[OPENMP][NVPTX] Support memory coalescing for globalized variables.

Added support for memory coalescing for better performance for
globalized variables. From now on all the globalized variables are
represented as arrays of 32 elements and each thread accesses these
elements using `tid & 31` as index.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@344049 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp b/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
index b75891e4f21e1c477fe36c3927e4639fc7010930..04de395662226c50e8ead1228e66e63c88ea5089 100644 (file)
--- a/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ b/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -169,7 +169,7 @@ enum MachineConfiguration : unsigned {
   LaneIDMask = WarpSize - 1,
 
   /// Global memory alignment for performance.
-  GlobalMemoryAlignment = 256,
+  GlobalMemoryAlignment = 128,
 };
 
 enum NamedBarrier : unsigned {
@@ -186,20 +186,30 @@ static bool stable_sort_comparator(const VarsDataTy P1, const VarsDataTy P2) {
 
 static RecordDecl *buildRecordForGlobalizedVars(
     ASTContext &C, ArrayRef<const ValueDecl *> EscapedDecls,
+    ArrayRef<const ValueDecl *> EscapedDeclsForTeams,
     llvm::SmallDenseMap<const ValueDecl *, const FieldDecl *>
         &MappedDeclsFields) {
-  if (EscapedDecls.empty())
+  if (EscapedDecls.empty() && EscapedDeclsForTeams.empty())
     return nullptr;
   SmallVector<VarsDataTy, 4> GlobalizedVars;
   for (const ValueDecl *D : EscapedDecls)
+    GlobalizedVars.emplace_back(
+        CharUnits::fromQuantity(std::max(
+            C.getDeclAlign(D).getQuantity(),
+            static_cast<CharUnits::QuantityType>(GlobalMemoryAlignment))),
+        D);
+  for (const ValueDecl *D : EscapedDeclsForTeams)
     GlobalizedVars.emplace_back(C.getDeclAlign(D), D);
   std::stable_sort(GlobalizedVars.begin(), GlobalizedVars.end(),
                    stable_sort_comparator);
   // Build struct _globalized_locals_ty {
-  //         /*  globalized vars  */
+  //         /*  globalized vars  */[32] align (max(decl_align, 128))
+  //         /*  globalized vars  */ for EscapedDeclsForTeams
   //       };
   RecordDecl *GlobalizedRD = C.buildImplicitRecord("_globalized_locals_ty");
   GlobalizedRD->startDefinition();
+  llvm::SmallPtrSet<const ValueDecl *, 16> SingleEscaped(
+      EscapedDeclsForTeams.begin(), EscapedDeclsForTeams.end());
   for (const auto &Pair : GlobalizedVars) {
     const ValueDecl *VD = Pair.second;
     QualType Type = VD->getType();
@@ -208,19 +218,39 @@ static RecordDecl *buildRecordForGlobalizedVars(
     else
       Type = Type.getNonReferenceType();
     SourceLocation Loc = VD->getLocation();
-    auto *Field =
-        FieldDecl::Create(C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
-                          C.getTrivialTypeSourceInfo(Type, SourceLocation()),
-                          /*BW=*/nullptr, /*Mutable=*/false,
-                          /*InitStyle=*/ICIS_NoInit);
-    Field->setAccess(AS_public);
-    GlobalizedRD->addDecl(Field);
-    if (VD->hasAttrs()) {
-      for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
-           E(VD->getAttrs().end());
-           I != E; ++I)
-        Field->addAttr(*I);
+    FieldDecl *Field;
+    if (SingleEscaped.count(VD)) {
+      Field = FieldDecl::Create(
+          C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
+          C.getTrivialTypeSourceInfo(Type, SourceLocation()),
+          /*BW=*/nullptr, /*Mutable=*/false,
+          /*InitStyle=*/ICIS_NoInit);
+      Field->setAccess(AS_public);
+      if (VD->hasAttrs()) {
+        for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()),
+             E(VD->getAttrs().end());
+             I != E; ++I)
+          Field->addAttr(*I);
+      }
+    } else {
+      llvm::APInt ArraySize(32, WarpSize);
+      Type = C.getConstantArrayType(Type, ArraySize, ArrayType::Normal, 0);
+      Field = FieldDecl::Create(
+          C, GlobalizedRD, Loc, Loc, VD->getIdentifier(), Type,
+          C.getTrivialTypeSourceInfo(Type, SourceLocation()),
+          /*BW=*/nullptr, /*Mutable=*/false,
+          /*InitStyle=*/ICIS_NoInit);
+      Field->setAccess(AS_public);
+      llvm::APInt Align(32, std::max(C.getDeclAlign(VD).getQuantity(),
+                                     static_cast<CharUnits::QuantityType>(
+                                         GlobalMemoryAlignment)));
+      Field->addAttr(AlignedAttr::CreateImplicit(
+          C, AlignedAttr::GNU_aligned, /*IsAlignmentExpr=*/true,
+          IntegerLiteral::Create(C, Align,
+                                 C.getIntTypeForBitwidth(32, /*Signed=*/0),
+                                 SourceLocation())));
     }
+    GlobalizedRD->addDecl(Field);
     MappedDeclsFields.try_emplace(VD, Field);
   }
   GlobalizedRD->completeDefinition();
@@ -344,7 +374,8 @@ class CheckVarsEscapingDeclContext final
     assert(!GlobalizedRD &&
            "Record for globalized variables is built already.");
     GlobalizedRD = ::buildRecordForGlobalizedVars(
-        CGF.getContext(), EscapedDecls.getArrayRef(), MappedDeclsFields);
+        CGF.getContext(), EscapedDecls.getArrayRef(), llvm::None,
+        MappedDeclsFields);
   }
 
 public:
@@ -1849,8 +1880,7 @@ getDistributeLastprivateVars(const OMPExecutableDirective &D,
   }
   if (!Dir)
     return;
-  for (const OMPLastprivateClause *C :
-       Dir->getClausesOfKind<OMPLastprivateClause>()) {
+  for (const auto *C : Dir->getClausesOfKind<OMPLastprivateClause>()) {
     for (const Expr *E : C->getVarRefs()) {
       const auto *DE = cast<DeclRefExpr>(E->IgnoreParens());
       Vars.push_back(cast<ValueDecl>(DE->getDecl()->getCanonicalDecl()));
@@ -1869,8 +1899,8 @@ llvm::Value *CGOpenMPRuntimeNVPTX::emitTeamsOutlinedFunction(
   if (getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_SPMD) {
     getDistributeLastprivateVars(D, LastPrivates);
     if (!LastPrivates.empty())
-      GlobalizedRD = buildRecordForGlobalizedVars(
-          CGM.getContext(), LastPrivates, MappedDeclsFields);
+      GlobalizedRD = ::buildRecordForGlobalizedVars(
+          CGM.getContext(), llvm::None, LastPrivates, MappedDeclsFields);
   }
 
   // Emit target region as a standalone region.
@@ -1899,9 +1929,9 @@ llvm::Value *CGOpenMPRuntimeNVPTX::emitTeamsOutlinedFunction(
         for (const auto &Pair : MappedDeclsFields) {
           assert(Pair.getFirst()->isCanonicalDecl() &&
                  "Expected canonical declaration");
-          Data.insert(std::make_pair(
-              Pair.getFirst(),
-              std::make_pair(Pair.getSecond(), Address::invalid())));
+          Data.insert(std::make_pair(Pair.getFirst(),
+                                     MappedVarData(Pair.getSecond(),
+                                                   /*IsOnePerTeam=*/true)));
         }
       }
       Rt.emitGenericVarsProlog(CGF, Loc);
@@ -1935,18 +1965,20 @@ void CGOpenMPRuntimeNVPTX::emitGenericVarsProlog(CodeGenFunction &CGF,
   if (I == FunctionGlobalizedDecls.end())
     return;
   if (const RecordDecl *GlobalizedVarsRecord = I->getSecond().GlobalRecord) {
-    QualType RecTy = CGM.getContext().getRecordType(GlobalizedVarsRecord);
+    QualType GlobalRecTy = CGM.getContext().getRecordType(GlobalizedVarsRecord);
 
     // Recover pointer to this function's global record. The runtime will
     // handle the specifics of the allocation of the memory.
     // Use actual memory size of the record including the padding
     // for alignment purposes.
     unsigned Alignment =
-        CGM.getContext().getTypeAlignInChars(RecTy).getQuantity();
+        CGM.getContext().getTypeAlignInChars(GlobalRecTy).getQuantity();
     unsigned GlobalRecordSize =
-        CGM.getContext().getTypeSizeInChars(RecTy).getQuantity();
+        CGM.getContext().getTypeSizeInChars(GlobalRecTy).getQuantity();
     GlobalRecordSize = llvm::alignTo(GlobalRecordSize, Alignment);
 
+    llvm::PointerType *GlobalRecPtrTy =
+        CGF.ConvertTypeForMem(GlobalRecTy)->getPointerTo();
     llvm::Value *GlobalRecCastAddr;
     if (WithSPMDCheck ||
         getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_Unknown) {
@@ -1959,7 +1991,8 @@ void CGOpenMPRuntimeNVPTX::emitGenericVarsProlog(CodeGenFunction &CGF,
       // There is no need to emit line number for unconditional branch.
       (void)ApplyDebugLocation::CreateEmpty(CGF);
       CGF.EmitBlock(SPMDBB);
-      Address RecPtr = CGF.CreateMemTemp(RecTy, "_local_stack");
+      Address RecPtr = Address(llvm::ConstantPointerNull::get(GlobalRecPtrTy),
+                               CharUnits::fromQuantity(Alignment));
       CGF.EmitBranch(ExitBB);
       // There is no need to emit line number for unconditional branch.
       (void)ApplyDebugLocation::CreateEmpty(CGF);
@@ -1974,9 +2007,9 @@ void CGOpenMPRuntimeNVPTX::emitGenericVarsProlog(CodeGenFunction &CGF,
                                   OMPRTL_NVPTX__kmpc_data_sharing_push_stack),
                               GlobalRecordSizeArg);
       GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
-          GlobalRecValue, CGF.ConvertTypeForMem(RecTy)->getPointerTo());
+          GlobalRecValue, GlobalRecPtrTy);
       CGF.EmitBlock(ExitBB);
-      auto *Phi = Bld.CreatePHI(GlobalRecCastAddr->getType(),
+      auto *Phi = Bld.CreatePHI(GlobalRecPtrTy,
                                 /*NumReservedValues=*/2, "_select_stack");
       Phi->addIncoming(RecPtr.getPointer(), SPMDBB);
       Phi->addIncoming(GlobalRecCastAddr, NonSPMDBB);
@@ -1994,12 +2027,12 @@ void CGOpenMPRuntimeNVPTX::emitGenericVarsProlog(CodeGenFunction &CGF,
                                   OMPRTL_NVPTX__kmpc_data_sharing_push_stack),
                               GlobalRecordSizeArg);
       GlobalRecCastAddr = Bld.CreatePointerBitCastOrAddrSpaceCast(
-          GlobalRecValue, CGF.ConvertTypeForMem(RecTy)->getPointerTo());
+          GlobalRecValue, GlobalRecPtrTy);
       I->getSecond().GlobalRecordAddr = GlobalRecValue;
       I->getSecond().IsInSPMDModeFlag = nullptr;
     }
     LValue Base =
-        CGF.MakeNaturalAlignPointeeAddrLValue(GlobalRecCastAddr, RecTy);
+        CGF.MakeNaturalAlignPointeeAddrLValue(GlobalRecCastAddr, GlobalRecTy);
 
     // Emit the "global alloca" which is a GEP from the global declaration
     // record using the pointer returned by the runtime.
@@ -2012,9 +2045,34 @@ void CGOpenMPRuntimeNVPTX::emitGenericVarsProlog(CodeGenFunction &CGF,
             CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType());
         ParValue = CGF.EmitLoadOfScalar(ParLVal, Loc);
       }
-      const FieldDecl *FD = Rec.second.first;
-      LValue VarAddr = CGF.EmitLValueForField(Base, FD);
-      Rec.second.second = VarAddr.getAddress();
+      LValue VarAddr = CGF.EmitLValueForField(Base, Rec.second.FD);
+      // Emit VarAddr basing on lane-id if required.
+      QualType VarTy;
+      if (Rec.second.IsOnePerTeam) {
+        Rec.second.PrivateAddr = VarAddr.getAddress();
+        VarTy = Rec.second.FD->getType();
+      } else {
+        llvm::Value *Ptr = CGF.Builder.CreateInBoundsGEP(
+            VarAddr.getAddress().getPointer(),
+            {Bld.getInt32(0), getNVPTXLaneID(CGF)});
+        Rec.second.PrivateAddr =
+            Address(Ptr, CGM.getContext().getDeclAlign(Rec.first));
+        VarTy =
+            Rec.second.FD->getType()->castAsArrayTypeUnsafe()->getElementType();
+        VarAddr = CGF.MakeAddrLValue(Rec.second.PrivateAddr, VarTy,
+                                     AlignmentSource::Decl);
+      }
+      if (WithSPMDCheck ||
+                getExecutionMode() == CGOpenMPRuntimeNVPTX::EM_Unknown) {
+        assert(I->getSecond().IsInSPMDModeFlag &&
+               "Expected unknown execution mode or required SPMD check.");
+        Address GlobalPtr = Rec.second.PrivateAddr;
+        Address LocalAddr = CGF.CreateMemTemp(VarTy, Rec.second.FD->getName());
+        Rec.second.PrivateAddr = Address(
+            Bld.CreateSelect(I->getSecond().IsInSPMDModeFlag,
+                             LocalAddr.getPointer(), GlobalPtr.getPointer()),
+            LocalAddr.getAlignment());
+      }
       if (EscapedParam) {
         const auto *VD = cast<VarDecl>(Rec.first);
         CGF.EmitStoreOfScalar(ParValue, VarAddr);
@@ -4047,7 +4105,7 @@ 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, std::make_pair(FD, Address::invalid())));
+    Data.insert(std::make_pair(VD, MappedVarData(FD)));
   }
   if (!NeedToDelayGlobalization) {
     emitGenericVarsProlog(CGF, D->getBeginLoc(), /*WithSPMDCheck=*/true);
@@ -4074,7 +4132,7 @@ Address CGOpenMPRuntimeNVPTX::getAddressOfLocalVariable(CodeGenFunction &CGF,
     return Address::invalid();
   auto VDI = I->getSecond().LocalVarData.find(VD);
   if (VDI != I->getSecond().LocalVarData.end())
-    return VDI->second.second;
+    return VDI->second.PrivateAddr;
   if (VD->hasAttrs()) {
     for (specific_attr_iterator<OMPReferencedVarAttr> IT(VD->attr_begin()),
          E(VD->attr_end());
@@ -4083,7 +4141,7 @@ Address CGOpenMPRuntimeNVPTX::getAddressOfLocalVariable(CodeGenFunction &CGF,
           cast<VarDecl>(cast<DeclRefExpr>(IT->getRef())->getDecl())
               ->getCanonicalDecl());
       if (VDI != I->getSecond().LocalVarData.end())
-        return VDI->second.second;
+        return VDI->second.PrivateAddr;
     }
   }
   return Address::invalid();

diff --git a/lib/CodeGen/CGOpenMPRuntimeNVPTX.h b/lib/CodeGen/CGOpenMPRuntimeNVPTX.h
index fc8cd2467bf255908fc301bb145c19f7f5958800..cdc856bbcd6fdabf2ae3546f20888999f3b55a99 100644 (file)
--- a/lib/CodeGen/CGOpenMPRuntimeNVPTX.h
+++ b/lib/CodeGen/CGOpenMPRuntimeNVPTX.h
@@ -373,9 +373,21 @@ private:
   llvm::Function *createParallelDataSharingWrapper(
       llvm::Function *OutlinedParallelFn, const OMPExecutableDirective &D);
 
+  /// The data for the single globalized variable.
+  struct MappedVarData {
+    /// Corresponding field in the global record.
+    const FieldDecl * FD = nullptr;
+    /// Corresponding address.
+    Address PrivateAddr = Address::invalid();
+    /// true, if only one element is required (for latprivates in SPMD mode),
+    /// false, if need to create based on the warp-size.
+    bool IsOnePerTeam = false;
+    MappedVarData() = delete;
+    MappedVarData(const FieldDecl *FD, bool IsOnePerTeam = false)
+        : FD(FD), IsOnePerTeam(IsOnePerTeam) {}
+  };
   /// The map of local variables to their addresses in the global memory.
-  using DeclToAddrMapTy = llvm::MapVector<const Decl *,
-      std::pair<const FieldDecl *, Address>>;
+  using DeclToAddrMapTy = llvm::MapVector<const Decl *, MappedVarData>;
   /// Set of the parameters passed by value escaping OpenMP context.
   using EscapedParamsTy = llvm::SmallPtrSet<const Decl *, 4>;
   struct FunctionData {

diff --git a/test/OpenMP/declare_target_codegen_globalization.cpp b/test/OpenMP/declare_target_codegen_globalization.cpp
index 6b08993c1946819bc31042e696c271b80951bb3e..8d2a4da77379d328ccecce14892c185905df7da3 100644 (file)
--- a/test/OpenMP/declare_target_codegen_globalization.cpp
+++ b/test/OpenMP/declare_target_codegen_globalization.cpp
@@ -35,16 +35,21 @@ int maini1() {
 // CHECK-NOT: @__kmpc_data_sharing_push_stack
 
 // CHECK: define {{.*}}[[BAR]]()
-// CHECK: [[STACK:%.+]] = alloca [[GLOBAL_ST:%.+]],
+// CHECK: alloca i32,
+// CHECK: [[A_LOCAL_ADDR:%.+]] = alloca i32,
 // CHECK: [[RES:%.+]] = call i8 @__kmpc_is_spmd_exec_mode()
 // CHECK: [[IS_SPMD:%.+]] = icmp ne i8 [[RES]], 0
 // CHECK: br i1 [[IS_SPMD]], label
 // CHECK: br label
-// CHECK: [[RES:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i64 4, i16 0)
-// CHECK: [[GLOBALS:%.+]] = bitcast i8* [[RES]] to [[GLOBAL_ST]]*
+// CHECK: [[RES:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i64 128, i16 0)
+// CHECK: [[GLOBALS:%.+]] = bitcast i8* [[RES]] to [[GLOBAL_ST:%.+]]*
 // CHECK: br label
-// CHECK: [[ITEMS:%.+]] = phi [[GLOBAL_ST]]* [ [[STACK]], {{.+}} ], [ [[GLOBALS]], {{.+}} ]
+// CHECK: [[ITEMS:%.+]] = phi [[GLOBAL_ST]]* [ null, {{.+}} ], [ [[GLOBALS]], {{.+}} ]
 // CHECK: [[A_ADDR:%.+]] = getelementptr inbounds [[GLOBAL_ST]], [[GLOBAL_ST]]* [[ITEMS]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CHECK: [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+// CHECK: [[LID:%.+]] = and i32 [[TID]], 31
+// CHECK: [[A_GLOBAL_ADDR:%.+]] = getelementptr inbounds [32 x i32], [32 x i32]* [[A_ADDR]], i32 0, i32 [[LID]]
+// CHECK: [[A_ADDR:%.+]] = select i1 [[IS_SPMD]], i32* [[A_LOCAL_ADDR]], i32* [[A_GLOBAL_ADDR]]
 // CHECK: call {{.*}}[[FOO]](i32* dereferenceable{{.*}} [[A_ADDR]])
 // CHECK: br i1 [[IS_SPMD]], label
 // CHECK: [[BC:%.+]] = bitcast [[GLOBAL_ST]]* [[ITEMS]] to i8*

diff --git a/test/OpenMP/nvptx_data_sharing.cpp b/test/OpenMP/nvptx_data_sharing.cpp
index 0acb1199152791c90924d826a71deefdc4f162bf..83206769243864401d7b1f6cf486e4d3cc307aa5 100644 (file)
--- a/test/OpenMP/nvptx_data_sharing.cpp
+++ b/test/OpenMP/nvptx_data_sharing.cpp
@@ -39,10 +39,16 @@ void test_ds(){
 // CK1: [[SHAREDARGS2:%.+]] = alloca i8**
 // CK1: call void @__kmpc_kernel_init
 // CK1: call void @__kmpc_data_sharing_init_stack
-// CK1: [[GLOBALSTACK:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i64 8, i16 0)
+// CK1: [[GLOBALSTACK:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i64 256, i16 0)
 // CK1: [[GLOBALSTACK2:%.+]] = bitcast i8* [[GLOBALSTACK]] to %struct._globalized_locals_ty*
-// CK1: [[A:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[GLOBALSTACK2]], i32 0, i32 0
-// CK1: [[B:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[GLOBALSTACK2]], i32 0, i32 1
+// CK1: [[A_ARR:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[GLOBALSTACK2]], i32 0, i32 0
+// CK1: [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+// CK1: [[LID:%.+]] = and i32 [[TID]], 31
+// CK1: [[A:%.+]] = getelementptr inbounds [32 x i32], [32 x i32]* [[A_ARR]], i32 0, i32 [[LID]]
+// CK1: [[B_ARR:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[GLOBALSTACK2]], i32 0, i32 1
+// CK1: [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+// CK1: [[LID:%.+]] = and i32 [[TID]], 31
+// CK1: [[B:%.+]] = getelementptr inbounds [32 x i32], [32 x i32]* [[B_ARR]], i32 0, i32 [[LID]]
 // CK1: store i32 10, i32* [[A]]
 // CK1: call void @__kmpc_kernel_prepare_parallel({{.*}}, i16 1)
 // CK1: call void @__kmpc_begin_sharing_variables(i8*** [[SHAREDARGS1]], i64 1)

diff --git a/test/OpenMP/nvptx_distribute_parallel_generic_mode_codegen.cpp b/test/OpenMP/nvptx_distribute_parallel_generic_mode_codegen.cpp
index 2b18f6d3f9bdfe873726aa3616d9997ffc2e425d..2a7344f77340cf5f840bc4bc0f4ec9e42a7e4edd 100644 (file)
--- a/test/OpenMP/nvptx_distribute_parallel_generic_mode_codegen.cpp
+++ b/test/OpenMP/nvptx_distribute_parallel_generic_mode_codegen.cpp
@@ -26,10 +26,13 @@ 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}} 84, i16 0)
+// CHECK: [[PTR:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{64|32}} 2688, i16 0)
 // CHECK: [[STACK:%.+]] = bitcast i8* [[PTR]] to %struct._globalized_locals_ty*
 // CHECK: [[ARGC:%.+]] = load i32, i32* %{{.+}}, align
-// CHECK: [[ARGC_ADDR:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[STACK]], i{{32|64}} 0, i{{32|64}} 0
+// 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: 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

diff --git a/test/OpenMP/nvptx_parallel_codegen.cpp b/test/OpenMP/nvptx_parallel_codegen.cpp
index d1a3104407d0987600a97656856934121dca4d01..978804d2495ab9bae7db3ddd40f44f24aa0286e5 100644 (file)
--- a/test/OpenMP/nvptx_parallel_codegen.cpp
+++ b/test/OpenMP/nvptx_parallel_codegen.cpp
@@ -318,11 +318,14 @@ int bar(int n){
 // CHECK-32: [[A_ADDR:%.+]] = alloca i32,
 // CHECK-64: [[A_ADDR:%.+]] = alloca i64,
 // CHECK-64: [[CONV:%.+]] = bitcast i64* [[A_ADDR]] to i32*
-// CHECK: [[STACK:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{64|32}} 4, i16 0)
+// CHECK: [[STACK:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{64|32}} 128, i16 0)
 // CHECK: [[BC:%.+]] = bitcast i8* [[STACK]] to %struct._globalized_locals_ty*
 // CHECK-32: [[A:%.+]] = load i32, i32* [[A_ADDR]],
 // CHECK-64: [[A:%.+]] = load i32, i32* [[CONV]],
-// CHECK: [[GLOBAL_A_ADDR:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[BC]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CHECK: [[GLOBAL_A_ADDR_ARR:%.+]] = getelementptr inbounds %struct._globalized_locals_ty, %struct._globalized_locals_ty* [[BC]], i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CHECK: [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+// CHECK: [[LID:%.+]] = and i32 [[TID]], 31
+// CHECK: [[GLOBAL_A_ADDR:%.+]] = getelementptr inbounds [32 x i32], [32 x i32]* [[GLOBAL_A_ADDR_ARR]], i32 0, i32 [[LID]]
 // CHECK: store i32 [[A]], i32* [[GLOBAL_A_ADDR]],
 
 // CHECK-LABEL: define internal void @{{.+}}(i32* noalias %{{.+}}, i32* noalias %{{.+}}, i32* dereferenceable{{.*}})

diff --git a/test/OpenMP/nvptx_target_codegen.cpp b/test/OpenMP/nvptx_target_codegen.cpp
index 01f346121a4df16760a7fe5f515a0f8f5fb57c76..69c338995b7e0bca30707ee291f17a9994f5f292 100644 (file)
--- a/test/OpenMP/nvptx_target_codegen.cpp
+++ b/test/OpenMP/nvptx_target_codegen.cpp
@@ -554,7 +554,8 @@ int baz(int f, double &a) {
   // CHECK: ret void
 
   // CHECK: define i32 [[BAZ]](i32 [[F:%.*]], double* dereferenceable{{.*}})
-  // CHECK: [[STACK:%.+]] = alloca [[GLOBAL_ST:%.+]],
+  // CHECK: alloca i32,
+  // CHECK: [[LOCAL_F_PTR:%.+]] = alloca i32,
   // CHECK: [[ZERO_ADDR:%.+]] = alloca i32,
   // CHECK: [[GTID:%.+]] = call i32 @__kmpc_global_thread_num(%struct.ident_t*
   // CHECK: store i32 0, i32* [[ZERO_ADDR]]
@@ -562,11 +563,15 @@ int baz(int f, double &a) {
   // CHECK: [[IS_SPMD:%.+]] = icmp ne i8 [[RES]], 0
   // CHECK: br i1 [[IS_SPMD]], label
   // CHECK: br label
-  // CHECK: [[PTR:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{64|32}} 4, i16 0)
-  // CHECK: [[REC_ADDR:%.+]] = bitcast i8* [[PTR]] to [[GLOBAL_ST]]*
+  // CHECK: [[PTR:%.+]] = call i8* @__kmpc_data_sharing_push_stack(i{{64|32}} 128, i16 0)
+  // CHECK: [[REC_ADDR:%.+]] = bitcast i8* [[PTR]] to [[GLOBAL_ST:%.+]]*
   // CHECK: br label
-  // CHECK: [[ITEMS:%.+]] = phi [[GLOBAL_ST]]* [ [[STACK]], {{.+}} ], [ [[REC_ADDR]], {{.+}} ]
-  // CHECK: [[F_PTR:%.+]] = getelementptr inbounds [[GLOBAL_ST]], [[GLOBAL_ST]]* [[ITEMS]], i32 0, i32 0
+  // CHECK: [[ITEMS:%.+]] = phi [[GLOBAL_ST]]* [ null, {{.+}} ], [ [[REC_ADDR]], {{.+}} ]
+  // CHECK: [[F_PTR_ARR:%.+]] = getelementptr inbounds [[GLOBAL_ST]], [[GLOBAL_ST]]* [[ITEMS]], i32 0, i32 0
+  // CHECK: [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+  // CHECK: [[LID:%.+]] = and i32 [[TID]], 31
+  // CHECK: [[GLOBAL_F_PTR:%.+]] = getelementptr inbounds [32 x i32], [32 x i32]* [[F_PTR_ARR]], i32 0, i32 [[LID]]
+  // CHECK: [[F_PTR:%.+]] = select i1 [[IS_SPMD]], i32* [[LOCAL_F_PTR]], i32* [[GLOBAL_F_PTR]]
   // CHECK: store i32 %{{.+}}, i32* [[F_PTR]],
 
   // CHECK: [[RES:%.+]] = call i8 @__kmpc_is_spmd_exec_mode()

diff --git a/test/OpenMP/nvptx_teams_codegen.cpp b/test/OpenMP/nvptx_teams_codegen.cpp
index 4e3f2674fc5f55169ce052b2673ba97deafab788..91b372c65b6a9823ee4625692e456bd3507b1b90 100644 (file)
--- a/test/OpenMP/nvptx_teams_codegen.cpp
+++ b/test/OpenMP/nvptx_teams_codegen.cpp
@@ -36,10 +36,13 @@ int main (int argc, char **argv) {
 // CK1:  store {{.+}} 0, {{.+}},
 // CK1:  store i{{[0-9]+}} [[ARGC]], i{{[0-9]+}}* [[ARGCADDR]],
 // CK1-64:  [[CONV:%.+]] = bitcast i{{[0-9]+}}* [[ARGCADDR]] to i{{[0-9]+}}*
-// CK1:  call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} 4, i16 0)
+// CK1:  call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} 128, i16 0)
 // CK1-64:  [[ARG:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[CONV]]
 // CK1-32:  [[ARG:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[ARGCADDR]]
-// CK1:  [[ARGCADDR:%.+]] = getelementptr inbounds %struct.{{.*}}, %struct.{{.*}}* %{{.*}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CK1:  [[ARGCADDR_ARR:%.+]] = getelementptr inbounds %struct.{{.*}}, %struct.{{.*}}* %{{.*}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CK1:  [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+// CK1:  [[LID:%.+]] = and i32 [[TID]], 31
+// CK1:  [[ARGCADDR:%.+]] = getelementptr inbounds [32 x i32], [32 x i32]* [[ARGCADDR_ARR]], i32 0, i32 [[LID]]
 // CK1:  store i{{[0-9]+}} [[ARG]], i{{[0-9]+}}* [[ARGCADDR]],
 // CK1:  store i{{[0-9]+}}* [[ARGCADDR]], i{{[0-9]+}}** [[ARGCADDR_PTR]],
 // CK1:  [[ARGCADDR_PTR_REF:%.+]] = load i{{[0-9]+}}*, i{{[0-9]+}}** [[ARGCADDR_PTR]],
@@ -53,9 +56,12 @@ int main (int argc, char **argv) {
 // CK1: [[ARGCADDR_PTR:%.+]] = alloca i{{.+}}***,
 // CK1: [[ARGCADDR:%.+]] = alloca i{{.+}}**,
 // CK1: store i{{.+}}** [[ARGC]], i{{.+}}*** [[ARGCADDR]]
-// CK1: call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} {{4|8}}, i16 0)
+// CK1: call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} {{128|256}}, i16 0)
 // CK1: [[ARG:%.+]] = load i{{[0-9]+}}**, i{{[0-9]+}}*** [[ARGCADDR]]
-// CK1: [[ARGCADDR:%.+]] = getelementptr inbounds %struct.{{.*}}, %struct.{{.*}}* %{{.*}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CK1: [[ARGCADDR_ARR:%.+]] = getelementptr inbounds %struct.{{.*}}, %struct.{{.*}}* %{{.*}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CK1: [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+// CK1: [[LID:%.+]] = and i32 [[TID]], 31
+// CK1: [[ARGCADDR:%.+]] = getelementptr inbounds [32 x i8**], [32 x i8**]* [[ARGCADDR_ARR]], i32 0, i32 [[LID]]
 // CK1: store i{{[0-9]+}}** [[ARG]], i{{[0-9]+}}*** [[ARGCADDR]],
 // CK1: store i8*** [[ARGCADDR]], i8**** [[ARGCADDR_PTR]],
 // CK1: [[ARGCADDR_PTR_REF:%.+]] = load i{{.+}}**, i{{.+}}*** [[ARGCADDR_PTR]],
@@ -111,10 +117,13 @@ int main (int argc, char **argv) {
 // CK2-64: [[ACONV:%.+]] = bitcast i64* [[AADDR]] to i32*
 // CK2-64: [[BCONV:%.+]] = bitcast i64* [[BADDR]] to i32*
 // CK2-64: [[CONV:%.+]] = bitcast i64* [[ARGCADDR]] to i32*
-// CK2:  call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} 4, i16 0)
+// CK2:  call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} 128, i16 0)
 // CK2-64:  [[ARG:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[CONV]]
 // CK2-32:  [[ARG:%.+]] = load i{{[0-9]+}}, i{{[0-9]+}}* [[ARGCADDR]]
-// CK2:  [[ARGCADDR:%.+]] = getelementptr inbounds %struct.{{.*}}, %struct.{{.*}}* %{{.*}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CK2:  [[ARGCADDR_ARR:%.+]] = getelementptr inbounds %struct.{{.*}}, %struct.{{.*}}* %{{.*}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CK2:  [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+// CK2:  [[LID:%.+]] = and i32 [[TID]], 31
+// CK2:  [[ARGCADDR:%.+]] = getelementptr inbounds [32 x i32], [32 x i32]* [[ARGCADDR_ARR]], i32 0, i32 [[LID]]
 // CK2:  store i{{[0-9]+}} [[ARG]], i{{[0-9]+}}* [[ARGCADDR]],
 // CK2:  {{%.+}} = call i32 @__kmpc_global_thread_num(
 // CK2:  store i{{[0-9]+}}* [[ARGCADDR]], i{{[0-9]+}}** [[ARGCADDR_PTR]],
@@ -132,9 +141,12 @@ int main (int argc, char **argv) {
 // CK2: store i{{[0-9]+}} [[A_IN]], i{{[0-9]+}}* [[AADDR]],
 // CK2: store i{{[0-9]+}} [[B_IN]], i{{[0-9]+}}* [[BADDR]],
 // CK2: store i{{[0-9]+}}** [[ARGC]], i{{[0-9]+}}*** [[ARGCADDR]],
-// CK2: call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} {{4|8}}, i16 0)
+// CK2: call i8* @__kmpc_data_sharing_push_stack(i{{[0-9]+}} {{128|256}}, i16 0)
 // CK2: [[ARG:%.+]] = load i{{[0-9]+}}**, i{{[0-9]+}}*** [[ARGCADDR]]
-// CK2: [[ARGCADDR:%.+]] = getelementptr inbounds %struct.{{.*}}, %struct.{{.*}}* %{{.*}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CK2: [[ARGCADDR_ARR:%.+]] = getelementptr inbounds %struct.{{.*}}, %struct.{{.*}}* %{{.*}}, i{{[0-9]+}} 0, i{{[0-9]+}} 0
+// CK2: [[TID:%.+]] = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
+// CK2: [[LID:%.+]] = and i32 [[TID]], 31
+// CK2: [[ARGCADDR:%.+]] = getelementptr inbounds [32 x i8**], [32 x i8**]* [[ARGCADDR_ARR]], i32 0, i32 [[LID]]
 // CK2: store i{{[0-9]+}}** [[ARG]], i{{[0-9]+}}*** [[ARGCADDR]],
 // CK2: {{%.+}} = call i32 @__kmpc_global_thread_num(
 // CK2: store i{{[0-9]+}}*** [[ARGCADDR]], i{{[0-9]+}}**** [[ARGCADDR_PTR]],

diff --git a/test/OpenMP/nvptx_teams_reduction_codegen.cpp b/test/OpenMP/nvptx_teams_reduction_codegen.cpp
index 6925b17163af8dbeb19d42d40c872cc35949a546..b73b5a8e93e53bebe771569274f66a5cc856d044 100644 (file)
--- a/test/OpenMP/nvptx_teams_reduction_codegen.cpp
+++ b/test/OpenMP/nvptx_teams_reduction_codegen.cpp
@@ -597,9 +597,9 @@ int bar(int n){
   // CHECK: [[OF:%.+]] = mul nuw i[[SZ]] [[NUM_TEAMS]], 1
   // CHECK: [[POS1:%.+]] = add nuw i[[SZ]] [[SCRATCHPAD]], [[OF]]
   // CHECK: [[POS2:%.+]] = sub nuw i[[SZ]] [[POS1]], 1
-  // CHECK: [[POS3:%.+]] = udiv i[[SZ]] [[POS2]], 256
+  // CHECK: [[POS3:%.+]] = udiv i[[SZ]] [[POS2]], 128
   // CHECK: [[POS4:%.+]] = add nuw i[[SZ]] [[POS3]], 1
-  // CHECK: [[SCRATCHPAD_NEXT:%.+]] = mul nuw i[[SZ]] [[POS4]], 256
+  // CHECK: [[SCRATCHPAD_NEXT:%.+]] = mul nuw i[[SZ]] [[POS4]], 128
   //
   // CHECK: [[ELT_REF:%.+]] = getelementptr inbounds [[RLT]], [[RLT]]* [[RED_LIST]], i[[SZ]] 0, i[[SZ]] 1
   // CHECK: [[ELT_VOID:%.+]] = load i8*, i8** [[ELT_REF]],
@@ -643,9 +643,9 @@ int bar(int n){
   // CHECK: [[OF:%.+]] = mul nuw i[[SZ]] [[NUM_TEAMS]], 1
   // CHECK: [[POS1:%.+]] = add nuw i[[SZ]] [[SCRATCHPAD]], [[OF]]
   // CHECK: [[POS2:%.+]] = sub nuw i[[SZ]] [[POS1]], 1
-  // CHECK: [[POS3:%.+]] = udiv i[[SZ]] [[POS2]], 256
+  // CHECK: [[POS3:%.+]] = udiv i[[SZ]] [[POS2]], 128
   // CHECK: [[POS4:%.+]] = add nuw i[[SZ]] [[POS3]], 1
-  // CHECK: [[SCRATCHPAD_NEXT:%.+]] = mul nuw i[[SZ]] [[POS4]], 256
+  // CHECK: [[SCRATCHPAD_NEXT:%.+]] = mul nuw i[[SZ]] [[POS4]], 128
   //
   // CHECK: [[P:%.+]] = mul nuw i[[SZ]] 4, [[TEAM]]
   // CHECK: [[SCRATCHPAD_ELT_PTR64:%.+]] = add nuw i[[SZ]] [[SCRATCHPAD_NEXT]], [[P]]
@@ -1024,9 +1024,9 @@ int bar(int n){
   // CHECK: [[OF:%.+]] = mul nuw i[[SZ]] [[NUM_TEAMS]], 4
   // CHECK: [[POS1:%.+]] = add nuw i[[SZ]] [[SCRATCHPAD]], [[OF]]
   // CHECK: [[POS2:%.+]] = sub nuw i[[SZ]] [[POS1]], 1
-  // CHECK: [[POS3:%.+]] = udiv i[[SZ]] [[POS2]], 256
+  // CHECK: [[POS3:%.+]] = udiv i[[SZ]] [[POS2]], 128
   // CHECK: [[POS4:%.+]] = add nuw i[[SZ]] [[POS3]], 1
-  // CHECK: [[SCRATCHPAD_NEXT:%.+]] = mul nuw i[[SZ]] [[POS4]], 256
+  // CHECK: [[SCRATCHPAD_NEXT:%.+]] = mul nuw i[[SZ]] [[POS4]], 128
   //
   // CHECK: [[ELT_REF:%.+]] = getelementptr inbounds [[RLT]], [[RLT]]* [[RED_LIST]], i[[SZ]] 0, i[[SZ]] 1
   // CHECK: [[ELT_VOID:%.+]] = load i8*, i8** [[ELT_REF]],
@@ -1072,9 +1072,9 @@ int bar(int n){
   // CHECK: [[OF:%.+]] = mul nuw i[[SZ]] [[NUM_TEAMS]], 4
   // CHECK: [[POS1:%.+]] = add nuw i[[SZ]] [[SCRATCHPAD]], [[OF]]
   // CHECK: [[POS2:%.+]] = sub nuw i[[SZ]] [[POS1]], 1
-  // CHECK: [[POS3:%.+]] = udiv i[[SZ]] [[POS2]], 256
+  // CHECK: [[POS3:%.+]] = udiv i[[SZ]] [[POS2]], 128
   // CHECK: [[POS4:%.+]] = add nuw i[[SZ]] [[POS3]], 1
-  // CHECK: [[SCRATCHPAD_NEXT:%.+]] = mul nuw i[[SZ]] [[POS4]], 256
+  // CHECK: [[SCRATCHPAD_NEXT:%.+]] = mul nuw i[[SZ]] [[POS4]], 128
   //
   // CHECK: [[P:%.+]] = mul nuw i[[SZ]] 2, [[TEAM]]
   // CHECK: [[SCRATCHPAD_ELT_PTR64:%.+]] = add nuw i[[SZ]] [[SCRATCHPAD_NEXT]], [[P]]