[OPENMP][NVPTX]Reduce memory use for globalized vars in

target/teams/distribute regions.

Previously introduced globalization scheme that uses memory coalescing
scheme may increase memory usage fr the variables that are devlared in
target/teams/distribute contexts. We don't need 32 copies of such
variables, just 1. Patch reduces memory use in this case.

llvm-svn: 344273

2 files changed, 17 insertions, 12 deletions

diff --git a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
index 04de3956622..944130b1abf 100644
--- a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -203,7 +203,8 @@ static RecordDecl *buildRecordForGlobalizedVars(
   std::stable_sort(GlobalizedVars.begin(), GlobalizedVars.end(),
                    stable_sort_comparator);
   // Build struct _globalized_locals_ty {
-  //         /*  globalized vars  */[32] align (max(decl_align, 128))
+  //         /*  globalized vars  */[WarSize] align (max(decl_align,
+  //         GlobalMemoryAlignment))
   //         /*  globalized vars  */ for EscapedDeclsForTeams
   //       };
   RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty");
@@ -370,11 +371,16 @@ class CheckVarsEscapingDeclContext final
     }
   }
 
-  void buildRecordForGlobalizedVars() {
+  void buildRecordForGlobalizedVars(bool IsInTargetMasterThreadRegion) {
     assert(!GlobalizedRD &&
            "Record for globalized variables is built already.");
+    ArrayRef<const ValueDecl *> EscapedDeclsForParallel, EscapedDeclsForTeams;
+    if (IsInTargetMasterThreadRegion)
+      EscapedDeclsForTeams = EscapedDecls.getArrayRef();
+    else
+      EscapedDeclsForParallel = EscapedDecls.getArrayRef();
     GlobalizedRD = ::buildRecordForGlobalizedVars(
-        CGF.getContext(), EscapedDecls.getArrayRef(), llvm::None,
+        CGF.getContext(), EscapedDeclsForParallel, EscapedDeclsForTeams,
         MappedDeclsFields);
   }
 
@@ -521,9 +527,9 @@ public:
 
   /// Returns the record that handles all the escaped local variables and used
   /// instead of their original storage.
-  const RecordDecl *getGlobalizedRecord() {
+  const RecordDecl *getGlobalizedRecord(bool IsInTargetMasterThreadRegion) {
     if (!GlobalizedRD)
-      buildRecordForGlobalizedVars();
+      buildRecordForGlobalizedVars(IsInTargetMasterThreadRegion);
     return GlobalizedRD;
   }
 
@@ -4087,7 +4093,8 @@ void CGOpenMPRuntimeNVPTX::emitFunctionProlog(CodeGenFunction &CGF,
     return;
   CheckVarsEscapingDeclContext VarChecker(CGF);
   VarChecker.Visit(Body);
-  const RecordDecl *GlobalizedVarsRecord = VarChecker.getGlobalizedRecord();
+  const RecordDecl *GlobalizedVarsRecord =
+      VarChecker.getGlobalizedRecord(IsInTargetMasterThreadRegion);
   ArrayRef<const ValueDecl *> EscapedVariableLengthDecls =
       VarChecker.getEscapedVariableLengthDecls();
   if (!GlobalizedVarsRecord && EscapedVariableLengthDecls.empty())
@@ -4105,7 +4112,8 @@ void CGOpenMPRuntimeNVPTX::emitFunctionProlog(CodeGenFunction &CGF,
   for (const ValueDecl *VD : VarChecker.getEscapedDecls()) {
     assert(VD->isCanonicalDecl() && "Expected canonical declaration");
     const FieldDecl *FD = VarChecker.getFieldForGlobalizedVar(VD);
-    Data.insert(std::make_pair(VD, MappedVarData(FD)));
+    Data.insert(
+        std::make_pair(VD, MappedVarData(FD, IsInTargetMasterThreadRegion)));
   }
   if (!NeedToDelayGlobalization) {
     emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true);
diff --git a/clang/test/OpenMP/nvptx_distribute_parallel_generic_mode_codegen.cpp b/clang/test/OpenMP/nvptx_distribute_parallel_generic_mode_codegen.cpp
index 2a7344f7734..2b18f6d3f9b 100644
--- a/clang/test/OpenMP/nvptx_distribute_parallel_generic_mode_codegen.cpp
+++ b/clang/test/OpenMP/nvptx_distribute_parallel_generic_mode_codegen.cpp
@@ -26,13 +26,10 @@ int main(int argc, char **argv) {
 // CHECK-LABEL: define internal void @__omp_offloading_{{.*}}_main_l17_worker(
 
 // CHECK: define weak void @__omp_offloading_{{.*}}_main_l17([10 x i32]* dereferenceable(40) %{{.+}}, [10 x i32]* dereferenceable(40) %{{.+}}, i32* dereferenceable(4) %{{.+}}, i{{64|32}} %{{.+}}, [10 x i32]* dereferenceable(40) %{{.+}})
-// CHECK: [[PTR:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{64|32}} 2688, i16 0)
+// CHECK: [[PTR:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{64|32}} 84, i16 0)
 // CHECK: [[STACK:%.+]] = bitcast i8* [[PTR]] to %struct._globalized_locals_ty*
 // CHECK: [[ARGC:%.+]] = load i32, i32* %{{.+}}, align
-// CHECK: [[ARGC_ARR_ADDR:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[STACK]], i{{32|64}} 0, i{{32|64}} 0
-// CHECK: [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
-// CHECK: [[LID:%.+]] = and i32 [[TID]], 31
-// CHECK: [[ARGC_ADDR:%.+]] = getelementptr inbounds [32 x i32], [32 x i32]* [[ARGC_ARR_ADDR]], i32 0, i32 [[LID]]
+// CHECK: [[ARGC_ADDR:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[STACK]], i{{32|64}} 0, i{{32|64}} 0
 // CHECK: store i32 [[ARGC]], i32* [[ARGC_ADDR]],
 // CHECK: getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[STACK]], i{{32|64}} 0, i{{32|64}} 1
 // CHECK: getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[STACK]], i{{32|64}} 0, i{{32|64}} 2