[OPENMP50]Initial codegen for 'affinity' clauses.

[ABataev]

Summary

Added initial codegen for 'affinity' clauses on task directives.
Emits next code:

kmp_task_affinity_info_t affs[<num_elems>];

void *td = __kmpc_task_alloc(..);

affs[<i>].base = &data_i;
affs[<i>].size = sizeof(data_i);
__kmpc_omp_reg_task_with_affinity(&loc, <gtid>, td, <num_elems>, affs);

The result returned by the call of __kmpc_omp_reg_task_with_affinity
function is ignored currently sincethe runtime currently ignores args
and returns 0 uncoditionally.

[ABataev]

Rebase

[jdoerfert]

Few comments, the alloca question is my only real concern. Parts of this are trivial NFC and can go in to make the patch smaller. Parts won't be needed after D80222 landed :)

clang/lib/CodeGen/CGOpenMPRuntime.cpp

2444

This will clash with D80222. I guess once that one is in you can just remove these parts and it will still work fine :)

5397

Feel free to commit the code movement parts separately as an NFC commit to make the diff cleaner.

clang/lib/CodeGen/CGOpenMPRuntime.h

470

We really need to move these into OMPKinds.def as well. Not in this patch but soon.

clang/test/OpenMP/task_affinity_codegen.cpp

55

I'm not so sure about dynamic allocas (without stack save/restore). If this happens in a loop we can run into problems fast, right?

133

Does it make sense to use the script to auto-generate the check lines? Makes updates way easier and cleaner. We could have the "C-code" explanation as comment above the pragma still.

[ABataev]

clang/test/OpenMP/task_affinity_codegen.cpp

55

There are stack saves and restores for this alloca, see the calls of @llvm.stacksave() and @llvm.stackrestore()

[jdoerfert]

LGTM. Let's wait until D80222 landed :)

clang/lib/CodeGen/CGOpenMPRuntime.h

470

Just a comment, not necessarily related to this review:
@clementval @DavidTruby @fghanim @jhuber6
We want to have all "internal" struct definitions in OMPKinds.def (as we have now), but also build boilerplate code for each of them once we use TableGen. We want to build a class STRUCT_TY which offers to create a local/global struct value, inspect the members, modify the members, create the IR type for it, etc.

clang/test/OpenMP/task_affinity_codegen.cpp

55

Right, missed them. Thx.

clang/lib/CodeGen/CGOpenMPRuntime.h

private:
   ///    kmp_intptr_t               base_addr;
   ///    size_t                     len;
   ///    struct {
   ///             bool                   in:1;
   ///             bool                   out:1;
   ///    } flags;
   /// } kmp_depend_info_t;
   QualType KmpDependInfoTy;
+  /// Type typedef struct kmp_task_affinity_info {
+  ///    kmp_intptr_t base_addr;
+  ///    size_t len;
+  ///    struct {
+  ///      bool flag1 : 1;
+  ///      bool flag2 : 1;
+  ///      kmp_int32 reserved : 30;
+  ///   } flags;
+  /// } kmp_task_affinity_info_t;
+  QualType KmpTaskAffinityInfoTy;
   /// struct kmp_dim {  // loop bounds info casted to kmp_int64
   ///  kmp_int64 lo; // lower
   ///  kmp_int64 up; // upper
   ///  kmp_int64 st; // stride
   /// };
   QualType KmpDimTy;
   /// Type struct __tgt_offload_entry{
   ///   void      *addr;       // Pointer to the offload entry info.

clang/lib/CodeGen/CGOpenMPRuntime.cpp

enum OpenMPRTLFunction {
   OMPRTL__kmpc_alloc,
   // Call to void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
   OMPRTL__kmpc_free,
   // Call to omp_allocator_handle_t __kmpc_init_allocator(int gtid,
   // omp_memspace_handle_t, int ntraits, omp_alloctrait_t traits[]);
   OMPRTL__kmpc_init_allocator,
   // Call to void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
   OMPRTL__kmpc_destroy_allocator,
+  // Call to kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref,
+  // kmp_int32 gtid, kmp_task_t *new_task, kmp_int32
+  // naffins,kmp_task_affinity_info_t *affin_list);
+  OMPRTL__kmpc_omp_reg_task_with_affinity,
 
   //
   // Offloading related calls
   //
   // Call to void __kmpc_push_target_tripcount(int64_t device_id, kmp_uint64
   // size);
   OMPRTL__kmpc_push_target_tripcount,
   // Call to int32_t __tgt_target(int64_t device_id, void *host_ptr, int32_t
case OMPRTL__kmpc_push_target_tripcount: {
     // Build void __kmpc_push_target_tripcount(int64_t device_id, kmp_uint64
     // size);
     llvm::Type *TypeParams[] = {CGM.Int64Ty, CGM.Int64Ty};
     llvm::FunctionType *FnTy =
         llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false);
     RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_target_tripcount");
     break;
   }
+  case OMPRTL__kmpc_omp_reg_task_with_affinity: {
+    // Build kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref,
+    // kmp_int32 gtid, kmp_task_t *new_task, kmp_int32
+    // naffins,kmp_task_affinity_info_t *affin_list);
+    llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty,
+                                CGM.VoidPtrTy, CGM.Int32Ty, CGM.VoidPtrTy};
+    llvm::FunctionType *FnTy =
+        llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false);
+    RTLFn =
+        CGM.CreateRuntimeFunction(FnTy, "__kmpc_omp_reg_task_with_affinity");
+    break;
+  }
   case OMPRTL__tgt_target: {
     // Build int32_t __tgt_target(int64_t device_id, void *host_ptr, int32_t
     // arg_num, void** args_base, void **args, int64_t *arg_sizes, int64_t
     // *arg_types);
     llvm::Type *TypeParams[] = {CGM.Int64Ty,
                                 CGM.VoidPtrTy,
                                 CGM.Int32Ty,
                                 CGM.VoidPtrPtrTy,
checkDestructorsRequired(const RecordDecl *KmpTaskTWithPrivatesQTyRD) {
   for (const FieldDecl *FD : PrivateRD->fields()) {
     NeedsCleanup = NeedsCleanup || FD->getType().isDestructedType();
     if (NeedsCleanup)
       break;
   }
   return NeedsCleanup;
 }
 
+namespace {
+/// Loop generator for OpenMP iterator expression.
+class OMPIteratorGeneratorScope final
+    : public CodeGenFunction::OMPPrivateScope {
+  CodeGenFunction &CGF;
+  const OMPIteratorExpr *E = nullptr;
+  SmallVector<CodeGenFunction::JumpDest, 4> ContDests;
+  SmallVector<CodeGenFunction::JumpDest, 4> ExitDests;
+  OMPIteratorGeneratorScope() = delete;
+  OMPIteratorGeneratorScope(OMPIteratorGeneratorScope &) = delete;
+
+public:
+  OMPIteratorGeneratorScope(CodeGenFunction &CGF, const OMPIteratorExpr *E)
+      : CodeGenFunction::OMPPrivateScope(CGF), CGF(CGF), E(E) {
+    if (!E)
+      return;
+    SmallVector<llvm::Value *, 4> Uppers;
+    for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) {
+      Uppers.push_back(CGF.EmitScalarExpr(E->getHelper(I).Upper));
+      const auto *VD = cast<VarDecl>(E->getIteratorDecl(I));
+      addPrivate(VD, [&CGF, VD]() {
+        return CGF.CreateMemTemp(VD->getType(), VD->getName());
+      });
+      const OMPIteratorHelperData &HelperData = E->getHelper(I);
+      addPrivate(HelperData.CounterVD, [&CGF, &HelperData]() {
+        return CGF.CreateMemTemp(HelperData.CounterVD->getType(),
+                                 "counter.addr");
+      });
+    }
+    Privatize();
+
+    for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) {
+      const OMPIteratorHelperData &HelperData = E->getHelper(I);
+      LValue CLVal =
+          CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(HelperData.CounterVD),
+                             HelperData.CounterVD->getType());
+      // Counter = 0;
+      CGF.EmitStoreOfScalar(
+          llvm::ConstantInt::get(CLVal.getAddress(CGF).getElementType(), 0),
+          CLVal);
+      CodeGenFunction::JumpDest &ContDest =
+          ContDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.cont"));
+      CodeGenFunction::JumpDest &ExitDest =
+          ExitDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.exit"));
+      // N = <number-of_iterations>;
+      llvm::Value *N = Uppers[I];
+      // cont:
+      // if (Counter < N) goto body; else goto exit;
+      CGF.EmitBlock(ContDest.getBlock());
+      auto *CVal =
+          CGF.EmitLoadOfScalar(CLVal, HelperData.CounterVD->getLocation());
+      llvm::Value *Cmp =
+          HelperData.CounterVD->getType()->isSignedIntegerOrEnumerationType()
+              ? CGF.Builder.CreateICmpSLT(CVal, N)
+              : CGF.Builder.CreateICmpULT(CVal, N);
+      llvm::BasicBlock *BodyBB = CGF.createBasicBlock("iter.body");
+      CGF.Builder.CreateCondBr(Cmp, BodyBB, ExitDest.getBlock());
+      // body:
+      CGF.EmitBlock(BodyBB);
+      // Iteri = Begini + Counter * Stepi;
+      CGF.EmitIgnoredExpr(HelperData.Update);
+    }
+  }
+  ~OMPIteratorGeneratorScope() {
+    if (!E)
+      return;
+    for (unsigned I = E->numOfIterators(); I > 0; --I) {
+      // Counter = Counter + 1;
+      const OMPIteratorHelperData &HelperData = E->getHelper(I - 1);
+      CGF.EmitIgnoredExpr(HelperData.CounterUpdate);
+      // goto cont;
+      CGF.EmitBranchThroughCleanup(ContDests[I - 1]);
+      // exit:
+      CGF.EmitBlock(ExitDests[I - 1].getBlock(), /*IsFinished=*/I == 1);
+    }
+  }
+};
+} // namespace
+
+static std::pair<llvm::Value *, llvm::Value *>
+getPointerAndSize(CodeGenFunction &CGF, const Expr *E) {
+  const auto *OASE = dyn_cast<OMPArrayShapingExpr>(E);
+  llvm::Value *Addr;
+  if (OASE) {
+    const Expr *Base = OASE->getBase();
+    Addr = CGF.EmitScalarExpr(Base);
+  } else {
+    Addr = CGF.EmitLValue(E).getPointer(CGF);
+  }
+  llvm::Value *SizeVal;
+  QualType Ty = E->getType();
+  if (OASE) {
+    SizeVal = CGF.getTypeSize(OASE->getBase()->getType()->getPointeeType());
+    for (const Expr *SE : OASE->getDimensions()) {
+      llvm::Value *Sz = CGF.EmitScalarExpr(SE);
+      Sz = CGF.EmitScalarConversion(
+          Sz, SE->getType(), CGF.getContext().getSizeType(), SE->getExprLoc());
+      SizeVal = CGF.Builder.CreateNUWMul(SizeVal, Sz);
+    }
+  } else if (const auto *ASE =
+                 dyn_cast<OMPArraySectionExpr>(E->IgnoreParenImpCasts())) {
+    LValue UpAddrLVal =
+        CGF.EmitOMPArraySectionExpr(ASE, /*IsLowerBound=*/false);
+    llvm::Value *UpAddr =
+        CGF.Builder.CreateConstGEP1_32(UpAddrLVal.getPointer(CGF), /*Idx0=*/1);
+    llvm::Value *LowIntPtr = CGF.Builder.CreatePtrToInt(Addr, CGF.SizeTy);
+    llvm::Value *UpIntPtr = CGF.Builder.CreatePtrToInt(UpAddr, CGF.SizeTy);
+    SizeVal = CGF.Builder.CreateNUWSub(UpIntPtr, LowIntPtr);
+  } else {
+    SizeVal = CGF.getTypeSize(Ty);
+  }
+  return std::make_pair(Addr, SizeVal);
+}
+
+/// Builds kmp_depend_info, if it is not built yet, and builds flags type.
+static void getKmpAffinityType(ASTContext &C, QualType &KmpTaskAffinityInfoTy) {
+  QualType FlagsTy = C.getIntTypeForBitwidth(32, /*Signed=*/false);
+  if (KmpTaskAffinityInfoTy.isNull()) {
+    RecordDecl *KmpAffinityInfoRD =
+        C.buildImplicitRecord("kmp_task_affinity_info_t");
+    KmpAffinityInfoRD->startDefinition();
+    addFieldToRecordDecl(C, KmpAffinityInfoRD, C.getIntPtrType());
+    addFieldToRecordDecl(C, KmpAffinityInfoRD, C.getSizeType());
+    addFieldToRecordDecl(C, KmpAffinityInfoRD, FlagsTy);
+    KmpAffinityInfoRD->completeDefinition();
+    KmpTaskAffinityInfoTy = C.getRecordType(KmpAffinityInfoRD);
+  }
+}
+
 CGOpenMPRuntime::TaskResultTy
 CGOpenMPRuntime::emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc,
                               const OMPExecutableDirective &D,
                               llvm::Function *TaskFunction, QualType SharedsTy,
                               Address Shareds, const OMPTaskDataTy &Data) {
   ASTContext &C = CGM.getContext();
   llvm::SmallVector<PrivateDataTy, 4> Privates;
   // Aggregate privates and sort them by the alignment.
if (const auto *DC = D.getSingleClause<OMPDetachClause>()) {
     Tid = CGF.Builder.CreateIntCast(Tid, CGF.IntTy, /*isSigned=*/false);
     llvm::Value *EvtVal = CGF.EmitRuntimeCall(
         createRuntimeFunction(OMPRTL__kmpc_task_allow_completion_event),
         {Loc, Tid, NewTask});
     EvtVal = CGF.EmitScalarConversion(EvtVal, C.VoidPtrTy, Evt->getType(),
                                       Evt->getExprLoc());
     CGF.EmitStoreOfScalar(EvtVal, EvtLVal);
   }
+  // Process affinity clauses.
+  if (D.hasClausesOfKind<OMPAffinityClause>()) {
+    // Process list of affinity data.
+    ASTContext &C = CGM.getContext();
+    Address AffinitiesArray = Address::invalid();
+    // Calculate number of elements to form the array of affinity data.
+    llvm::Value *NumOfElements = nullptr;
+    unsigned NumAffinities = 0;
+    for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) {
+      if (const Expr *Modifier = C->getModifier()) {
+        const auto *IE = cast<OMPIteratorExpr>(Modifier->IgnoreParenImpCasts());
+        for (unsigned I = 0, E = IE->numOfIterators(); I < E; ++I) {
+          llvm::Value *Sz = CGF.EmitScalarExpr(IE->getHelper(I).Upper);
+          Sz = CGF.Builder.CreateIntCast(Sz, CGF.SizeTy, /*isSigned=*/false);
+          NumOfElements =
+              NumOfElements ? CGF.Builder.CreateNUWMul(NumOfElements, Sz) : Sz;
+        }
+      } else {
+        NumAffinities += C->varlist_size();
+      }
+    }
+    getKmpAffinityType(CGM.getContext(), KmpTaskAffinityInfoTy);
+    /// Fields ids in kmp_task_affinity_info record.
+    enum RTLAffinityInfoFieldsTy { BaseAddr, Len, Flags };
+
+    QualType KmpTaskAffinityInfoArrayTy;
+    if (NumOfElements) {
+      NumOfElements = CGF.Builder.CreateNUWAdd(
+          llvm::ConstantInt::get(CGF.SizeTy, NumAffinities), NumOfElements);
+      OpaqueValueExpr OVE(
+          Loc,
+          C.getIntTypeForBitwidth(C.getTypeSize(C.getSizeType()), /*Signed=*/0),
+          VK_RValue);
+      CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE,
+                                                    RValue::get(NumOfElements));
+      KmpTaskAffinityInfoArrayTy =
+          C.getVariableArrayType(KmpTaskAffinityInfoTy, &OVE, ArrayType::Normal,
+                                 /*IndexTypeQuals=*/0, SourceRange(Loc, Loc));
+      // Properly emit variable-sized array.
+      auto *PD = ImplicitParamDecl::Create(C, KmpTaskAffinityInfoArrayTy,
+                                           ImplicitParamDecl::Other);
+      CGF.EmitVarDecl(*PD);
+      AffinitiesArray = CGF.GetAddrOfLocalVar(PD);
+      NumOfElements = CGF.Builder.CreateIntCast(NumOfElements, CGF.Int32Ty,
+                                                /*isSigned=*/false);
+    } else {
+      KmpTaskAffinityInfoArrayTy = C.getConstantArrayType(
+          KmpTaskAffinityInfoTy,
+          llvm::APInt(C.getTypeSize(C.getSizeType()), NumAffinities), nullptr,
+          ArrayType::Normal, /*IndexTypeQuals=*/0);
+      AffinitiesArray =
+          CGF.CreateMemTemp(KmpTaskAffinityInfoArrayTy, ".affs.arr.addr");
+      AffinitiesArray = CGF.Builder.CreateConstArrayGEP(AffinitiesArray, 0);
+      NumOfElements = llvm::ConstantInt::get(CGM.Int32Ty, NumAffinities,
+                                             /*isSigned=*/false);
+    }
+
+    const auto *KmpAffinityInfoRD = KmpTaskAffinityInfoTy->getAsRecordDecl();
+    // Fill array by elements without iterators.
+    unsigned Pos = 0;
+    bool HasIterator = false;
+    for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) {
+      if (C->getModifier()) {
+        HasIterator = true;
+        continue;
+      }
+      for (const Expr *E : C->varlists()) {
+        llvm::Value *Addr;
+        llvm::Value *Size;
+        std::tie(Addr, Size) = getPointerAndSize(CGF, E);
+        LValue Base =
+            CGF.MakeAddrLValue(CGF.Builder.CreateConstGEP(AffinitiesArray, Pos),
+                               KmpTaskAffinityInfoTy);
+        // affs[i].base_addr = &<Affinities[i].second>;
+        LValue BaseAddrLVal = CGF.EmitLValueForField(
+            Base, *std::next(KmpAffinityInfoRD->field_begin(), BaseAddr));
+        CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy),
+                              BaseAddrLVal);
+        // affs[i].len = sizeof(<Affinities[i].second>);
+        LValue LenLVal = CGF.EmitLValueForField(
+            Base, *std::next(KmpAffinityInfoRD->field_begin(), Len));
+        CGF.EmitStoreOfScalar(Size, LenLVal);
+        ++Pos;
+      }
+    }
+    LValue PosLVal;
+    if (HasIterator) {
+      PosLVal = CGF.MakeAddrLValue(
+          CGF.CreateMemTemp(C.getSizeType(), "affs.counter.addr"),
+          C.getSizeType());
+      CGF.EmitStoreOfScalar(llvm::ConstantInt::get(CGF.SizeTy, Pos), PosLVal);
+    }
+    // Process elements with iterators.
+    for (const auto *C : D.getClausesOfKind<OMPAffinityClause>()) {
+      const Expr *Modifier = C->getModifier();
+      if (!Modifier)
+        continue;
+      OMPIteratorGeneratorScope IteratorScope(
+          CGF, cast_or_null<OMPIteratorExpr>(Modifier->IgnoreParenImpCasts()));
+      for (const Expr *E : C->varlists()) {
+        llvm::Value *Addr;
+        llvm::Value *Size;
+        std::tie(Addr, Size) = getPointerAndSize(CGF, E);
+        llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc());
+        LValue Base = CGF.MakeAddrLValue(
+            Address(CGF.Builder.CreateGEP(AffinitiesArray.getPointer(), Idx),
+                    AffinitiesArray.getAlignment()),
+            KmpTaskAffinityInfoTy);
+        // affs[i].base_addr = &<Affinities[i].second>;
+        LValue BaseAddrLVal = CGF.EmitLValueForField(
+            Base, *std::next(KmpAffinityInfoRD->field_begin(), BaseAddr));
+        CGF.EmitStoreOfScalar(CGF.Builder.CreatePtrToInt(Addr, CGF.IntPtrTy),
+                              BaseAddrLVal);
+        // affs[i].len = sizeof(<Affinities[i].second>);
+        LValue LenLVal = CGF.EmitLValueForField(
+            Base, *std::next(KmpAffinityInfoRD->field_begin(), Len));
+        CGF.EmitStoreOfScalar(Size, LenLVal);
+        Idx = CGF.Builder.CreateNUWAdd(
+            Idx, llvm::ConstantInt::get(Idx->getType(), 1));
+        CGF.EmitStoreOfScalar(Idx, PosLVal);
+      }
+    }
+    // Call to kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref,
+    // kmp_int32 gtid, kmp_task_t *new_task, kmp_int32
+    // naffins, kmp_task_affinity_info_t *affin_list);
+    llvm::Value *LocRef = emitUpdateLocation(CGF, Loc);
+    llvm::Value *GTid = getThreadID(CGF, Loc);
+    llvm::Value *AffinListPtr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
+        AffinitiesArray.getPointer(), CGM.VoidPtrTy);
+    // FIXME: Emit the function and ignore its result for now unless the
+    // runtime function is properly implemented.
+    (void)CGF.EmitRuntimeCall(
+        createRuntimeFunction(OMPRTL__kmpc_omp_reg_task_with_affinity),
+        {LocRef, GTid, NewTask, NumOfElements, AffinListPtr});
+  }
   llvm::Value *NewTaskNewTaskTTy =
       CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(
           NewTask, KmpTaskTWithPrivatesPtrTy);
   LValue Base = CGF.MakeNaturalAlignAddrLValue(NewTaskNewTaskTTy,
                                                KmpTaskTWithPrivatesQTy);
   LValue TDBase =
       CGF.EmitLValueForField(Base, *KmpTaskTWithPrivatesQTyRD->field_begin());
   // Fill the data in the resulting kmp_task_t record.
LValue NumDepsBase = CGF.MakeAddrLValue(
       Base.getBaseInfo(), Base.getTBAAInfo());
   // NumDeps = deps[i].base_addr;
   LValue BaseAddrLVal = CGF.EmitLValueForField(
       NumDepsBase, *std::next(KmpDependInfoRD->field_begin(), BaseAddr));
   llvm::Value *NumDeps = CGF.EmitLoadOfScalar(BaseAddrLVal, Loc);
   return std::make_pair(NumDeps, Base);
 }
 
-namespace {
-/// Loop generator for OpenMP iterator expression.
-class OMPIteratorGeneratorScope final
-    : public CodeGenFunction::OMPPrivateScope {
-  CodeGenFunction &CGF;
-  const OMPIteratorExpr *E = nullptr;
-  SmallVector<CodeGenFunction::JumpDest, 4> ContDests;
-  SmallVector<CodeGenFunction::JumpDest, 4> ExitDests;
-  OMPIteratorGeneratorScope() = delete;
-  OMPIteratorGeneratorScope(OMPIteratorGeneratorScope &) = delete;
-
-public:
-  OMPIteratorGeneratorScope(CodeGenFunction &CGF, const OMPIteratorExpr *E)
-      : CodeGenFunction::OMPPrivateScope(CGF), CGF(CGF), E(E) {
-    if (!E)
-      return;
-    SmallVector<llvm::Value *, 4> Uppers;
-    for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) {
-      Uppers.push_back(CGF.EmitScalarExpr(E->getHelper(I).Upper));
-      const auto *VD = cast<VarDecl>(E->getIteratorDecl(I));
-      addPrivate(VD, [&CGF, VD]() {
-        return CGF.CreateMemTemp(VD->getType(), VD->getName());
-      });
-      const OMPIteratorHelperData &HelperData = E->getHelper(I);
-      addPrivate(HelperData.CounterVD, [&CGF, &HelperData]() {
-        return CGF.CreateMemTemp(HelperData.CounterVD->getType(),
-                                 "counter.addr");
-      });
-    }
-    Privatize();
-
-    for (unsigned I = 0, End = E->numOfIterators(); I < End; ++I) {
-      const OMPIteratorHelperData &HelperData = E->getHelper(I);
-      LValue CLVal =
-          CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(HelperData.CounterVD),
-                             HelperData.CounterVD->getType());
-      // Counter = 0;
-      CGF.EmitStoreOfScalar(
-          llvm::ConstantInt::get(CLVal.getAddress(CGF).getElementType(), 0),
-          CLVal);
-      CodeGenFunction::JumpDest &ContDest =
-          ContDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.cont"));
-      CodeGenFunction::JumpDest &ExitDest =
-          ExitDests.emplace_back(CGF.getJumpDestInCurrentScope("iter.exit"));
-      // N = <number-of_iterations>;
-      llvm::Value *N = Uppers[I];
-      // cont:
-      // if (Counter < N) goto body; else goto exit;
-      CGF.EmitBlock(ContDest.getBlock());
-      auto *CVal =
-          CGF.EmitLoadOfScalar(CLVal, HelperData.CounterVD->getLocation());
-      llvm::Value *Cmp =
-          HelperData.CounterVD->getType()->isSignedIntegerOrEnumerationType()
-              ? CGF.Builder.CreateICmpSLT(CVal, N)
-              : CGF.Builder.CreateICmpULT(CVal, N);
-      llvm::BasicBlock *BodyBB = CGF.createBasicBlock("iter.body");
-      CGF.Builder.CreateCondBr(Cmp, BodyBB, ExitDest.getBlock());
-      // body:
-      CGF.EmitBlock(BodyBB);
-      // Iteri = Begini + Counter * Stepi;
-      CGF.EmitIgnoredExpr(HelperData.Update);
-    }
-  }
-  ~OMPIteratorGeneratorScope() {
-    if (!E)
-      return;
-    for (unsigned I = E->numOfIterators(); I > 0; --I) {
-      // Counter = Counter + 1;
-      const OMPIteratorHelperData &HelperData = E->getHelper(I - 1);
-      CGF.EmitIgnoredExpr(HelperData.CounterUpdate);
-      // goto cont;
-      CGF.EmitBranchThroughCleanup(ContDests[I - 1]);
-      // exit:
-      CGF.EmitBlock(ExitDests[I - 1].getBlock(), /*IsFinished=*/I == 1);
-    }
-  }
-};
-} // namespace
-
 static void emitDependData(CodeGenFunction &CGF, QualType &KmpDependInfoTy,
                            llvm::PointerUnion<unsigned *, LValue *> Pos,
                            const OMPTaskDataTy::DependData &Data,
                            Address DependenciesArray) {
   CodeGenModule &CGM = CGF.CGM;
   ASTContext &C = CGM.getContext();
   QualType FlagsTy;
   getDependTypes(C, KmpDependInfoTy, FlagsTy);
   RecordDecl *KmpDependInfoRD =
       cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl());
   llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy);
 
   OMPIteratorGeneratorScope IteratorScope(
       CGF, cast_or_null<OMPIteratorExpr>(
                Data.IteratorExpr ? Data.IteratorExpr->IgnoreParenImpCasts()
                                  : nullptr));
   for (const Expr *E : Data.DepExprs) {
-    const auto *OASE = dyn_cast<OMPArrayShapingExpr>(E);
     llvm::Value *Addr;
-    if (OASE) {
-      const Expr *Base = OASE->getBase();
-      Addr = CGF.EmitScalarExpr(Base);
-    } else {
-      Addr = CGF.EmitLValue(E).getPointer(CGF);
-    }
     llvm::Value *Size;
-    QualType Ty = E->getType();
-    if (OASE) {
-      Size = CGF.getTypeSize(OASE->getBase()->getType()->getPointeeType());
-      for (const Expr *SE : OASE->getDimensions()) {
-        llvm::Value *Sz = CGF.EmitScalarExpr(SE);
-        Sz = CGF.EmitScalarConversion(Sz, SE->getType(),
-                                      CGF.getContext().getSizeType(),
-                                      SE->getExprLoc());
-        Size = CGF.Builder.CreateNUWMul(Size, Sz);
-      }
-    } else if (const auto *ASE =
-                   dyn_cast<OMPArraySectionExpr>(E->IgnoreParenImpCasts())) {
-      LValue UpAddrLVal =
-          CGF.EmitOMPArraySectionExpr(ASE, /*IsLowerBound=*/false);
-      llvm::Value *UpAddr = CGF.Builder.CreateConstGEP1_32(
-          UpAddrLVal.getPointer(CGF), /*Idx0=*/1);
-      llvm::Value *LowIntPtr = CGF.Builder.CreatePtrToInt(Addr, CGM.SizeTy);
-      llvm::Value *UpIntPtr = CGF.Builder.CreatePtrToInt(UpAddr, CGM.SizeTy);
-      Size = CGF.Builder.CreateNUWSub(UpIntPtr, LowIntPtr);
-    } else {
-      Size = CGF.getTypeSize(Ty);
-    }
+    std::tie(Addr, Size) = getPointerAndSize(CGF, E);
     LValue Base;
     if (unsigned *P = Pos.dyn_cast<unsigned *>()) {
       Base = CGF.MakeAddrLValue(
           CGF.Builder.CreateConstGEP(DependenciesArray, *P), KmpDependInfoTy);
     } else {
       LValue &PosLVal = *Pos.get<LValue *>();
       llvm::Value *Idx = CGF.EmitLoadOfScalar(PosLVal, E->getExprLoc());
       Base = CGF.MakeAddrLValue(

clang/test/OpenMP/task_affinity_codegen.cpp

+// RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp -fopenmp-version=50 -x c++ -emit-llvm %s -o - | FileCheck %s
+// RUN: %clang_cc1 -fopenmp -fopenmp-version=50 -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s
+// RUN: %clang_cc1 -fopenmp -fopenmp-version=50 -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - | FileCheck %s
+//
+// RUN: %clang_cc1 -verify -triple x86_64-apple-darwin10 -fopenmp-simd -fopenmp-version=50 -x c++ -emit-llvm %s -o - | FileCheck --check-prefix SIMD-ONLY0 %s
+// RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=50 -x c++ -triple x86_64-apple-darwin10 -emit-pch -o %t %s
+// RUN: %clang_cc1 -fopenmp-simd -fopenmp-version=50 -x c++ -triple x86_64-apple-darwin10 -include-pch %t -verify %s -emit-llvm -o - | FileCheck --check-prefix SIMD-ONLY0 %s
+// SIMD-ONLY0-NOT: {{__kmpc|__tgt}}
+// expected-no-diagnostics
+#ifndef HEADER
+#define HEADER
+
+// CHECK-LABEL: @main
+int main() {
+  float *p;
+  int a = 10;
+  // kmp_task_affinity_info_t affs[1];
+  // CHECK: [[AFFS_ADDR:%.+]] = alloca [1 x %struct.kmp_task_affinity_info_t],
+  // CHECK: [[TD:%.+]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @{{.+}}, i32 [[GTID:%.+]], i32 1, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %{{.+}}*)* @{{.+}} to i32 (i32, i8*)*))
+  // CHECK: [[AFFINE_LST_ADDR:%.+]] = getelementptr inbounds [1 x %struct.kmp_task_affinity_info_t], [1 x %struct.kmp_task_affinity_info_t]* [[AFFS_ADDR]], i64 0, i64 0
+  // CHECK: [[P:%.+]] = load float*, float** [[P_ADDR:%.+]],
+  // CHECK: [[A_VAL:%.+]] = load i32, i32* [[A_ADDR:%.+]],
+  // CHECK: [[A_SZ:%.+]] = sext i32 [[A_VAL]] to i64
+  // CHECK: [[BYTES:%.+]] = mul nuw i64 4, [[A_SZ]]
+  // CHECK: [[SZ:%.+]] = mul nuw i64 [[BYTES]], 10
+  // CHECK: [[A_VAL:%.+]] = load i32, i32* [[A_ADDR]],
+  // CHECK: [[A_SZ1:%.+]] = sext i32 [[A_VAL]] to i64
+  // CHECK: [[SIZE:%.+]] = mul nuw i64 [[SZ]], [[A_SZ]]
+  // CHECK: [[AFFS_0_ADDR:%.+]] = getelementptr %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFINE_LST_ADDR]], i64 0
+
+  // affs[0].base = p;
+  // CHECK: [[AFFS_0_BASE_ADDR:%.+]] = getelementptr inbounds %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFS_0_ADDR]], i32 0, i32 0
+  // CHECK: [[P_INTPTR:%.+]] = ptrtoint float* [[P]] to i64
+  // CHECK: store i64 [[P_INTPTR]], i64* [[AFFS_0_BASE_ADDR]],
+
+  // affs[0].size = sizeof(*p) * a * 10 * a;
+  // CHECK: [[AFFS_0_SIZE_ADDR:%.+]] = getelementptr inbounds %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFS_0_ADDR]], i32 0, i32 1
+  // CHECK: store i64 [[SIZE]], i64* [[AFFS_0_SIZE_ADDR]],
+  // CHECK: [[BC:%.+]] = bitcast %struct.kmp_task_affinity_info_t* [[AFFINE_LST_ADDR]] to i8*
+  // CHECK: call i32 @__kmpc_omp_reg_task_with_affinity(%struct.ident_t* @{{.+}}, i32 [[GTID]], i8* [[TD]], i32 1, i8* [[BC]])
+#pragma omp task affinity(([a][10][a])p)
+  ;
+  // CHECK: [[TD:%.+]] = call i8* @__kmpc_omp_task_alloc(%struct.ident_t* @{{.+}}, i32 [[GTID]], i32 1, i64 40, i64 1, i32 (i32, i8*)* bitcast (i32 (i32, %{{.+}}*)* @{{.+}} to i32 (i32, i8*)*))
+  // CHECK: [[A_VAL:%.+]] = load i32, i32* [[A_ADDR]],
+  // CHECK: [[SUB:%.+]] = sub nsw i32 [[A_VAL]], 0
+  // CHECK: [[CONV:%.+]] = zext i32 [[SUB]] to i64
+
+  // <num_elem> = <num_iters> + 1 constant affinity for affinity(a)
+  // CHECK: [[NUM_ELEMS:%.+]] = add nuw i64 1, [[CONV]]
+  // CHECK: [[SV:%.+]] = call i8* @llvm.stacksave()
+  // CHECK: store i8* [[SV]], i8** [[SV_ADDR:%.+]],
+
+  // kmp_task_affinity_info_t affs[<num_elem>];
+  // CHECK: [[AFFS_ADDR:%.+]] = alloca %struct.kmp_task_affinity_info_t, i64 [[NUM_ELEMS]],
+  // store i64 %21, i64* %__vla_expr0, align 8
+  // CHECK: [[NAFFS:%.+]] = trunc i64 [[NUM_ELEMS]] to i32
+  // CHECK: [[AFFS_0_ADDR:%.+]] = getelementptr %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFS_ADDR]], i64 0
+
+  // affs[0].base = &a;
+  // CHECK: [[AFFS_0_BASE_ADDR:%.+]] = getelementptr inbounds %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFS_0_ADDR]], i32 0, i32 0
+  // CHECK: [[A_INTPTR:%.+]] = ptrtoint i32* [[A_ADDR]] to i64
+  // CHECK: store i64 [[A_INTPTR]], i64* [[AFFS_0_BASE_ADDR]],
+
+  // affs[0].size = sizeof(a);
+  // CHECK: [[AFFS_0_SIZE_ADDR:%.+]] = getelementptr inbounds %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFS_0_ADDR]], i32 0, i32 1
+  // CHECK: store i64 4, i64* [[AFFS_0_SIZE_ADDR]],
+
+  // affs_cnt = 1;
+  // CHECK: store i64 1, i64* [[AFFS_CNT_ADDR:%.+]],
+  // CHECK: [[A_VAL:%.+]] = load i32, i32* [[A_ADDR]],
+  // CHECK: [[NITERS:%.+]] = sub nsw i32 [[A_VAL]], 0
+  // CHECK: store i32 0, i32* [[CNT_ADDR:%.+]],
+  // CHECK: br label %[[CONT:[^,]+]]
+
+  //for (int cnt = 0; cnt < (a-0); ++cnt) {
+  // int i = cnt + 0;
+  // affs[affs_cnt].base = &p[i];
+  // affs[affs_cnt].size = sizeof(p[i]);
+  // ++affs_cnt;
+  // }
+
+  // CHECK: [[CONT]]:
+  // CHECK: [[CNT:%.+]] = load i32, i32* [[CNT_ADDR]],
+  // CHECK: [[CMP:%.+]] = icmp slt i32 [[CNT]], [[NITERS]]
+  // CHECK: br i1 [[CMP]], label %[[BODY:[^,]+]], label %[[DONE:[^,]+]]
+
+  // CHECK: [[BODY]]:
+  // i = cnt + 0;
+  // CHECK: [[CNT:%.+]] = load i32, i32* [[CNT_ADDR]],
+  // CHECK: [[VAL:%.+]] = add nsw i32 0, [[CNT]]
+  // CHECK: store i32 [[VAL]], i32* [[I_ADDR:%.+]],
+
+  // &p[i]
+  // CHECK: [[P:%.+]] = load float*, float** [[P_ADDR]],
+  // CHECK: [[I:%.+]] = load i32, i32* [[I_ADDR]],
+  // CHECK: [[IDX:%.+]] = sext i32 [[I]] to i64
+  // CHECK: [[P_I_ADDR:%.+]] = getelementptr inbounds float, float* [[P]], i64 [[IDX]]
+
+  // affs[affs_cnt]
+  // CHECK: [[AFFS_CNT:%.+]] = load i64, i64* [[AFFS_CNT_ADDR]],
+  // CHECK: [[AFFS_ELEM_ADDR:%.+]] = getelementptr %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFS_ADDR]], i64 [[AFFS_CNT]]
+
+  // affs[affs_cnt].base = &p[i];
+  // CHECK: [[AFFS_ELEM_BASE_ADDR:%.+]] = getelementptr inbounds %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFS_ELEM_ADDR]], i32 0, i32 0
+  // CHECK: [[CAST:%.+]] = ptrtoint float* [[P_I_ADDR]] to i64
+  // CHECK: store i64 [[CAST]], i64* [[AFFS_ELEM_BASE_ADDR]],
+
+  // affs[affs_cnt].size = sizeof(p[i]);
+  // CHECK: [[AFFS_ELEM_SIZE_ADDR:%.+]] = getelementptr inbounds %struct.kmp_task_affinity_info_t, %struct.kmp_task_affinity_info_t* [[AFFS_ELEM_ADDR]], i32 0, i32 1
+  // CHECK: store i64 4, i64* [[AFFS_ELEM_SIZE_ADDR]],
+
+  // ++affs_cnt;
+  // CHECK: [[AFFS_CNT_NEXT:%.+]] = add nuw i64 [[AFFS_CNT]], 1
+  // CHECK: store i64 [[AFFS_CNT_NEXT]], i64* [[AFFS_CNT_ADDR]],
+
+  // ++cnt;
+  // CHECK: [[CNT:%.+]] = load i32, i32* [[CNT_ADDR]],
+  // CHECK: [[CNT_NEXT:%.+]] = add nsw i32 [[CNT]], 1
+  // CHECK: store i32 [[CNT_NEXT]], i32* [[CNT_ADDR]],
+  // CHECK: br label %[[CONT]]
+
+  // CHECK: [[DONE]]:
+  // CHECK: [[BC:%.+]] = bitcast %struct.kmp_task_affinity_info_t* [[AFFS_ADDR]] to i8*
+  // CHECK: call i32 @__kmpc_omp_reg_task_with_affinity(%struct.ident_t* @{{.+}} i32 [[GTID]], i8* [[TD]], i32 [[NAFFS]], i8* [[BC]])
+  // CHECK: [[SV:%.+]] = load i8*, i8** [[SV_ADDR]],
+  // CHECK: call void @llvm.stackrestore(i8* [[SV]])
+#pragma omp task affinity(iterator(i=0:a): p[i]) affinity(a)
+  ;
+  return 0;
+}
+
+#endif

llvm/include/llvm/Frontend/OpenMP/OMPKinds.def

 __OMP_RTL(__kmpc_fork_call, true, Void, IdentPtr, Int32, ParallelTaskPtr)
 __OMP_RTL(__kmpc_omp_taskwait, false, Int32, IdentPtr, Int32)
 __OMP_RTL(__kmpc_omp_taskyield, false, Int32, IdentPtr, Int32, Int32)
 __OMP_RTL(__kmpc_push_num_threads, false, Void, IdentPtr, Int32,
           /* Int */ Int32)
 __OMP_RTL(__kmpc_push_proc_bind, false, Void, IdentPtr, Int32, /* Int */ Int32)
 __OMP_RTL(__kmpc_serialized_parallel, false, Void, IdentPtr, Int32)
 __OMP_RTL(__kmpc_end_serialized_parallel, false, Void, IdentPtr, Int32)
+__OMP_RTL(__kmpc_omp_reg_task_with_affinity, false, Int32, IdentPtr, Int32,
+          Int8Ptr, Int32, Int8Ptr)
 
 __OMP_RTL(omp_get_thread_num, false, Int32, )
 __OMP_RTL(omp_get_num_threads, false, Int32, )
 __OMP_RTL(omp_get_max_threads, false, Int32, )
 __OMP_RTL(omp_in_parallel, false, Int32, )
 __OMP_RTL(omp_get_dynamic, false, Int32, )
 __OMP_RTL(omp_get_cancellation, false, Int32, )
 __OMP_RTL(omp_get_nested, false, Int32, )