[Flang] Add standalone tile support

flang/lib/Lower/OpenMP/ClauseProcessor.cpp

@@ -431,6 +431,19 @@ bool ClauseProcessor::processNumTasks(
   return false;
 }
 
+bool ClauseProcessor::processSizes(StatementContext &stmtCtx,
+                                   mlir::omp::SizesClauseOps &result) const {
+  if (auto *clause = findUniqueClause<omp::clause::Sizes>()) {
+    result.sizes.reserve(clause->v.size());
+    for (const ExprTy &vv : clause->v)
+      result.sizes.push_back(fir::getBase(converter.genExprValue(vv, stmtCtx)));
+
+    return true;
+  }
+
+  return false;
+}
+
 bool ClauseProcessor::processNumTeams(
     lower::StatementContext &stmtCtx,
     mlir::omp::NumTeamsClauseOps &result) const {

flang/lib/Lower/OpenMP/ClauseProcessor.h

@@ -66,6 +66,8 @@ class ClauseProcessor {
                   mlir::omp::LoopRelatedClauseOps &loopResult,
                   mlir::omp::CollapseClauseOps &collapseResult,
                   llvm::SmallVectorImpl<const semantics::Symbol *> &iv) const;
+  bool processSizes(StatementContext &stmtCtx,
+                    mlir::omp::SizesClauseOps &result) const;
   bool processDevice(lower::StatementContext &stmtCtx,
                      mlir::omp::DeviceClauseOps &result) const;
   bool processDeviceType(mlir::omp::DeviceTypeClauseOps &result) const;

flang/lib/Lower/OpenMP/Utils.cpp

@@ -652,7 +652,6 @@ int64_t collectLoopRelatedInfo(
     mlir::omp::LoopRelatedClauseOps &result,
     llvm::SmallVectorImpl<const semantics::Symbol *> &iv) {
   int64_t numCollapse = 1;
-  fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
 
   // Collect the loops to collapse.
   lower::pft::Evaluation *doConstructEval = &eval.getFirstNestedEvaluation();
@@ -667,6 +666,25 @@ int64_t collectLoopRelatedInfo(
     numCollapse = collapseValue;
   }
 
+  collectLoopRelatedInfo(converter, currentLocation, eval, numCollapse, result,
+                         iv);
+  return numCollapse;
+}
+
+void collectLoopRelatedInfo(
+    lower::AbstractConverter &converter, mlir::Location currentLocation,
+    lower::pft::Evaluation &eval, int64_t numCollapse,
+    mlir::omp::LoopRelatedClauseOps &result,
+    llvm::SmallVectorImpl<const semantics::Symbol *> &iv) {
+
+  fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
+
+  // Collect the loops to collapse.
+  lower::pft::Evaluation *doConstructEval = &eval.getFirstNestedEvaluation();
+  if (doConstructEval->getIf<parser::DoConstruct>()->IsDoConcurrent()) {
+    TODO(currentLocation, "Do Concurrent in Worksharing loop construct");
+  }
+
   // Collect sizes from tile directive if present.
   std::int64_t sizesLengthValue = 0l;
   if (auto *ompCons{eval.getIf<parser::OpenMPConstruct>()}) {
@@ -676,7 +694,7 @@ int64_t collectLoopRelatedInfo(
         });
   }
 
-  collapseValue = std::max(collapseValue, sizesLengthValue);
+  std::int64_t collapseValue = std::max(numCollapse, sizesLengthValue);
   std::size_t loopVarTypeSize = 0;
   do {
     lower::pft::Evaluation *doLoop =
@@ -709,8 +727,6 @@ int64_t collectLoopRelatedInfo(
   } while (collapseValue > 0);
 
   convertLoopBounds(converter, currentLocation, result, loopVarTypeSize);
-
-  return numCollapse;
 }
 
 } // namespace omp

flang/lib/Lower/OpenMP/Utils.h

@@ -165,6 +165,13 @@ int64_t collectLoopRelatedInfo(
     mlir::omp::LoopRelatedClauseOps &result,
     llvm::SmallVectorImpl<const semantics::Symbol *> &iv);
 
+void collectLoopRelatedInfo(
+    lower::AbstractConverter &converter, mlir::Location currentLocation,
+    lower::pft::Evaluation &eval, std::int64_t collapseValue,
+    // const omp::List<omp::Clause> &clauses,
+    mlir::omp::LoopRelatedClauseOps &result,
+    llvm::SmallVectorImpl<const semantics::Symbol *> &iv);
+
 void collectTileSizesFromOpenMPConstruct(
     const parser::OpenMPConstruct *ompCons,
     llvm::SmallVectorImpl<int64_t> &tileSizes,

flang/lib/Semantics/check-directive-structure.h

@@ -383,7 +383,8 @@ class DirectiveStructureChecker : public virtual BaseChecker {
       const C &clause, const parser::ScalarIntConstantExpr &i);
 
   void RequiresPositiveParameter(const C &clause,
-      const parser::ScalarIntExpr &i, llvm::StringRef paramName = "parameter");
+      const parser::ScalarIntExpr &i, llvm::StringRef paramName = "parameter",
+      bool allowZero = true);
 
   void OptionalConstantPositiveParameter(
       const C &clause, const std::optional<parser::ScalarIntConstantExpr> &o);
@@ -657,9 +658,9 @@ void DirectiveStructureChecker<D, C, PC, ClauseEnumSize>::SayNotMatching(
 template <typename D, typename C, typename PC, std::size_t ClauseEnumSize>
 void DirectiveStructureChecker<D, C, PC,
     ClauseEnumSize>::RequiresPositiveParameter(const C &clause,
-    const parser::ScalarIntExpr &i, llvm::StringRef paramName) {
+    const parser::ScalarIntExpr &i, llvm::StringRef paramName, bool allowZero) {
   if (const auto v{GetIntValue(i)}) {
-    if (*v < 0) {
+    if (*v < (allowZero ? 0 : 1)) {
       context_.Say(GetContext().clauseSource,
           "The %s of the %s clause must be "
           "a positive integer expression"_err_en_US,

flang/lib/Semantics/check-omp-structure.cpp

@@ -2749,6 +2749,13 @@ void OmpStructureChecker::Enter(const parser::OmpClause &x) {
   }
 }
 
+void OmpStructureChecker::Enter(const parser::OmpClause::Sizes &c) {
+  CheckAllowedClause(llvm::omp::Clause::OMPC_sizes);
+  for (const parser::Cosubscript &v : c.v)
+    RequiresPositiveParameter(llvm::omp::Clause::OMPC_sizes, v,
+        /*paramName=*/"parameter", /*allowZero=*/false);
+}
+
 // Following clauses do not have a separate node in parse-tree.h.
 CHECK_SIMPLE_CLAUSE(Absent, OMPC_absent)
 CHECK_SIMPLE_CLAUSE(Affinity, OMPC_affinity)
@@ -2790,7 +2797,6 @@ CHECK_SIMPLE_CLAUSE(Notinbranch, OMPC_notinbranch)
 CHECK_SIMPLE_CLAUSE(Partial, OMPC_partial)
 CHECK_SIMPLE_CLAUSE(ProcBind, OMPC_proc_bind)
 CHECK_SIMPLE_CLAUSE(Simd, OMPC_simd)
-CHECK_SIMPLE_CLAUSE(Sizes, OMPC_sizes)
 CHECK_SIMPLE_CLAUSE(Permutation, OMPC_permutation)
 CHECK_SIMPLE_CLAUSE(Uniform, OMPC_uniform)
 CHECK_SIMPLE_CLAUSE(Unknown, OMPC_unknown)

flang/lib/Semantics/resolve-directives.cpp

@@ -149,6 +149,9 @@ template <typename T> class DirectiveAttributeVisitor {
     dataSharingAttributeObjects_.clear();
   }
   bool HasDataSharingAttributeObject(const Symbol &);
+  std::tuple<const parser::Name *, const parser::ScalarExpr *,
+      const parser::ScalarExpr *, const parser::ScalarExpr *>
+  GetLoopBounds(const parser::DoConstruct &);
   const parser::Name *GetLoopIndex(const parser::DoConstruct &);
   const parser::DoConstruct *GetDoConstructIf(
       const parser::ExecutionPartConstruct &);
@@ -933,6 +936,13 @@ class OmpAttributeVisitor : DirectiveAttributeVisitor<llvm::omp::Directive> {
     privateDataSharingAttributeObjects_.clear();
   }
 
+  /// Check that loops in the loop nest are perfectly nested, as well that lower
+  /// bound, upper bound, and step expressions do not use the iv
+  /// of a surrounding loop of the associated loops nest.
+  /// We do not support non-perfectly nested loops not non-rectangular loops yet
+  /// (both introduced in OpenMP 5.0)
+  void CheckPerfectNestAndRectangularLoop(const parser::OpenMPLoopConstruct &x);
+
   // Predetermined DSA rules
   void PrivatizeAssociatedLoopIndexAndCheckLoopLevel(
       const parser::OpenMPLoopConstruct &);
@@ -1009,23 +1019,31 @@ bool DirectiveAttributeVisitor<T>::HasDataSharingAttributeObject(
 }
 
 template <typename T>
-const parser::Name *DirectiveAttributeVisitor<T>::GetLoopIndex(
-    const parser::DoConstruct &x) {
+std::tuple<const parser::Name *, const parser::ScalarExpr *,
+    const parser::ScalarExpr *, const parser::ScalarExpr *>
+DirectiveAttributeVisitor<T>::GetLoopBounds(const parser::DoConstruct &x) {
   using Bounds = parser::LoopControl::Bounds;
   if (x.GetLoopControl()) {
     if (const Bounds * b{std::get_if<Bounds>(&x.GetLoopControl()->u)}) {
-      return &b->name.thing;
-    } else {
-      return nullptr;
+      auto &&step = b->step;
+      return {&b->name.thing, &b->lower, &b->upper,
+          step.has_value() ? &step.value() : nullptr};
     }
   } else {
     context_
         .Say(std::get<parser::Statement<parser::NonLabelDoStmt>>(x.t).source,
             "Loop control is not present in the DO LOOP"_err_en_US)
         .Attach(GetContext().directiveSource,
             "associated with the enclosing LOOP construct"_en_US);
-    return nullptr;
   }
+  return {nullptr, nullptr, nullptr, nullptr};
+}
+
+template <typename T>
+const parser::Name *DirectiveAttributeVisitor<T>::GetLoopIndex(
+    const parser::DoConstruct &x) {
+  auto &&[iv, lb, ub, step] = GetLoopBounds(x);
+  return iv;
 }
 
 template <typename T>
@@ -1957,6 +1975,10 @@ bool OmpAttributeVisitor::Pre(const parser::OpenMPLoopConstruct &x) {
       }
     }
   }
+
+  // Must be done before iv privatization
+  CheckPerfectNestAndRectangularLoop(x);
+
   PrivatizeAssociatedLoopIndexAndCheckLoopLevel(x);
   ordCollapseLevel = GetNumAffectedLoopsFromLoopConstruct(x) + 1;
   return true;
@@ -2152,6 +2174,114 @@ void OmpAttributeVisitor::CollectNumAffectedLoopsFromClauses(
   }
 }
 
+void OmpAttributeVisitor::CheckPerfectNestAndRectangularLoop(
+    const parser::OpenMPLoopConstruct &x) {
+  auto &dirContext = GetContext();
+  std::int64_t dirDepth{dirContext.associatedLoopLevel};
+  if (dirDepth <= 0)
+    return;
+
+  auto checkExprHasSymbols = [&](llvm::SmallVector<Symbol *> &ivs,
+                                 const parser::ScalarExpr *bound) {
+    if (ivs.empty())
+      return;
+    auto boundExpr{semantics::AnalyzeExpr(context_, *bound)};
+    if (!boundExpr)
+      return;
+    semantics::UnorderedSymbolSet boundSyms =
+        evaluate::CollectSymbols(*boundExpr);
+    if (boundSyms.empty())
+      return;
+    for (Symbol *iv : ivs) {
+      if (boundSyms.count(*iv) != 0) {
+        // TODO: Point to occurence of iv in boundExpr, directiveSource as a
+        //       note
+        context_.Say(dirContext.directiveSource,
+            "Trip count must be computable and invariant"_err_en_US);
+      }
+    }
+  };
+
+  // Skip over loop transformation directives
+  const parser::OpenMPLoopConstruct *innerMostLoop = &x;
+  const parser::NestedConstruct *innerMostNest = nullptr;
+  while (auto &optLoopCons{
+      std::get<std::optional<parser::NestedConstruct>>(innerMostLoop->t)}) {
+    innerMostNest = &(optLoopCons.value());
+    if (const auto *innerLoop{
+            std::get_if<common::Indirection<parser::OpenMPLoopConstruct>>(
+                innerMostNest)}) {
+      innerMostLoop = &(innerLoop->value());
+    } else {
+      break;
+    }
+  }
+
+  if (!innerMostNest)
+    return;
+  const auto &outer{std::get_if<parser::DoConstruct>(innerMostNest)};
+  if (!outer)
+    return;
+
+  llvm::SmallVector<Symbol *> ivs;
+  int curLevel{0};
+  const parser::DoConstruct *loop{outer};
+  while (true) {
+    auto [iv, lb, ub, step] = GetLoopBounds(*loop);
+
+    if (lb)
+      checkExprHasSymbols(ivs, lb);
+    if (ub)
+      checkExprHasSymbols(ivs, ub);
+    if (step)
+      checkExprHasSymbols(ivs, step);
+    if (iv) {
+      if (auto *symbol{currScope().FindSymbol(iv->source)})
+        ivs.push_back(symbol);
+    }
+
+    // Stop after processing all affected loops
+    if (curLevel + 1 >= dirDepth)
+      break;
+
+    // Recurse into nested loop
+    const auto &block{std::get<parser::Block>(loop->t)};
+    if (block.empty()) {
+      // Insufficient number of nested loops already reported by
+      // CheckAssocLoopLevel()
+      break;
+    }
+
+    loop = GetDoConstructIf(block.front());
+    if (!loop) {
+      // Insufficient number of nested loops already reported by
+      // CheckAssocLoopLevel()
+      break;
+    }
+
+    auto checkPerfectNest = [&, this]() {
+      auto blockSize = block.size();
+      if (blockSize <= 1)
+        return;
+
+      if (parser::Unwrap<parser::ContinueStmt>(x))
+        blockSize -= 1;
+
+      if (blockSize <= 1)
+        return;
+
+      // Non-perfectly nested loop
+      // TODO: Point to non-DO statement, directiveSource as a note
+      context_.Say(dirContext.directiveSource,
+          "Canonical loop nest must be perfectly nested."_err_en_US);
+    };
+
+    checkPerfectNest();
+
+    ++curLevel;
+  }
+}
+
 // 2.15.1.1 Data-sharing Attribute Rules - Predetermined
 //   - The loop iteration variable(s) in the associated do-loop(s) of a do,
 //     parallel do, taskloop, or distribute construct is (are) private.
@@ -2258,10 +2388,18 @@ void OmpAttributeVisitor::PrivatizeAssociatedLoopIndexAndCheckLoopLevel(
 void OmpAttributeVisitor::CheckAssocLoopLevel(
     std::int64_t level, const parser::OmpClause *clause) {
   if (clause && level != 0) {
-    context_.Say(clause->source,
-        "The value of the parameter in the COLLAPSE or ORDERED clause must"
-        " not be larger than the number of nested loops"
-        " following the construct."_err_en_US);
+    switch (clause->Id()) {
+    case llvm::omp::OMPC_sizes:
+      context_.Say(clause->source,
+          "The SIZES clause has more entries than there are nested canonical loops."_err_en_US);
+      break;
+    default:
+      context_.Say(clause->source,
+          "The value of the parameter in the COLLAPSE or ORDERED clause must"
+          " not be larger than the number of nested loops"
+          " following the construct."_err_en_US);
+      break;
+    }
   }
 }