegen.cpp
index 0a8e56106ed7a68616f089c5c70e08cef3cc8ba0..23dd23e86cd181951610f7dc56757b3cc972b372 100644 (file)
--- a/test/OpenMP/nvptx_target_codegen.cpp
+++ b/test/OpenMP/nvptx_target_codegen.cpp
@@ -561,6 +561,16 @@ int baz(int f, double &a) {
   // CHECK: [[F_PTR:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[REC_ADDR]], i32 0, i32 0
   // CHECK: store i32 %{{.+}}, i32* [[F_PTR]],
   // CHECK: store i32 [[GTID]], i32* [[GTID_ADDR]],
+
+  // CHECK: [[RES:%.+]] = call i8 @__kmpc_is_spmd_exec_mode()
+  // CHECK: icmp ne i8 [[RES]], 0
+  // CHECK: br i1
+
+  // CHECK: call void @__kmpc_serialized_parallel(%struct.ident_t* @{{.+}}, i32 [[GTID]])
+  // CHECK: call void [[OUTLINED:@.+]](i32* [[GTID_ADDR]], i32* [[ZERO_ADDR]], i32* [[F_PTR]], double* %{{.+}})
+  // CHECK: call void @__kmpc_end_serialized_parallel(%struct.ident_t* @{{.+}}, i32 [[GTID]])
+  // CHECK: br label
+
   // CHECK: icmp eq i32
   // CHECK: br i1