[OpenACC] Add 'collapse' clause AST/basic Sema implementation (#109461)

The 'collapse' clause on a 'loop' construct is used to specify how many
nested loops are associated with the 'loop' construct. It takes an
optional 'force' tag, and an integer constant expression as arguments.

There are many other restrictions based on the contents of the loop/etc,
but those are implemented in followup patches, for now, this patch just
adds the AST node and does basic argument checking on the loop-count.

Author: erichkeane

clang/include/clang/AST/OpenACCClause.h

@@ -547,6 +547,32 @@ class OpenACCAsyncClause : public OpenACCClauseWithSingleIntExpr {
                                     SourceLocation EndLoc);
 };
 
+/// Represents a 'collapse' clause on a 'loop' construct. This clause takes an
+/// integer constant expression 'N' that represents how deep to collapse the
+/// construct. It also takes an optional 'force' tag that permits intervening
+/// code in the loops.
+class OpenACCCollapseClause : public OpenACCClauseWithSingleIntExpr {
+  bool HasForce = false;
+
+  OpenACCCollapseClause(SourceLocation BeginLoc, SourceLocation LParenLoc,
+                        bool HasForce, Expr *LoopCount, SourceLocation EndLoc);
+
+public:
+  const Expr *getLoopCount() const { return getIntExpr(); }
+  Expr *getLoopCount() { return getIntExpr(); }
+
+  bool hasForce() const { return HasForce; }
+
+  static bool classof(const OpenACCClause *C) {
+    return C->getClauseKind() == OpenACCClauseKind::Collapse;
+  }
+
+  static OpenACCCollapseClause *Create(const ASTContext &C,
+                                       SourceLocation BeginLoc,
+                                       SourceLocation LParenLoc, bool HasForce,
+                                       Expr *LoopCount, SourceLocation EndLoc);
+};
+
 /// Represents a clause with one or more 'var' objects, represented as an expr,
 /// as its arguments. Var-list is expected to be stored in trailing storage.
 /// For now, we're just storing the original expression in its entirety, unlike

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -12606,6 +12606,9 @@ def note_acc_construct_here : Note<"'%0' construct is here">;
 def err_acc_loop_spec_conflict
     : Error<"OpenACC clause '%0' on '%1' construct conflicts with previous "
             "data dependence clause">;
+def err_acc_collapse_loop_count
+    : Error<"OpenACC 'collapse' clause loop count must be a %select{constant "
+            "expression|positive integer value, evaluated to %1}0">;
 
 // AMDGCN builtins diagnostics
 def err_amdgcn_global_load_lds_size_invalid_value : Error<"invalid size value">;

clang/include/clang/Basic/OpenACCClauses.def

@@ -24,6 +24,7 @@
 VISIT_CLAUSE(Auto)
 VISIT_CLAUSE(Async)
 VISIT_CLAUSE(Attach)
+VISIT_CLAUSE(Collapse)
 VISIT_CLAUSE(Copy)
 CLAUSE_ALIAS(PCopy, Copy, true)
 CLAUSE_ALIAS(PresentOrCopy, Copy, true)

clang/include/clang/Sema/SemaOpenACC.h

@@ -87,9 +87,14 @@ class SemaOpenACC : public SemaBase {
       SmallVector<Expr *> VarList;
     };
 
+    struct CollapseDetails {
+      bool IsForce;
+      Expr *LoopCount;
+    };
+
     std::variant<std::monostate, DefaultDetails, ConditionDetails,
                  IntExprDetails, VarListDetails, WaitDetails, DeviceTypeDetails,
-                 ReductionDetails>
+                 ReductionDetails, CollapseDetails>
         Details = std::monostate{};
 
   public:
@@ -246,6 +251,18 @@ class SemaOpenACC : public SemaBase {
       return std::get<VarListDetails>(Details).IsZero;
     }
 
+    bool isForce() const {
+      assert(ClauseKind == OpenACCClauseKind::Collapse &&
+             "Only 'collapse' has a force tag");
+      return std::get<CollapseDetails>(Details).IsForce;
+    }
+
+    Expr *getLoopCount() const {
+      assert(ClauseKind == OpenACCClauseKind::Collapse &&
+             "Only 'collapse' has a loop count");
+      return std::get<CollapseDetails>(Details).LoopCount;
+    }
+
     ArrayRef<DeviceTypeArgument> getDeviceTypeArchitectures() const {
       assert((ClauseKind == OpenACCClauseKind::DeviceType ||
               ClauseKind == OpenACCClauseKind::DType) &&
@@ -384,6 +401,12 @@ class SemaOpenACC : public SemaBase {
              "Only 'device_type'/'dtype' has a device-type-arg list");
       Details = DeviceTypeDetails{std::move(Archs)};
     }
+
+    void setCollapseDetails(bool IsForce, Expr *LoopCount) {
+      assert(ClauseKind == OpenACCClauseKind::Collapse &&
+             "Only 'collapse' has collapse details");
+      Details = CollapseDetails{IsForce, LoopCount};
+    }
   };
 
   SemaOpenACC(Sema &S);
@@ -448,6 +471,8 @@ class SemaOpenACC : public SemaBase {
                                    Expr *LowerBound,
                                    SourceLocation ColonLocFirst, Expr *Length,
                                    SourceLocation RBLoc);
+  /// Checks the loop depth value for a collapse clause.
+  ExprResult CheckCollapseLoopCount(Expr *LoopCount);
 
   /// Helper type for the registration/assignment of constructs that need to
   /// 'know' about their parent constructs and hold a reference to them, such as

clang/lib/AST/OpenACCClause.cpp

@@ -43,7 +43,7 @@ bool OpenACCClauseWithCondition::classof(const OpenACCClause *C) {
 bool OpenACCClauseWithSingleIntExpr::classof(const OpenACCClause *C) {
   return OpenACCNumWorkersClause::classof(C) ||
          OpenACCVectorLengthClause::classof(C) ||
-         OpenACCAsyncClause::classof(C);
+         OpenACCCollapseClause::classof(C) || OpenACCAsyncClause::classof(C);
 }
 OpenACCDefaultClause *OpenACCDefaultClause::Create(const ASTContext &C,
                                                    OpenACCDefaultClauseKind K,
@@ -134,6 +134,30 @@ OpenACCNumWorkersClause::Create(const ASTContext &C, SourceLocation BeginLoc,
       OpenACCNumWorkersClause(BeginLoc, LParenLoc, IntExpr, EndLoc);
 }
 
+OpenACCCollapseClause::OpenACCCollapseClause(SourceLocation BeginLoc,
+                                             SourceLocation LParenLoc,
+                                             bool HasForce, Expr *LoopCount,
+                                             SourceLocation EndLoc)
+    : OpenACCClauseWithSingleIntExpr(OpenACCClauseKind::Collapse, BeginLoc,
+                                     LParenLoc, LoopCount, EndLoc),
+      HasForce(HasForce) {
+  assert(LoopCount && "LoopCount required");
+}
+
+OpenACCCollapseClause *
+OpenACCCollapseClause::Create(const ASTContext &C, SourceLocation BeginLoc,
+                              SourceLocation LParenLoc, bool HasForce,
+                              Expr *LoopCount, SourceLocation EndLoc) {
+  assert(
+      LoopCount &&
+      (LoopCount->isInstantiationDependent() || isa<ConstantExpr>(LoopCount)) &&
+      "Loop count not constant expression");
+  void *Mem =
+      C.Allocate(sizeof(OpenACCCollapseClause), alignof(OpenACCCollapseClause));
+  return new (Mem)
+      OpenACCCollapseClause(BeginLoc, LParenLoc, HasForce, LoopCount, EndLoc);
+}
+
 OpenACCVectorLengthClause::OpenACCVectorLengthClause(SourceLocation BeginLoc,
                                                      SourceLocation LParenLoc,
                                                      Expr *IntExpr,
@@ -550,3 +574,11 @@ void OpenACCClausePrinter::VisitIndependentClause(
 void OpenACCClausePrinter::VisitSeqClause(const OpenACCSeqClause &C) {
   OS << "seq";
 }
+
+void OpenACCClausePrinter::VisitCollapseClause(const OpenACCCollapseClause &C) {
+  OS << "collapse(";
+  if (C.hasForce())
+    OS << "force:";
+  printExpr(C.getLoopCount());
+  OS << ")";
+}

clang/lib/AST/StmtProfile.cpp

@@ -2558,6 +2558,12 @@ void OpenACCClauseProfiler::VisitNumWorkersClause(
   Profiler.VisitStmt(Clause.getIntExpr());
 }
 
+void OpenACCClauseProfiler::VisitCollapseClause(
+    const OpenACCCollapseClause &Clause) {
+  assert(Clause.getLoopCount() && "collapse clause requires a valid int expr");
+  Profiler.VisitStmt(Clause.getLoopCount());
+}
+
 void OpenACCClauseProfiler::VisitPrivateClause(
     const OpenACCPrivateClause &Clause) {
   for (auto *E : Clause.getVarList())

clang/lib/AST/TextNodeDumper.cpp

@@ -419,6 +419,12 @@ void TextNodeDumper::Visit(const OpenACCClause *C) {
       // but print 'clause' here so it is clear what is happening from the dump.
       OS << " clause";
       break;
+    case OpenACCClauseKind::Collapse:
+      OS << " clause";
+      if (cast<OpenACCCollapseClause>(C)->hasForce())
+        OS << ": force";
+      break;
+
     case OpenACCClauseKind::CopyIn:
     case OpenACCClauseKind::PCopyIn:
     case OpenACCClauseKind::PresentOrCopyIn:

clang/lib/Parse/ParseOpenACC.cpp

@@ -976,14 +976,25 @@ Parser::OpenACCClauseParseResult Parser::ParseOpenACCClauseParams(
                                      /*IsReadOnly=*/false, /*IsZero=*/false);
       break;
     case OpenACCClauseKind::Collapse: {
-      tryParseAndConsumeSpecialTokenKind(*this, OpenACCSpecialTokenKind::Force,
-                                         ClauseKind);
-      ExprResult NumLoops =
+      bool HasForce = tryParseAndConsumeSpecialTokenKind(
+          *this, OpenACCSpecialTokenKind::Force, ClauseKind);
+      ExprResult LoopCount =
           getActions().CorrectDelayedTyposInExpr(ParseConstantExpression());
-      if (NumLoops.isInvalid()) {
+      if (LoopCount.isInvalid()) {
         Parens.skipToEnd();
         return OpenACCCanContinue();
       }
+
+      LoopCount = getActions().OpenACC().ActOnIntExpr(
+          OpenACCDirectiveKind::Invalid, ClauseKind,
+          LoopCount.get()->getBeginLoc(), LoopCount.get());
+
+      if (LoopCount.isInvalid()) {
+        Parens.skipToEnd();
+        return OpenACCCanContinue();
+      }
+
+      ParsedClause.setCollapseDetails(HasForce, LoopCount.get());
       break;
     }
     case OpenACCClauseKind::Bind: {

clang/lib/Sema/SemaOpenACC.cpp

@@ -343,6 +343,18 @@ bool doesClauseApplyToDirective(OpenACCDirectiveKind DirectiveKind,
       return false;
     }
 
+  case OpenACCClauseKind::Collapse: {
+    switch (DirectiveKind) {
+    case OpenACCDirectiveKind::Loop:
+    case OpenACCDirectiveKind::ParallelLoop:
+    case OpenACCDirectiveKind::SerialLoop:
+    case OpenACCDirectiveKind::KernelsLoop:
+      return true;
+    default:
+      return false;
+    }
+  }
+
   default:
     // Do nothing so we can go to the 'unimplemented' diagnostic instead.
     return true;
@@ -1037,6 +1049,26 @@ OpenACCClause *SemaOpenACCClauseVisitor::VisitReductionClause(
       ValidVars, Clause.getEndLoc());
 }
 
+OpenACCClause *SemaOpenACCClauseVisitor::VisitCollapseClause(
+    SemaOpenACC::OpenACCParsedClause &Clause) {
+  // Duplicates here are not really sensible.  We could possible permit
+  // multiples if they all had the same value, but there isn't really a good
+  // reason to do so. Also, this simplifies the suppression of duplicates, in
+  // that we know if we 'find' one after instantiation, that it is the same
+  // clause, which simplifies instantiation/checking/etc.
+  if (checkAlreadyHasClauseOfKind(SemaRef, ExistingClauses, Clause))
+    return nullptr;
+
+  ExprResult LoopCount = SemaRef.CheckCollapseLoopCount(Clause.getLoopCount());
+
+  if (!LoopCount.isUsable())
+    return nullptr;
+
+  return OpenACCCollapseClause::Create(Ctx, Clause.getBeginLoc(),
+                                       Clause.getLParenLoc(), Clause.isForce(),
+                                       LoopCount.get(), Clause.getEndLoc());
+}
+
 } // namespace
 
 SemaOpenACC::SemaOpenACC(Sema &S) : SemaBase(S) {}
@@ -1273,6 +1305,9 @@ ExprResult SemaOpenACC::ActOnIntExpr(OpenACCDirectiveKind DK,
     }
   } IntExprDiagnoser(DK, CK, IntExpr);
 
+  if (!IntExpr)
+    return ExprError();
+
   ExprResult IntExprResult = SemaRef.PerformContextualImplicitConversion(
       Loc, IntExpr, IntExprDiagnoser);
   if (IntExprResult.isInvalid())
@@ -1583,6 +1618,34 @@ ExprResult SemaOpenACC::ActOnArraySectionExpr(Expr *Base, SourceLocation LBLoc,
                        OK_Ordinary, ColonLoc, RBLoc);
 }
 
+ExprResult SemaOpenACC::CheckCollapseLoopCount(Expr *LoopCount) {
+  if (!LoopCount)
+    return ExprError();
+
+  assert((LoopCount->isInstantiationDependent() ||
+          LoopCount->getType()->isIntegerType()) &&
+         "Loop argument non integer?");
+
+  // If this is dependent, there really isn't anything we can check.
+  if (LoopCount->isInstantiationDependent())
+    return ExprResult{LoopCount};
+
+  std::optional<llvm::APSInt> ICE =
+      LoopCount->getIntegerConstantExpr(getASTContext());
+
+  // OpenACC 3.3: 2.9.1
+  // The argument to the collapse clause must be a constant positive integer
+  // expression.
+  if (!ICE || *ICE <= 0) {
+    Diag(LoopCount->getBeginLoc(), diag::err_acc_collapse_loop_count)
+        << ICE.has_value() << ICE.value_or(llvm::APSInt{}).getExtValue();
+    return ExprError();
+  }
+
+  return ExprResult{
+      ConstantExpr::Create(getASTContext(), LoopCount, APValue{*ICE})};
+}
+
 bool SemaOpenACC::ActOnStartStmtDirective(OpenACCDirectiveKind K,
                                           SourceLocation StartLoc) {
   SemaRef.DiscardCleanupsInEvaluationContext();

clang/lib/Sema/TreeTransform.h

@@ -11822,6 +11822,31 @@ void OpenACCClauseTransform<Derived>::VisitReductionClause(
       ParsedClause.getLParenLoc(), C.getReductionOp(), ValidVars,
       ParsedClause.getEndLoc());
 }
+
+template <typename Derived>
+void OpenACCClauseTransform<Derived>::VisitCollapseClause(
+    const OpenACCCollapseClause &C) {
+  Expr *LoopCount = const_cast<Expr *>(C.getLoopCount());
+  assert(LoopCount && "collapse clause constructed with invalid loop count");
+
+  ExprResult NewLoopCount = Self.TransformExpr(LoopCount);
+
+  NewLoopCount = Self.getSema().OpenACC().ActOnIntExpr(
+      OpenACCDirectiveKind::Invalid, ParsedClause.getClauseKind(),
+      NewLoopCount.get()->getBeginLoc(), NewLoopCount.get());
+
+  NewLoopCount =
+      Self.getSema().OpenACC().CheckCollapseLoopCount(NewLoopCount.get());
+
+  if (!NewLoopCount.isUsable())
+    return;
+
+  ParsedClause.setCollapseDetails(C.hasForce(), NewLoopCount.get());
+  NewClause = OpenACCCollapseClause::Create(
+      Self.getSema().getASTContext(), ParsedClause.getBeginLoc(),
+      ParsedClause.getLParenLoc(), ParsedClause.isForce(),
+      ParsedClause.getLoopCount(), ParsedClause.getEndLoc());
+}
 } // namespace
 template <typename Derived>
 OpenACCClause *TreeTransform<Derived>::TransformOpenACCClause(