[OPENMP] Captured arguments of the capturable clauses by value.

If the clause is applied to the combined construct and has captured
expression, try to capture this expression by value rather than by
reference.

llvm-svn: 321386

4 files changed, 49 insertions, 26 deletions

diff --git a/clang/lib/Sema/SemaOpenMP.cpp b/clang/lib/Sema/SemaOpenMP.cpp
index b34bb3388d7..0880b2d7906 100644
--- a/clang/lib/Sema/SemaOpenMP.cpp
+++ b/clang/lib/Sema/SemaOpenMP.cpp
@@ -1290,9 +1290,14 @@ bool Sema::IsOpenMPCapturedByRef(ValueDecl *D, unsigned Level) {
   }
 
   if (IsByRef && Ty.getNonReferenceType()->isScalarType()) {
-    IsByRef = !DSAStack->hasExplicitDSA(
-        D, [](OpenMPClauseKind K) -> bool { return K == OMPC_firstprivate; },
-        Level, /*NotLastprivate=*/true);
+    IsByRef =
+        !DSAStack->hasExplicitDSA(
+            D,
+            [](OpenMPClauseKind K) -> bool { return K == OMPC_firstprivate; },
+            Level, /*NotLastprivate=*/true) &&
+        // If the variable is artificial and must be captured by value - try to
+        // capture by value.
+        !(isa<OMPCapturedExprDecl>(D) && D->hasAttr<OMPCaptureKindAttr>());
   }
 
   // When passing data by copy, we need to make sure it fits the uintptr size
@@ -2321,10 +2326,11 @@ static OMPCapturedExprDecl *buildCaptureDecl(Sema &S, IdentifierInfo *Id,
   ASTContext &C = S.getASTContext();
   Expr *Init = AsExpression ? CaptureExpr : CaptureExpr->IgnoreImpCasts();
   QualType Ty = Init->getType();
+  Attr *OMPCaptureKind = nullptr;
   if (CaptureExpr->getObjectKind() == OK_Ordinary && CaptureExpr->isGLValue()) {
-    if (S.getLangOpts().CPlusPlus)
+    if (S.getLangOpts().CPlusPlus) {
       Ty = C.getLValueReferenceType(Ty);
-    else {
+    } else {
       Ty = C.getPointerType(Ty);
       ExprResult Res =
           S.CreateBuiltinUnaryOp(CaptureExpr->getExprLoc(), UO_AddrOf, Init);
@@ -2333,11 +2339,16 @@ static OMPCapturedExprDecl *buildCaptureDecl(Sema &S, IdentifierInfo *Id,
       Init = Res.get();
     }
     WithInit = true;
+  } else if (AsExpression) {
+    // This variable must be captured by value.
+    OMPCaptureKind = OMPCaptureKindAttr::CreateImplicit(C, OMPC_unknown);
   }
   auto *CED = OMPCapturedExprDecl::Create(C, S.CurContext, Id, Ty,
                                           CaptureExpr->getLocStart());
   if (!WithInit)
     CED->addAttr(OMPCaptureNoInitAttr::CreateImplicit(C, SourceRange()));
+  if (OMPCaptureKind)
+    CED->addAttr(OMPCaptureKind);
   S.CurContext->addHiddenDecl(CED);
   S.AddInitializerToDecl(CED, Init, /*DirectInit=*/false);
   return CED;
@@ -2346,31 +2357,34 @@ static OMPCapturedExprDecl *buildCaptureDecl(Sema &S, IdentifierInfo *Id,
 static DeclRefExpr *buildCapture(Sema &S, ValueDecl *D, Expr *CaptureExpr,
                                  bool WithInit) {
   OMPCapturedExprDecl *CD;
-  if (auto *VD = S.IsOpenMPCapturedDecl(D))
+  if (auto *VD = S.IsOpenMPCapturedDecl(D)) {
     CD = cast<OMPCapturedExprDecl>(VD);
-  else
+  } else {
     CD = buildCaptureDecl(S, D->getIdentifier(), CaptureExpr, WithInit,
                           /*AsExpression=*/false);
+  }
   return buildDeclRefExpr(S, CD, CD->getType().getNonReferenceType(),
                           CaptureExpr->getExprLoc());
 }
 
 static ExprResult buildCapture(Sema &S, Expr *CaptureExpr, DeclRefExpr *&Ref) {
+  CaptureExpr = S.DefaultLvalueConversion(CaptureExpr).get();
   if (!Ref) {
-    auto *CD =
-        buildCaptureDecl(S, &S.getASTContext().Idents.get(".capture_expr."),
-                         CaptureExpr, /*WithInit=*/true, /*AsExpression=*/true);
+    OMPCapturedExprDecl *CD = buildCaptureDecl(
+        S, &S.getASTContext().Idents.get(".capture_expr."), CaptureExpr,
+        /*WithInit=*/true, /*AsExpression=*/true);
     Ref = buildDeclRefExpr(S, CD, CD->getType().getNonReferenceType(),
                            CaptureExpr->getExprLoc());
   }
   ExprResult Res = Ref;
   if (!S.getLangOpts().CPlusPlus &&
       CaptureExpr->getObjectKind() == OK_Ordinary && CaptureExpr->isGLValue() &&
-      Ref->getType()->isPointerType())
+      Ref->getType()->isPointerType()) {
     Res = S.CreateBuiltinUnaryOp(CaptureExpr->getExprLoc(), UO_Deref, Ref);
-  if (!Res.isUsable())
-    return ExprError();
-  return CaptureExpr->isGLValue() ? Res : S.DefaultLvalueConversion(Res.get());
+    if (!Res.isUsable())
+      return ExprError();
+  }
+  return S.DefaultLvalueConversion(Res.get());
 }
 
 namespace {
@@ -8117,12 +8131,13 @@ OMPClause *Sema::ActOnOpenMPIfClause(OpenMPDirectiveKind NameModifier,
     if (Val.isInvalid())
       return nullptr;
 
-    ValExpr = MakeFullExpr(Val.get()).get();
+    ValExpr = Val.get();
 
     OpenMPDirectiveKind DKind = DSAStack->getCurrentDirective();
     CaptureRegion =
         getOpenMPCaptureRegionForClause(DKind, OMPC_if, NameModifier);
     if (CaptureRegion != OMPD_unknown && !CurContext->isDependentContext()) {
+      ValExpr = MakeFullExpr(ValExpr).get();
       llvm::MapVector<Expr *, DeclRefExpr *> Captures;
       ValExpr = tryBuildCapture(*this, ValExpr, Captures).get();
       HelperValStmt = buildPreInits(Context, Captures);
@@ -8239,6 +8254,7 @@ OMPClause *Sema::ActOnOpenMPNumThreadsClause(Expr *NumThreads,
   OpenMPDirectiveKind CaptureRegion =
       getOpenMPCaptureRegionForClause(DKind, OMPC_num_threads);
   if (CaptureRegion != OMPD_unknown && !CurContext->isDependentContext()) {
+    ValExpr = MakeFullExpr(ValExpr).get();
     llvm::MapVector<Expr *, DeclRefExpr *> Captures;
     ValExpr = tryBuildCapture(*this, ValExpr, Captures).get();
     HelperValStmt = buildPreInits(Context, Captures);
@@ -8666,6 +8682,7 @@ OMPClause *Sema::ActOnOpenMPScheduleClause(
                      DSAStack->getCurrentDirective(), OMPC_schedule) !=
                      OMPD_unknown &&
                  !CurContext->isDependentContext()) {
+        ValExpr = MakeFullExpr(ValExpr).get();
         llvm::MapVector<Expr *, DeclRefExpr *> Captures;
         ValExpr = tryBuildCapture(*this, ValExpr, Captures).get();
         HelperValStmt = buildPreInits(Context, Captures);
@@ -11355,6 +11372,7 @@ OMPClause *Sema::ActOnOpenMPDeviceClause(Expr *Device, SourceLocation StartLoc,
   OpenMPDirectiveKind CaptureRegion =
       getOpenMPCaptureRegionForClause(DKind, OMPC_device);
   if (CaptureRegion != OMPD_unknown && !CurContext->isDependentContext()) {
+    ValExpr = MakeFullExpr(ValExpr).get();
     llvm::MapVector<Expr *, DeclRefExpr *> Captures;
     ValExpr = tryBuildCapture(*this, ValExpr, Captures).get();
     HelperValStmt = buildPreInits(Context, Captures);
@@ -12378,6 +12396,7 @@ OMPClause *Sema::ActOnOpenMPNumTeamsClause(Expr *NumTeams,
   OpenMPDirectiveKind CaptureRegion =
       getOpenMPCaptureRegionForClause(DKind, OMPC_num_teams);
   if (CaptureRegion != OMPD_unknown && !CurContext->isDependentContext()) {
+    ValExpr = MakeFullExpr(ValExpr).get();
     llvm::MapVector<Expr *, DeclRefExpr *> Captures;
     ValExpr = tryBuildCapture(*this, ValExpr, Captures).get();
     HelperValStmt = buildPreInits(Context, Captures);
@@ -12404,6 +12423,7 @@ OMPClause *Sema::ActOnOpenMPThreadLimitClause(Expr *ThreadLimit,
   OpenMPDirectiveKind CaptureRegion =
       getOpenMPCaptureRegionForClause(DKind, OMPC_thread_limit);
   if (CaptureRegion != OMPD_unknown && !CurContext->isDependentContext()) {
+    ValExpr = MakeFullExpr(ValExpr).get();
     llvm::MapVector<Expr *, DeclRefExpr *> Captures;
     ValExpr = tryBuildCapture(*this, ValExpr, Captures).get();
     HelperValStmt = buildPreInits(Context, Captures);
@@ -12514,6 +12534,7 @@ OMPClause *Sema::ActOnOpenMPDistScheduleClause(
                      DSAStack->getCurrentDirective(), OMPC_dist_schedule) !=
                      OMPD_unknown &&
                  !CurContext->isDependentContext()) {
+        ValExpr = MakeFullExpr(ValExpr).get();
         llvm::MapVector<Expr *, DeclRefExpr *> Captures;
         ValExpr = tryBuildCapture(*this, ValExpr, Captures).get();
         HelperValStmt = buildPreInits(Context, Captures);
diff --git a/clang/test/OpenMP/parallel_for_codegen.cpp b/clang/test/OpenMP/parallel_for_codegen.cpp
index 1773619bce2..db3a2de580d 100644
--- a/clang/test/OpenMP/parallel_for_codegen.cpp
+++ b/clang/test/OpenMP/parallel_for_codegen.cpp
@@ -15,10 +15,12 @@ void with_var_schedule() {
   double a = 5;
 // CHECK: [[CHUNK_SIZE:%.+]] = fptosi double %{{.+}}to i8
 // CHECK: store i8 %{{.+}}, i8* [[CHUNK:%.+]],
-// CHECK: call void {{.+}} @__kmpc_fork_call({{.+}}, i8* [[CHUNK]])
+// CHECK: [[VAL:%.+]] = load i8, i8* [[CHUNK]],
+// CHECK: store i8 [[VAL]], i8*
+// CHECK: [[CHUNK:%.+]] = load i64, i64* %
+// CHECK: call void {{.+}} @__kmpc_fork_call({{.+}}, i64 [[CHUNK]])
 
-// CHECK: [[CHUNK:%.+]] = load i8*, i8** %
-// CHECK: [[CHUNK_VAL:%.+]] = load i8, i8* [[CHUNK]],
+// CHECK: [[CHUNK_VAL:%.+]] = load i8, i8* %
 // CHECK: [[CHUNK_SIZE:%.+]] = sext i8 [[CHUNK_VAL]] to i64
 // CHECK: call void @__kmpc_for_static_init_8u([[IDENT_T_TY]]* [[LOOP_LOC]], i32 [[GTID:%[^,]+]], i32 33, i32* [[IS_LAST:%[^,]+]], i64* [[OMP_LB:%[^,]+]], i64* [[OMP_UB:%[^,]+]], i64* [[OMP_ST:%[^,]+]], i64 1, i64 [[CHUNK_SIZE]])
 // CHECK: call void @__kmpc_for_static_fini([[IDENT_T_TY]]* [[LOOP_LOC]], i32 [[GTID]])
diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_num_threads_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_num_threads_codegen.cpp
index 7c0ac461220..304cc8b3f4a 100644
--- a/clang/test/OpenMP/teams_distribute_parallel_for_num_threads_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_num_threads_codegen.cpp
@@ -60,8 +60,8 @@ int main() {
 #pragma omp teams distribute parallel for num_threads(a)
   for (int i = 0; i < 100; i++) {
     // CHECK: define{{.+}} void [[OMP_TEAMS_OUTLINED_1]](
-    // CHECK-DAG: [[A_ADDR:%.+]] = alloca i8*,
-    // CHECK-DAG: [[A_REF:%.+]] = load i8*, i8** [[A_ADDR]],
+    // CHECK-DAG: [[A_ADDR:%.+]] = alloca i64,
+    // CHECK-DAG: [[A_REF:%.+]] = bitcast i64* [[A_ADDR]] to i8*
     // CHECK-DAG: [[A_VAL:%.+]] = load i8, i8* [[A_REF]],
     // CHECK-DAG: [[A_EXT:%.+]] = sext i8 [[A_VAL]] to {{.+}}
     // CHECK: call {{.*}}void @__kmpc_push_num_threads([[IDENT_T_TY]]* [[DEF_LOC_2]], i32 {{.+}}, i32 [[A_EXT]])
@@ -110,9 +110,9 @@ int main() {
 // CHECK: call {{.*}}void {{.+}} @__kmpc_fork_teams({{.+}}, i32 {{.+}}, {{.+}}* [[T_OMP_TEAMS_OUTLINED_3:@.+]] to {{.+}})
 
 // CHECK: define{{.+}} void [[T_OMP_TEAMS_OUTLINED_3]]({{.+}}, {{.+}}, {{.+}} [[NUM_TH_CPT_IN:%.+]])
-// CHECK: [[NUM_TH_CPT:%.+]] = alloca i8*,
+// CHECK: [[NUM_TH_CPT:%.+]] = alloca i64,
 // CHECK: store {{.+}} [[NUM_TH_CPT_IN]], {{.+}} [[NUM_TH_CPT]],
-// CHECK: [[NUM_TH_REF:%.+]] = load{{.+}}, {{.+}} [[NUM_TH_CPT]],
+// CHECK: [[NUM_TH_REF:%.+]] = bitcast i64* [[NUM_TH_CPT]] to i8*
 // CHECK-DAG:   [[NUM_TH_VAL:%.+]] = load {{.+}}, {{.+}} [[NUM_TH_REF]],
 // CHECK-DAG:   [[NUM_TH_SEXT:%.+]] = sext i8 [[NUM_TH_VAL]] to {{.+}}
 // CHECK:       call {{.*}}void @__kmpc_push_num_threads([[IDENT_T_TY]]* [[DEF_LOC_2]], i32 {{.+}}, i32 [[NUM_TH_SEXT]])
diff --git a/clang/test/OpenMP/teams_distribute_parallel_for_simd_num_threads_codegen.cpp b/clang/test/OpenMP/teams_distribute_parallel_for_simd_num_threads_codegen.cpp
index 93051af6161..2576241725c 100644
--- a/clang/test/OpenMP/teams_distribute_parallel_for_simd_num_threads_codegen.cpp
+++ b/clang/test/OpenMP/teams_distribute_parallel_for_simd_num_threads_codegen.cpp
@@ -60,8 +60,8 @@ int main() {
 #pragma omp teams distribute parallel for simd num_threads(a)
   for (int i = 0; i < 100; i++) {
     // CHECK: define{{.+}} void [[OMP_TEAMS_OUTLINED_1]](
-    // CHECK-DAG: [[A_ADDR:%.+]] = alloca i8*,
-    // CHECK-DAG: [[A_REF:%.+]] = load i8*, i8** [[A_ADDR]],
+    // CHECK-DAG: [[A_ADDR:%.+]] = alloca i64,
+    // CHECK-DAG: [[A_REF:%.+]] = bitcast i64* [[A_ADDR]] to i8*
     // CHECK-DAG: [[A_VAL:%.+]] = load i8, i8* [[A_REF]],
     // CHECK-DAG: [[A_EXT:%.+]] = sext i8 [[A_VAL]] to {{.+}}
     // CHECK: call {{.*}}void @__kmpc_push_num_threads([[IDENT_T_TY]]* [[DEF_LOC_2]], i32 {{.+}}, i32 [[A_EXT]])
@@ -110,9 +110,9 @@ int main() {
 // CHECK: call {{.*}}void {{.+}} @__kmpc_fork_teams({{.+}}, i32 {{.+}}, {{.+}}* [[T_OMP_TEAMS_OUTLINED_3:@.+]] to {{.+}})
 
 // CHECK: define{{.+}} void [[T_OMP_TEAMS_OUTLINED_3]]({{.+}}, {{.+}}, {{.+}} [[NUM_TH_CPT_IN:%.+]])
-// CHECK: [[NUM_TH_CPT:%.+]] = alloca i8*,
+// CHECK: [[NUM_TH_CPT:%.+]] = alloca i64,
 // CHECK: store {{.+}} [[NUM_TH_CPT_IN]], {{.+}} [[NUM_TH_CPT]],
-// CHECK: [[NUM_TH_REF:%.+]] = load{{.+}}, {{.+}} [[NUM_TH_CPT]],
+// CHECK: [[NUM_TH_REF:%.+]] = bitcast i64* [[NUM_TH_CPT]] to i8*
 // CHECK-DAG:   [[NUM_TH_VAL:%.+]] = load {{.+}}, {{.+}} [[NUM_TH_REF]],
 // CHECK-DAG:   [[NUM_TH_SEXT:%.+]] = sext i8 [[NUM_TH_VAL]] to {{.+}}
 // CHECK:       call {{.*}}void @__kmpc_push_num_threads([[IDENT_T_TY]]* [[DEF_LOC_2]], i32 {{.+}}, i32 [[NUM_TH_SEXT]])