[flang-commits] [flang] [llvm] [mlir] [flang][OpenMP] Enable tiling (PR #143715)

Jan Leyonberg via flang-commits flang-commits at lists.llvm.org
Fri Jun 13 06:53:42 PDT 2025


https://github.com/jsjodin updated https://github.com/llvm/llvm-project/pull/143715

>From bd8edf7ad3561fbbf46a48dc2c91afaf7cafa30f Mon Sep 17 00:00:00 2001
From: Jan Leyonberg <jan_sjodin at yahoo.com>
Date: Fri, 14 Mar 2025 11:06:51 -0400
Subject: [PATCH 1/5] Initial implementation of tiling.

---
 flang/include/flang/Lower/OpenMP.h            |   1 -
 flang/include/flang/Parser/parse-tree.h       |   5 +-
 flang/lib/Lower/OpenMP/OpenMP.cpp             |  70 +++++--
 flang/lib/Lower/OpenMP/Utils.cpp              |  33 +++-
 flang/lib/Semantics/canonicalize-omp.cpp      |  46 +++--
 flang/lib/Semantics/resolve-directives.cpp    | 172 +++++++++++++++---
 .../Frontend/OpenMP/ConstructDecompositionT.h |  24 +++
 .../llvm/Frontend/OpenMP/OMPIRBuilder.h       |   9 +
 llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp     |  24 +++
 llvm/lib/Transforms/Utils/CodeExtractor.cpp   |   7 +-
 .../mlir/Dialect/OpenMP/OpenMPClauses.td      |  32 ++++
 mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td |   4 +-
 .../Conversion/SCFToOpenMP/SCFToOpenMP.cpp    |   3 +-
 mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp  |  52 +++++-
 .../OpenMP/OpenMPToLLVMIRTranslation.cpp      |  65 +++++--
 15 files changed, 461 insertions(+), 86 deletions(-)

diff --git a/flang/include/flang/Lower/OpenMP.h b/flang/include/flang/Lower/OpenMP.h
index 6e150ef4e8e82..4d77fff912f4f 100644
--- a/flang/include/flang/Lower/OpenMP.h
+++ b/flang/include/flang/Lower/OpenMP.h
@@ -79,7 +79,6 @@ void genOpenMPDeclarativeConstruct(AbstractConverter &,
 void genOpenMPSymbolProperties(AbstractConverter &converter,
                                const pft::Variable &var);
 
-int64_t getCollapseValue(const Fortran::parser::OmpClauseList &clauseList);
 void genThreadprivateOp(AbstractConverter &, const pft::Variable &);
 void genDeclareTargetIntGlobal(AbstractConverter &, const pft::Variable &);
 bool isOpenMPTargetConstruct(const parser::OpenMPConstruct &);
diff --git a/flang/include/flang/Parser/parse-tree.h b/flang/include/flang/Parser/parse-tree.h
index c99006f0c1c22..4ae1c354d504a 100644
--- a/flang/include/flang/Parser/parse-tree.h
+++ b/flang/include/flang/Parser/parse-tree.h
@@ -5075,9 +5075,10 @@ struct OpenMPBlockConstruct {
 struct OpenMPLoopConstruct {
   TUPLE_CLASS_BOILERPLATE(OpenMPLoopConstruct);
   OpenMPLoopConstruct(OmpBeginLoopDirective &&a)
-      : t({std::move(a), std::nullopt, std::nullopt}) {}
+    : t({std::move(a), std::nullopt, std::nullopt, std::nullopt}) {}
   std::tuple<OmpBeginLoopDirective, std::optional<DoConstruct>,
-      std::optional<OmpEndLoopDirective>>
+             std::optional<common::Indirection<OpenMPLoopConstruct>>,
+             std::optional<OmpEndLoopDirective>>
       t;
 };
 
diff --git a/flang/lib/Lower/OpenMP/OpenMP.cpp b/flang/lib/Lower/OpenMP/OpenMP.cpp
index 6892e571e62a3..d6f6ba70c9401 100644
--- a/flang/lib/Lower/OpenMP/OpenMP.cpp
+++ b/flang/lib/Lower/OpenMP/OpenMP.cpp
@@ -451,6 +451,7 @@ static void processHostEvalClauses(lower::AbstractConverter &converter,
       return;
 
     const parser::OmpClauseList *beginClauseList = nullptr;
+    const parser::OmpClauseList *middleClauseList = nullptr;
     const parser::OmpClauseList *endClauseList = nullptr;
     common::visit(
         common::visitors{
@@ -468,6 +469,22 @@ static void processHostEvalClauses(lower::AbstractConverter &converter,
               beginClauseList =
                   &std::get<parser::OmpClauseList>(beginDirective.t);
 
+              // FIXME(JAN): For now we check if there is an inner
+              // OpenMPLoopConstruct, and extract the size clause from there
+              const auto &innerOptional = std::get<std::optional<
+                  common::Indirection<parser::OpenMPLoopConstruct>>>(
+                  ompConstruct.t);
+              if (innerOptional.has_value()) {
+                const auto &innerLoopDirective = innerOptional.value().value();
+                const auto &innerBegin =
+                  std::get<parser::OmpBeginLoopDirective>(innerLoopDirective.t);
+                const auto &innerDirective =
+                  std::get<parser::OmpLoopDirective>(innerBegin.t);
+                if (innerDirective.v == llvm::omp::Directive::OMPD_tile) {
+                  middleClauseList =
+                      &std::get<parser::OmpClauseList>(innerBegin.t);
+                }
+              }
               if (auto &endDirective =
                       std::get<std::optional<parser::OmpEndLoopDirective>>(
                           ompConstruct.t))
@@ -480,6 +497,9 @@ static void processHostEvalClauses(lower::AbstractConverter &converter,
     assert(beginClauseList && "expected begin directive");
     clauses.append(makeClauses(*beginClauseList, semaCtx));
 
+    if (middleClauseList)
+      clauses.append(makeClauses(*middleClauseList, semaCtx));
+
     if (endClauseList)
       clauses.append(makeClauses(*endClauseList, semaCtx));
   };
@@ -955,6 +975,7 @@ static void genLoopVars(
     storeOp =
         createAndSetPrivatizedLoopVar(converter, loc, indexVal, argSymbol);
   }
+
   firOpBuilder.setInsertionPointAfter(storeOp);
 }
 
@@ -1697,6 +1718,23 @@ genLoopNestClauses(lower::AbstractConverter &converter,
     cp.processCollapse(loc, eval, clauseOps, iv);
 
   clauseOps.loopInclusive = converter.getFirOpBuilder().getUnitAttr();
+
+  fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
+  for (auto &clause : clauses) {
+    if (clause.id == llvm::omp::Clause::OMPC_collapse) {
+      const auto &collapse = std::get<clause::Collapse>(clause.u);
+      int64_t collapseValue = evaluate::ToInt64(collapse.v).value();
+      clauseOps.numCollapse = firOpBuilder.getI64IntegerAttr(collapseValue);
+    } else if (clause.id == llvm::omp::Clause::OMPC_sizes) {
+      const auto &sizes = std::get<clause::Sizes>(clause.u);
+      llvm::SmallVector<int64_t> sizeValues;
+      for (auto &size : sizes.v) {
+        int64_t sizeValue = evaluate::ToInt64(size).value();
+        sizeValues.push_back(sizeValue);
+      }
+      clauseOps.tileSizes = sizeValues;
+    }
+  }
 }
 
 static void genLoopClauses(
@@ -2069,9 +2107,9 @@ static mlir::omp::LoopNestOp genLoopNestOp(
     return llvm::SmallVector<const semantics::Symbol *>(iv);
   };
 
-  auto *nestedEval =
-      getCollapsedLoopEval(eval, getCollapseValue(item->clauses));
-
+  uint64_t nestValue = getCollapseValue(item->clauses);
+  nestValue = nestValue < iv.size() ? iv.size() : nestValue;
+  auto *nestedEval = getCollapsedLoopEval(eval, nestValue);
   return genOpWithBody<mlir::omp::LoopNestOp>(
       OpWithBodyGenInfo(converter, symTable, semaCtx, loc, *nestedEval,
                         directive)
@@ -4385,6 +4423,20 @@ static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable,
       std::get<parser::OmpBeginLoopDirective>(loopConstruct.t);
   List<Clause> clauses = makeClauses(
       std::get<parser::OmpClauseList>(beginLoopDirective.t), semaCtx);
+
+  const auto &innerOptional = std::get<std::optional<common::Indirection<parser::OpenMPLoopConstruct>>>(loopConstruct.t);
+  if (innerOptional.has_value()) {
+    const auto &innerLoopDirective = innerOptional.value().value();
+    const auto &innerBegin =
+      std::get<parser::OmpBeginLoopDirective>(innerLoopDirective.t);
+    const auto &innerDirective =
+      std::get<parser::OmpLoopDirective>(innerBegin.t);
+    if (innerDirective.v == llvm::omp::Directive::OMPD_tile) {
+      clauses.append(
+          makeClauses(std::get<parser::OmpClauseList>(innerBegin.t), semaCtx));
+    }
+  }
+
   if (auto &endLoopDirective =
           std::get<std::optional<parser::OmpEndLoopDirective>>(
               loopConstruct.t)) {
@@ -4505,18 +4557,6 @@ void Fortran::lower::genOpenMPSymbolProperties(
     lower::genDeclareTargetIntGlobal(converter, var);
 }
 
-int64_t
-Fortran::lower::getCollapseValue(const parser::OmpClauseList &clauseList) {
-  for (const parser::OmpClause &clause : clauseList.v) {
-    if (const auto &collapseClause =
-            std::get_if<parser::OmpClause::Collapse>(&clause.u)) {
-      const auto *expr = semantics::GetExpr(collapseClause->v);
-      return evaluate::ToInt64(*expr).value();
-    }
-  }
-  return 1;
-}
-
 void Fortran::lower::genThreadprivateOp(lower::AbstractConverter &converter,
                                         const lower::pft::Variable &var) {
   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
diff --git a/flang/lib/Lower/OpenMP/Utils.cpp b/flang/lib/Lower/OpenMP/Utils.cpp
index c226c2558e7aa..453dec9fbf3d7 100644
--- a/flang/lib/Lower/OpenMP/Utils.cpp
+++ b/flang/lib/Lower/OpenMP/Utils.cpp
@@ -38,14 +38,22 @@ namespace lower {
 namespace omp {
 
 int64_t getCollapseValue(const List<Clause> &clauses) {
-  auto iter = llvm::find_if(clauses, [](const Clause &clause) {
-    return clause.id == llvm::omp::Clause::OMPC_collapse;
-  });
-  if (iter != clauses.end()) {
-    const auto &collapse = std::get<clause::Collapse>(iter->u);
-    return evaluate::ToInt64(collapse.v).value();
+  int64_t collapseValue = 1;
+  int64_t numTileSizes = 0;
+  for (auto &clause : clauses) {
+    if (clause.id == llvm::omp::Clause::OMPC_collapse) {
+      const auto &collapse = std::get<clause::Collapse>(clause.u);
+      collapseValue = evaluate::ToInt64(collapse.v).value();
+    } else if (clause.id == llvm::omp::Clause::OMPC_sizes) {
+      const auto &sizes = std::get<clause::Sizes>(clause.u);
+      numTileSizes = sizes.v.size();
+    }
   }
-  return 1;
+
+  collapseValue = collapseValue - numTileSizes;
+  int64_t result =
+    collapseValue > numTileSizes ? collapseValue : numTileSizes;
+  return result;
 }
 
 void genObjectList(const ObjectList &objects,
@@ -611,6 +619,7 @@ bool collectLoopRelatedInfo(
     lower::pft::Evaluation &eval, const omp::List<omp::Clause> &clauses,
     mlir::omp::LoopRelatedClauseOps &result,
     llvm::SmallVectorImpl<const semantics::Symbol *> &iv) {
+
   bool found = false;
   fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
 
@@ -626,7 +635,16 @@ bool collectLoopRelatedInfo(
     collapseValue = evaluate::ToInt64(clause->v).value();
     found = true;
   }
+  std::int64_t sizesLengthValue = 0l;
+  if (auto *clause =
+          ClauseFinder::findUniqueClause<omp::clause::Sizes>(clauses)) {
+    sizesLengthValue = clause->v.size();
+    found = true;
+  }
 
+  collapseValue = collapseValue - sizesLengthValue;
+  collapseValue =
+      collapseValue < sizesLengthValue ? sizesLengthValue : collapseValue;
   std::size_t loopVarTypeSize = 0;
   do {
     lower::pft::Evaluation *doLoop =
@@ -659,7 +677,6 @@ bool collectLoopRelatedInfo(
   } while (collapseValue > 0);
 
   convertLoopBounds(converter, currentLocation, result, loopVarTypeSize);
-
   return found;
 }
 } // namespace omp
diff --git a/flang/lib/Semantics/canonicalize-omp.cpp b/flang/lib/Semantics/canonicalize-omp.cpp
index 5164f1dc6faab..44e8045167798 100644
--- a/flang/lib/Semantics/canonicalize-omp.cpp
+++ b/flang/lib/Semantics/canonicalize-omp.cpp
@@ -8,7 +8,7 @@
 
 #include "canonicalize-omp.h"
 #include "flang/Parser/parse-tree-visitor.h"
-
+# include <stack>
 // After Loop Canonicalization, rewrite OpenMP parse tree to make OpenMP
 // Constructs more structured which provide explicit scopes for later
 // structural checks and semantic analysis.
@@ -112,15 +112,17 @@ class CanonicalizationOfOmp {
     // in the same iteration
     //
     // Original:
-    //   ExecutableConstruct -> OpenMPConstruct -> OpenMPLoopConstruct
-    //     OmpBeginLoopDirective
+    //   ExecutableConstruct -> OpenMPConstruct -> OpenMPLoopConstruct t->
+    //     OmpBeginLoopDirective t-> OmpLoopDirective
+    //   [ExecutableConstruct -> OpenMPConstruct -> OpenMPLoopConstruct u->
+    ///    OmpBeginLoopDirective t-> OmpLoopDirective t-> Tile v-> OMP_tile]
     //   ExecutableConstruct -> DoConstruct
+    //   [ExecutableConstruct -> OmpEndLoopDirective]
     //   ExecutableConstruct -> OmpEndLoopDirective (if available)
     //
     // After rewriting:
-    //   ExecutableConstruct -> OpenMPConstruct -> OpenMPLoopConstruct
-    //     OmpBeginLoopDirective
-    //     DoConstruct
+    //   ExecutableConstruct -> OpenMPConstruct -> OpenMPLoopConstruct t->
+    //     OmpBeginLoopDirective t -> OmpLoopDirective -> DoConstruct
     //     OmpEndLoopDirective (if available)
     parser::Block::iterator nextIt;
     auto &beginDir{std::get<parser::OmpBeginLoopDirective>(x.t)};
@@ -132,19 +134,39 @@ class CanonicalizationOfOmp {
       if (GetConstructIf<parser::CompilerDirective>(*nextIt))
         continue;
 
+      // Keep track of the loops to handle the end loop directives
+      std::stack<parser::OpenMPLoopConstruct *> loops;
+      loops.push(&x);
+      while (auto *innerConstruct{
+          GetConstructIf<parser::OpenMPConstruct>(*nextIt)}) {
+        if (auto *innerOmpLoop{
+                std::get_if<parser::OpenMPLoopConstruct>(&innerConstruct->u)}) {
+          std::get<
+              std::optional<common::Indirection<parser::OpenMPLoopConstruct>>>(
+              loops.top()->t) = std::move(*innerOmpLoop);
+          // Retrieveing the address so that DoConstruct or inner loop can be
+          // set later.
+          loops.push(&(std::get<std::optional<
+                           common::Indirection<parser::OpenMPLoopConstruct>>>(
+              loops.top()->t)
+                           .value()
+                           .value()));
+          nextIt = block.erase(nextIt);
+        }
+      }
       if (auto *doCons{GetConstructIf<parser::DoConstruct>(*nextIt)}) {
         if (doCons->GetLoopControl()) {
-          // move DoConstruct
-          std::get<std::optional<parser::DoConstruct>>(x.t) =
+          std::get<std::optional<parser::DoConstruct>>(loops.top()->t) =
               std::move(*doCons);
           nextIt = block.erase(nextIt);
           // try to match OmpEndLoopDirective
-          if (nextIt != block.end()) {
+          while (nextIt != block.end() && !loops.empty()) {
             if (auto *endDir{
                     GetConstructIf<parser::OmpEndLoopDirective>(*nextIt)}) {
-              std::get<std::optional<parser::OmpEndLoopDirective>>(x.t) =
-                  std::move(*endDir);
-              block.erase(nextIt);
+              std::get<std::optional<parser::OmpEndLoopDirective>>(
+                  loops.top()->t) = std::move(*endDir);
+              nextIt = block.erase(nextIt);
+              loops.pop();
             }
           }
         } else {
diff --git a/flang/lib/Semantics/resolve-directives.cpp b/flang/lib/Semantics/resolve-directives.cpp
index ec385b7baf8bd..2cfa6a2867e64 100644
--- a/flang/lib/Semantics/resolve-directives.cpp
+++ b/flang/lib/Semantics/resolve-directives.cpp
@@ -743,7 +743,28 @@ class OmpAttributeVisitor : DirectiveAttributeVisitor<llvm::omp::Directive> {
   const parser::OmpClause *GetAssociatedClause() { return associatedClause; }
 
 private:
+  std::int64_t SetAssociatedMaxClause(llvm::SmallVector<std::int64_t> &,
+      llvm::SmallVector<const parser::OmpClause *> &);
+  std::int64_t GetAssociatedLoopLevelFromLoopConstruct(
+      const parser::OpenMPLoopConstruct &);
   std::int64_t GetAssociatedLoopLevelFromClauses(const parser::OmpClauseList &);
+  void CollectAssociatedLoopLevelsFromLoopConstruct(
+      const parser::OpenMPLoopConstruct &, llvm::SmallVector<std::int64_t> &,
+      llvm::SmallVector<const parser::OmpClause *> &);
+  void CollectAssociatedLoopLevelsFromInnerLoopContruct(
+      const parser::OpenMPLoopConstruct &, llvm::SmallVector<std::int64_t> &,
+      llvm::SmallVector<const parser::OmpClause *> &);
+  template <typename T>
+  void CollectAssociatedLoopLevelFromClauseValue(
+      const parser::OmpClause &clause, llvm::SmallVector<std::int64_t> &,
+      llvm::SmallVector<const parser::OmpClause *> &);
+  template <typename T>
+  void CollectAssociatedLoopLevelFromClauseSize(const parser::OmpClause &,
+      llvm::SmallVector<std::int64_t> &,
+      llvm::SmallVector<const parser::OmpClause *> &);
+  void CollectAssociatedLoopLevelsFromClauses(const parser::OmpClauseList &,
+      llvm::SmallVector<std::int64_t> &,
+      llvm::SmallVector<const parser::OmpClause *> &);
 
   Symbol::Flags dataSharingAttributeFlags{Symbol::Flag::OmpShared,
       Symbol::Flag::OmpPrivate, Symbol::Flag::OmpFirstPrivate,
@@ -1740,7 +1761,6 @@ bool OmpAttributeVisitor::Pre(
 bool OmpAttributeVisitor::Pre(const parser::OpenMPLoopConstruct &x) {
   const auto &beginLoopDir{std::get<parser::OmpBeginLoopDirective>(x.t)};
   const auto &beginDir{std::get<parser::OmpLoopDirective>(beginLoopDir.t)};
-  const auto &clauseList{std::get<parser::OmpClauseList>(beginLoopDir.t)};
   switch (beginDir.v) {
   case llvm::omp::Directive::OMPD_distribute:
   case llvm::omp::Directive::OMPD_distribute_parallel_do:
@@ -1791,7 +1811,7 @@ bool OmpAttributeVisitor::Pre(const parser::OpenMPLoopConstruct &x) {
       beginDir.v == llvm::omp::Directive::OMPD_target_loop)
     IssueNonConformanceWarning(beginDir.v, beginDir.source);
   ClearDataSharingAttributeObjects();
-  SetContextAssociatedLoopLevel(GetAssociatedLoopLevelFromClauses(clauseList));
+  SetContextAssociatedLoopLevel(GetAssociatedLoopLevelFromLoopConstruct(x));
 
   if (beginDir.v == llvm::omp::Directive::OMPD_do) {
     if (const auto &doConstruct{
@@ -1802,7 +1822,7 @@ bool OmpAttributeVisitor::Pre(const parser::OpenMPLoopConstruct &x) {
     }
   }
   PrivatizeAssociatedLoopIndexAndCheckLoopLevel(x);
-  ordCollapseLevel = GetAssociatedLoopLevelFromClauses(clauseList) + 1;
+  ordCollapseLevel = GetAssociatedLoopLevelFromLoopConstruct(x) + 1;
   return true;
 }
 
@@ -1890,39 +1910,124 @@ bool OmpAttributeVisitor::Pre(const parser::DoConstruct &x) {
   return true;
 }
 
+static bool isSizesClause(const parser::OmpClause *clause) {
+  return std::holds_alternative<parser::OmpClause::Sizes>(clause->u);
+}
+
+std::int64_t OmpAttributeVisitor::SetAssociatedMaxClause(
+    llvm::SmallVector<std::int64_t> &levels,
+    llvm::SmallVector<const parser::OmpClause *> &clauses) {
+
+  // Find the tile level to know how much to reduce the level for collapse
+  std::int64_t tileLevel = 0;
+  for (auto [level, clause] : llvm::zip_equal(levels, clauses)) {
+    if (isSizesClause(clause)) {
+      tileLevel = level;
+    }
+  }
+
+  std::int64_t maxLevel = 1;
+  const parser::OmpClause *maxClause = nullptr;
+  for (auto [level, clause] : llvm::zip_equal(levels, clauses)) {
+    if (tileLevel > 0 && tileLevel < level) {
+      context_.Say(clause->source,
+          "The value of the parameter in the COLLAPSE clause must"
+          " not be larger than the number of the number of tiled loops"
+          " because collapse relies on independent loop iterations."_err_en_US);
+      return 1;
+    }
+
+    if (!isSizesClause(clause)) {
+      level = level - tileLevel;
+    }
+
+    if (level > maxLevel) {
+      maxLevel = level;
+      maxClause = clause;
+    }
+  }
+  if (maxClause)
+    SetAssociatedClause(*maxClause);
+  return maxLevel;
+}
+
+std::int64_t OmpAttributeVisitor::GetAssociatedLoopLevelFromLoopConstruct(
+    const parser::OpenMPLoopConstruct &x) {
+  llvm::SmallVector<std::int64_t> levels;
+  llvm::SmallVector<const parser::OmpClause *> clauses;
+
+  CollectAssociatedLoopLevelsFromLoopConstruct(x, levels, clauses);
+  return SetAssociatedMaxClause(levels, clauses);
+}
+
 std::int64_t OmpAttributeVisitor::GetAssociatedLoopLevelFromClauses(
     const parser::OmpClauseList &x) {
-  std::int64_t orderedLevel{0};
-  std::int64_t collapseLevel{0};
+  llvm::SmallVector<std::int64_t> levels;
+  llvm::SmallVector<const parser::OmpClause *> clauses;
 
-  const parser::OmpClause *ordClause{nullptr};
-  const parser::OmpClause *collClause{nullptr};
+  CollectAssociatedLoopLevelsFromClauses(x, levels, clauses);
+  return SetAssociatedMaxClause(levels, clauses);
+}
 
-  for (const auto &clause : x.v) {
-    if (const auto *orderedClause{
-            std::get_if<parser::OmpClause::Ordered>(&clause.u)}) {
-      if (const auto v{EvaluateInt64(context_, orderedClause->v)}) {
-        orderedLevel = *v;
-      }
-      ordClause = &clause;
-    }
-    if (const auto *collapseClause{
-            std::get_if<parser::OmpClause::Collapse>(&clause.u)}) {
-      if (const auto v{EvaluateInt64(context_, collapseClause->v)}) {
-        collapseLevel = *v;
-      }
-      collClause = &clause;
+void OmpAttributeVisitor::CollectAssociatedLoopLevelsFromLoopConstruct(
+    const parser::OpenMPLoopConstruct &x,
+    llvm::SmallVector<std::int64_t> &levels,
+    llvm::SmallVector<const parser::OmpClause *> &clauses) {
+  const auto &beginLoopDir{std::get<parser::OmpBeginLoopDirective>(x.t)};
+  const auto &clauseList{std::get<parser::OmpClauseList>(beginLoopDir.t)};
+
+  CollectAssociatedLoopLevelsFromClauses(clauseList, levels, clauses);
+  CollectAssociatedLoopLevelsFromInnerLoopContruct(x, levels, clauses);
+}
+
+void OmpAttributeVisitor::CollectAssociatedLoopLevelsFromInnerLoopContruct(
+    const parser::OpenMPLoopConstruct &x,
+    llvm::SmallVector<std::int64_t> &levels,
+    llvm::SmallVector<const parser::OmpClause *> &clauses) {
+  const auto &innerOptional =
+      std::get<std::optional<common::Indirection<parser::OpenMPLoopConstruct>>>(
+          x.t);
+  if (innerOptional.has_value()) {
+    CollectAssociatedLoopLevelsFromLoopConstruct(
+        innerOptional.value().value(), levels, clauses);
+  }
+}
+
+template <typename T>
+void OmpAttributeVisitor::CollectAssociatedLoopLevelFromClauseValue(
+    const parser::OmpClause &clause, llvm::SmallVector<std::int64_t> &levels,
+    llvm::SmallVector<const parser::OmpClause *> &clauses) {
+  if (const auto tclause{std::get_if<T>(&clause.u)}) {
+    std::int64_t level = 0;
+    if (const auto v{EvaluateInt64(context_, tclause->v)}) {
+      level = *v;
     }
+    levels.push_back(level);
+    clauses.push_back(&clause);
   }
+}
 
-  if (orderedLevel && (!collapseLevel || orderedLevel >= collapseLevel)) {
-    SetAssociatedClause(*ordClause);
-    return orderedLevel;
-  } else if (!orderedLevel && collapseLevel) {
-    SetAssociatedClause(*collClause);
-    return collapseLevel;
-  } // orderedLevel < collapseLevel is an error handled in structural checks
-  return 1; // default is outermost loop
+template <typename T>
+void OmpAttributeVisitor::CollectAssociatedLoopLevelFromClauseSize(
+    const parser::OmpClause &clause, llvm::SmallVector<std::int64_t> &levels,
+    llvm::SmallVector<const parser::OmpClause *> &clauses) {
+  if (const auto tclause{std::get_if<T>(&clause.u)}) {
+    levels.push_back(tclause->v.size());
+    clauses.push_back(&clause);
+  }
+}
+
+void OmpAttributeVisitor::CollectAssociatedLoopLevelsFromClauses(
+    const parser::OmpClauseList &x, llvm::SmallVector<std::int64_t> &levels,
+    llvm::SmallVector<const parser::OmpClause *> &clauses) {
+  for (const auto &clause : x.v) {
+    CollectAssociatedLoopLevelFromClauseValue<parser::OmpClause::Ordered>(
+        clause, levels, clauses);
+    CollectAssociatedLoopLevelFromClauseValue<parser::OmpClause::Collapse>(
+        clause, levels, clauses);
+    CollectAssociatedLoopLevelFromClauseSize<parser::OmpClause::Sizes>(
+        clause, levels, clauses);
+  }
 }
 
 // 2.15.1.1 Data-sharing Attribute Rules - Predetermined
@@ -1949,7 +2054,14 @@ void OmpAttributeVisitor::PrivatizeAssociatedLoopIndexAndCheckLoopLevel(
     ivDSA = Symbol::Flag::OmpLastPrivate;
   }
 
-  const auto &outer{std::get<std::optional<parser::DoConstruct>>(x.t)};
+  const parser::OpenMPLoopConstruct *innerMostLoop = &x;
+  while (auto &innerLoop{
+      std::get<std::optional<common::Indirection<parser::OpenMPLoopConstruct>>>(
+          innerMostLoop->t)}) {
+    innerMostLoop = &innerLoop.value().value();
+  }
+
+  const auto &outer{std::get<std::optional<parser::DoConstruct>>(innerMostLoop->t)};
   if (outer.has_value()) {
     for (const parser::DoConstruct *loop{&*outer}; loop && level > 0; --level) {
       // go through all the nested do-loops and resolve index variables
diff --git a/llvm/include/llvm/Frontend/OpenMP/ConstructDecompositionT.h b/llvm/include/llvm/Frontend/OpenMP/ConstructDecompositionT.h
index cdc80c88b7425..dea8cf0bada10 100644
--- a/llvm/include/llvm/Frontend/OpenMP/ConstructDecompositionT.h
+++ b/llvm/include/llvm/Frontend/OpenMP/ConstructDecompositionT.h
@@ -209,6 +209,8 @@ struct ConstructDecompositionT {
 
   bool applyClause(const tomp::clause::CollapseT<TypeTy, IdTy, ExprTy> &clause,
                    const ClauseTy *);
+  bool applyClause(const tomp::clause::SizesT<TypeTy, IdTy, ExprTy> &clause,
+                   const ClauseTy *);
   bool applyClause(const tomp::clause::PrivateT<TypeTy, IdTy, ExprTy> &clause,
                    const ClauseTy *);
   bool
@@ -482,6 +484,28 @@ bool ConstructDecompositionT<C, H>::applyClause(
 
   return false;
 }
+// FIXME(JAN): Do the correct thing, but for now we'll do the same as collapse
+template <typename C, typename H>
+bool ConstructDecompositionT<C, H>::applyClause(
+    const tomp::clause::SizesT<TypeTy, IdTy, ExprTy> &clause,
+    const ClauseTy *node) {
+  // Apply "sizes" to the innermost directive. If it's not one that
+  // allows it flag an error.
+  if (!leafs.empty()) {
+    auto &last = leafs.back();
+
+    if (llvm::omp::isAllowedClauseForDirective(last.id, node->id, version)) {
+      last.clauses.push_back(node);
+      return true;
+    } else {
+      llvm::errs() << "** OVERRIDING isAllowedClauseForDirective **\n";
+      last.clauses.push_back(node);
+      return true;
+    }
+  }
+
+  return false;
+}
 
 // PRIVATE
 // [5.2:111:5-7]
diff --git a/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h b/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
index e4b1241151e9d..1aa70f04d0fc6 100644
--- a/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
+++ b/llvm/include/llvm/Frontend/OpenMP/OMPIRBuilder.h
@@ -2162,6 +2162,9 @@ class OpenMPIRBuilder {
 
     /// Return the function that contains the region to be outlined.
     Function *getFunction() const { return EntryBB->getParent(); }
+
+    /// Dump the info in a somewhat readable way
+    void dump();
   };
 
   /// Collection of regions that need to be outlined during finalization.
@@ -2179,6 +2182,9 @@ class OpenMPIRBuilder {
   /// Add a new region that will be outlined later.
   void addOutlineInfo(OutlineInfo &&OI) { OutlineInfos.emplace_back(OI); }
 
+  /// Dump outline infos
+  void dumpOutlineInfos();
+
   /// An ordered map of auto-generated variables to their unique names.
   /// It stores variables with the following names: 1) ".gomp_critical_user_" +
   /// <critical_section_name> + ".var" for "omp critical" directives; 2)
@@ -3762,6 +3768,9 @@ class CanonicalLoopInfo {
   /// Invalidate this loop. That is, the underlying IR does not fulfill the
   /// requirements of an OpenMP canonical loop anymore.
   LLVM_ABI void invalidate();
+
+  /// Dump the info in a somewhat readable way
+  void dump();
 };
 
 } // end namespace llvm
diff --git a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
index ca3d8438654dc..da6adcada6236 100644
--- a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
+++ b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
@@ -8530,6 +8530,15 @@ Error OpenMPIRBuilder::emitOffloadingArrays(
   return Error::success();
 }
 
+void OpenMPIRBuilder::dumpOutlineInfos() {
+  errs() << "=== Outline Infos Begin ===\n";
+  for (auto En : enumerate(OutlineInfos)) {
+    errs() << "[" << En.index() << "]: ";
+    En.value().dump();
+  }
+  errs() << "=== Outline Infos End ===\n";
+}
+
 void OpenMPIRBuilder::emitBranch(BasicBlock *Target) {
   BasicBlock *CurBB = Builder.GetInsertBlock();
 
@@ -9438,6 +9447,14 @@ void OpenMPIRBuilder::OutlineInfo::collectBlocks(
   }
 }
 
+void OpenMPIRBuilder::OutlineInfo::dump() {
+  errs() << "=== OutilneInfo == "
+         << " EntryBB: " << (EntryBB ? EntryBB->getName() : "n\a")
+         << " ExitBB: " << (ExitBB ? ExitBB->getName() : "n\a")
+         << " OuterAllocaBB: "
+         << (OuterAllocaBB ? OuterAllocaBB->getName() : "n/a") << "\n";
+}
+
 void OpenMPIRBuilder::createOffloadEntry(Constant *ID, Constant *Addr,
                                          uint64_t Size, int32_t Flags,
                                          GlobalValue::LinkageTypes,
@@ -10215,3 +10232,10 @@ void CanonicalLoopInfo::invalidate() {
   Latch = nullptr;
   Exit = nullptr;
 }
+
+void CanonicalLoopInfo::dump() {
+  errs() << "CanonicaLoop == Header: " << (Header ? Header->getName() : "n/a")
+         << " Cond: " << (Cond ? Cond->getName() : "n/a")
+         << " Latch: " << (Latch ? Latch->getName() : "n/a")
+         << " Exit: " << (Exit ? Exit->getName() : "n/a") << "\n";
+}
diff --git a/llvm/lib/Transforms/Utils/CodeExtractor.cpp b/llvm/lib/Transforms/Utils/CodeExtractor.cpp
index c4894c90c127f..0a1c6ce3972e3 100644
--- a/llvm/lib/Transforms/Utils/CodeExtractor.cpp
+++ b/llvm/lib/Transforms/Utils/CodeExtractor.cpp
@@ -815,7 +815,11 @@ void CodeExtractor::severSplitPHINodesOfExits() {
 }
 
 void CodeExtractor::splitReturnBlocks() {
-  for (BasicBlock *Block : Blocks)
+  for (BasicBlock *Block : Blocks) {
+    if (!Block->getTerminator()) {
+      errs() << "====== No terminator in block: " << Block->getName()
+             << "======\n";
+    }
     if (ReturnInst *RI = dyn_cast<ReturnInst>(Block->getTerminator())) {
       BasicBlock *New =
           Block->splitBasicBlock(RI->getIterator(), Block->getName() + ".ret");
@@ -832,6 +836,7 @@ void CodeExtractor::splitReturnBlocks() {
           DT->changeImmediateDominator(I, NewNode);
       }
     }
+  }
 }
 
 Function *CodeExtractor::constructFunctionDeclaration(
diff --git a/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td b/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
index 16c14ef085d6d..09c77beb857d4 100644
--- a/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
+++ b/mlir/include/mlir/Dialect/OpenMP/OpenMPClauses.td
@@ -266,6 +266,38 @@ class OpenMP_DeviceClauseSkip<
 
 def OpenMP_DeviceClause : OpenMP_DeviceClauseSkip<>;
 
+//===----------------------------------------------------------------------===//
+// V5.2: [XX.X] `collapse` clause
+//===----------------------------------------------------------------------===//
+
+class OpenMP_CollapseClauseSkip<
+    bit traits = false, bit arguments = false, bit assemblyFormat = false,
+    bit description = false, bit extraClassDeclaration = false
+  > : OpenMP_Clause<traits, arguments, assemblyFormat, description,
+                    extraClassDeclaration> {
+  let arguments = (ins
+      DefaultValuedOptionalAttr<I64Attr, "1">:$num_collapse
+  );
+}
+
+def OpenMP_CollapseClause : OpenMP_CollapseClauseSkip<>;
+
+//===----------------------------------------------------------------------===//
+// V5.2: [xx.x] `sizes` clause
+//===----------------------------------------------------------------------===//
+
+class OpenMP_TileSizesClauseSkip<
+    bit traits = false, bit arguments = false, bit assemblyFormat = false,
+    bit description = false, bit extraClassDeclaration = false
+  > : OpenMP_Clause<traits, arguments, assemblyFormat, description,
+                    extraClassDeclaration> {
+  let arguments = (ins
+      OptionalAttr<DenseI64ArrayAttr>:$tile_sizes
+  );
+}
+
+def OpenMP_TileSizesClause : OpenMP_TileSizesClauseSkip<>;
+
 //===----------------------------------------------------------------------===//
 // V5.2: [11.6.1] `dist_schedule` clause
 //===----------------------------------------------------------------------===//
diff --git a/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td b/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td
index 036c6a6e350a8..7867209c872f2 100644
--- a/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td
+++ b/mlir/include/mlir/Dialect/OpenMP/OpenMPOps.td
@@ -407,7 +407,9 @@ def WorkshareLoopWrapperOp : OpenMP_Op<"workshare.loop_wrapper", traits = [
 def LoopNestOp : OpenMP_Op<"loop_nest", traits = [
     RecursiveMemoryEffects, SameVariadicOperandSize
   ], clauses = [
-    OpenMP_LoopRelatedClause
+    OpenMP_LoopRelatedClause,
+    OpenMP_CollapseClause,
+    OpenMP_TileSizesClause
   ], singleRegion = true> {
   let summary = "rectangular loop nest";
   let description = [{
diff --git a/mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp b/mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp
index 71786e856c6db..d60578dc9a539 100644
--- a/mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp
+++ b/mlir/lib/Conversion/SCFToOpenMP/SCFToOpenMP.cpp
@@ -493,7 +493,8 @@ struct ParallelOpLowering : public OpRewritePattern<scf::ParallelOp> {
         // Create loop nest and populate region with contents of scf.parallel.
         auto loopOp = rewriter.create<omp::LoopNestOp>(
             parallelOp.getLoc(), parallelOp.getLowerBound(),
-            parallelOp.getUpperBound(), parallelOp.getStep());
+            parallelOp.getUpperBound(), parallelOp.getStep(), false, 1,
+            nullptr);
 
         rewriter.inlineRegionBefore(parallelOp.getRegion(), loopOp.getRegion(),
                                     loopOp.getRegion().begin());
diff --git a/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp b/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
index e94d570b57122..53e3bf9793569 100644
--- a/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
+++ b/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
@@ -55,6 +55,11 @@ makeDenseBoolArrayAttr(MLIRContext *ctx, const ArrayRef<bool> boolArray) {
   return boolArray.empty() ? nullptr : DenseBoolArrayAttr::get(ctx, boolArray);
 }
 
+static DenseI64ArrayAttr
+makeDenseI64ArrayAttr(MLIRContext *ctx, const ArrayRef<int64_t> intArray) {
+  return intArray.empty() ? nullptr : DenseI64ArrayAttr::get(ctx, intArray);
+}
+
 namespace {
 struct MemRefPointerLikeModel
     : public PointerLikeType::ExternalModel<MemRefPointerLikeModel,
@@ -2939,10 +2944,10 @@ ParseResult LoopNestOp::parse(OpAsmParser &parser, OperationState &result) {
   for (auto &iv : ivs)
     iv.type = loopVarType;
 
+  auto ctx = parser.getBuilder().getContext();
   // Parse "inclusive" flag.
   if (succeeded(parser.parseOptionalKeyword("inclusive")))
-    result.addAttribute("loop_inclusive",
-                        UnitAttr::get(parser.getBuilder().getContext()));
+    result.addAttribute("loop_inclusive", UnitAttr::get(ctx));
 
   // Parse step values.
   SmallVector<OpAsmParser::UnresolvedOperand> steps;
@@ -2950,6 +2955,38 @@ ParseResult LoopNestOp::parse(OpAsmParser &parser, OperationState &result) {
       parser.parseOperandList(steps, ivs.size(), OpAsmParser::Delimiter::Paren))
     return failure();
 
+  // Parse collapse
+  int64_t value = 0;
+  if (!parser.parseOptionalKeyword("collapse") &&
+      (parser.parseLParen() || parser.parseInteger(value) ||
+       parser.parseRParen()))
+    return failure();
+  if (value > 1) {
+    result.addAttribute(
+        "num_collapse",
+        IntegerAttr::get(parser.getBuilder().getI64Type(), value));
+  }
+
+  // Parse tiles
+  SmallVector<int64_t> tiles;
+  auto parseTiles = [&]() -> ParseResult {
+    int64_t tile;
+    if (parser.parseInteger(tile))
+      return failure();
+    tiles.push_back(tile);
+    return success();
+  };
+
+  if (!parser.parseOptionalKeyword("tiles") &&
+      (parser.parseLParen() ||
+      parser.parseCommaSeparatedList(parseTiles) ||
+       parser.parseRParen()))
+    return failure();
+
+  if (tiles.size() > 0) {
+    result.addAttribute("tile_sizes", DenseI64ArrayAttr::get(ctx, tiles));
+  }
+
   // Parse the body.
   Region *region = result.addRegion();
   if (parser.parseRegion(*region, ivs))
@@ -2973,14 +3010,23 @@ void LoopNestOp::print(OpAsmPrinter &p) {
   if (getLoopInclusive())
     p << "inclusive ";
   p << "step (" << getLoopSteps() << ") ";
+  if (int64_t numCollapse = getNumCollapse()) {
+    if (numCollapse > 1)
+      p << "collapse(" << numCollapse << ") ";
+  }
+  if (const auto tiles = getTileSizes()) {
+    p << "tiles(" << tiles.value() << ") ";
+  }
   p.printRegion(region, /*printEntryBlockArgs=*/false);
 }
 
 void LoopNestOp::build(OpBuilder &builder, OperationState &state,
                        const LoopNestOperands &clauses) {
+  MLIRContext *ctx = builder.getContext();
   LoopNestOp::build(builder, state, clauses.loopLowerBounds,
                     clauses.loopUpperBounds, clauses.loopSteps,
-                    clauses.loopInclusive);
+                    clauses.loopInclusive, clauses.numCollapse,
+                    makeDenseI64ArrayAttr(ctx, clauses.tileSizes));
 }
 
 LogicalResult LoopNestOp::verify() {
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index ec1a9c909c94a..0ded962434938 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2923,10 +2923,9 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
 /// Converts an OpenMP loop nest into LLVM IR using OpenMPIRBuilder.
 static LogicalResult
 convertOmpLoopNest(Operation &opInst, llvm::IRBuilderBase &builder,
-                   LLVM::ModuleTranslation &moduleTranslation) {
+                        LLVM::ModuleTranslation &moduleTranslation) {
   llvm::OpenMPIRBuilder *ompBuilder = moduleTranslation.getOpenMPBuilder();
   auto loopOp = cast<omp::LoopNestOp>(opInst);
-
   // Set up the source location value for OpenMP runtime.
   llvm::OpenMPIRBuilder::LocationDescription ompLoc(builder);
 
@@ -2992,18 +2991,60 @@ convertOmpLoopNest(Operation &opInst, llvm::IRBuilderBase &builder,
     loopInfos.push_back(*loopResult);
   }
 
-  // Collapse loops. Store the insertion point because LoopInfos may get
-  // invalidated.
+  //  llvm::OpenMPIRBuilder::InsertPointTy afterIP = builder.saveIP();
   llvm::OpenMPIRBuilder::InsertPointTy afterIP =
-      loopInfos.front()->getAfterIP();
+    loopInfos.front()->getAfterIP();
 
-  // Update the stack frame created for this loop to point to the resulting loop
-  // after applying transformations.
-  moduleTranslation.stackWalk<OpenMPLoopInfoStackFrame>(
-      [&](OpenMPLoopInfoStackFrame &frame) {
-        frame.loopInfo = ompBuilder->collapseLoops(ompLoc.DL, loopInfos, {});
-        return WalkResult::interrupt();
-      });
+  // Initialize the new loop info to the current one, in case there
+  // are no loop transformations done.
+  llvm::CanonicalLoopInfo *NewTopLoopInfo = nullptr;
+
+  // Do tiling
+  if (const auto &tiles = loopOp.getTileSizes()) {
+    llvm::Type *IVType = loopInfos.front()->getIndVarType();
+    SmallVector<llvm::Value *> TileSizes;
+
+    for (auto tile : tiles.value()) {
+      llvm::Value *TileVal =  llvm::ConstantInt::get(IVType, tile);
+      TileSizes.push_back(TileVal);
+    }
+
+    std::vector<llvm::CanonicalLoopInfo*> NewLoops =
+      ompBuilder->tileLoops(ompLoc.DL, loopInfos, TileSizes);
+
+    // Collapse loops. Store the insertion point because LoopInfos may get
+    // invalidated.
+    auto AfterBB = NewLoops.front()->getAfter();
+    auto AfterAfterBB = AfterBB->getSingleSuccessor();
+    afterIP = {AfterAfterBB, AfterAfterBB->begin()};
+    NewTopLoopInfo = NewLoops[0];
+
+    // Update the loop infos
+    loopInfos.clear();
+    for (const auto& newLoop : NewLoops) {
+      loopInfos.push_back(newLoop);
+    }
+  } // Tiling done
+
+  // Do collapse
+  if (const auto &numCollapse = loopOp.getNumCollapse()) {
+    SmallVector<llvm::CanonicalLoopInfo *> collapseLoopInfos(
+        loopInfos.begin(), loopInfos.begin() + (numCollapse));
+
+    auto newLoopInfo =
+        ompBuilder->collapseLoops(ompLoc.DL, collapseLoopInfos, {});
+    NewTopLoopInfo = newLoopInfo;
+  } // Collapse done
+
+  // Update the stack frame created for this loop to point to the resulting
+  // loop after applying transformations.
+  if (NewTopLoopInfo) {
+    moduleTranslation.stackWalk<OpenMPLoopInfoStackFrame>(
+        [&](OpenMPLoopInfoStackFrame &frame) {
+          frame.loopInfo = NewTopLoopInfo;
+          return WalkResult::interrupt();
+        });
+  }
 
   // Continue building IR after the loop. Note that the LoopInfo returned by
   // `collapseLoops` points inside the outermost loop and is intended for

>From ffd9374f2e6aa726dbf7b2dc47beea05a00e82fb Mon Sep 17 00:00:00 2001
From: Jan Leyonberg <jan_sjodin at yahoo.com>
Date: Tue, 10 Jun 2025 09:51:02 -0400
Subject: [PATCH 2/5] Fix tests and limit the nesting of construct to only
 tiling.

---
 flang/lib/Semantics/canonicalize-omp.cpp      | 34 ++++++++++++-------
 .../Lower/OpenMP/parallel-wsloop-lastpriv.f90 |  4 +--
 flang/test/Lower/OpenMP/simd.f90              |  2 +-
 flang/test/Lower/OpenMP/wsloop-variable.f90   |  2 +-
 flang/test/Semantics/OpenMP/do-collapse.f90   |  1 +
 .../LLVMIR/omptarget-wsloop-collapsed.mlir    |  2 +-
 mlir/test/Target/LLVMIR/openmp-llvm.mlir      | 12 +++----
 7 files changed, 33 insertions(+), 24 deletions(-)

diff --git a/flang/lib/Semantics/canonicalize-omp.cpp b/flang/lib/Semantics/canonicalize-omp.cpp
index 44e8045167798..67d2bbf31d0e2 100644
--- a/flang/lib/Semantics/canonicalize-omp.cpp
+++ b/flang/lib/Semantics/canonicalize-omp.cpp
@@ -133,27 +133,32 @@ class CanonicalizationOfOmp {
       // Ignore compiler directives.
       if (GetConstructIf<parser::CompilerDirective>(*nextIt))
         continue;
-
       // Keep track of the loops to handle the end loop directives
       std::stack<parser::OpenMPLoopConstruct *> loops;
       loops.push(&x);
-      while (auto *innerConstruct{
+      if (auto *innerConstruct{
           GetConstructIf<parser::OpenMPConstruct>(*nextIt)}) {
         if (auto *innerOmpLoop{
                 std::get_if<parser::OpenMPLoopConstruct>(&innerConstruct->u)}) {
-          std::get<
-              std::optional<common::Indirection<parser::OpenMPLoopConstruct>>>(
-              loops.top()->t) = std::move(*innerOmpLoop);
-          // Retrieveing the address so that DoConstruct or inner loop can be
-          // set later.
-          loops.push(&(std::get<std::optional<
-                           common::Indirection<parser::OpenMPLoopConstruct>>>(
-              loops.top()->t)
-                           .value()
-                           .value()));
-          nextIt = block.erase(nextIt);
+          auto &innerBeginDir{
+              std::get<parser::OmpBeginLoopDirective>(innerOmpLoop->t)};
+          auto &innerDir{std::get<parser::OmpLoopDirective>(innerBeginDir.t)};
+          if (innerDir.v == llvm::omp::Directive::OMPD_tile) {
+            std::get<std::optional<
+                common::Indirection<parser::OpenMPLoopConstruct>>>(
+                loops.top()->t) = std::move(*innerOmpLoop);
+            // Retrieveing the address so that DoConstruct or inner loop can be
+            // set later.
+            loops.push(&(std::get<std::optional<
+                             common::Indirection<parser::OpenMPLoopConstruct>>>(
+                loops.top()->t)
+                             .value()
+                             .value()));
+            nextIt = block.erase(nextIt);
+          }
         }
       }
+
       if (auto *doCons{GetConstructIf<parser::DoConstruct>(*nextIt)}) {
         if (doCons->GetLoopControl()) {
           std::get<std::optional<parser::DoConstruct>>(loops.top()->t) =
@@ -167,6 +172,9 @@ class CanonicalizationOfOmp {
                   loops.top()->t) = std::move(*endDir);
               nextIt = block.erase(nextIt);
               loops.pop();
+            } else {
+              // If there is a mismatch bail out.
+              break;
             }
           }
         } else {
diff --git a/flang/test/Lower/OpenMP/parallel-wsloop-lastpriv.f90 b/flang/test/Lower/OpenMP/parallel-wsloop-lastpriv.f90
index 2890e78e9d17f..faf8f717f6308 100644
--- a/flang/test/Lower/OpenMP/parallel-wsloop-lastpriv.f90
+++ b/flang/test/Lower/OpenMP/parallel-wsloop-lastpriv.f90
@@ -108,7 +108,7 @@ subroutine omp_do_lastprivate_collapse2(a)
   ! CHECK-NEXT: %[[UB2:.*]] = fir.load %[[ARG0_DECL]]#0 : !fir.ref<i32>
   ! CHECK-NEXT: %[[STEP2:.*]] = arith.constant 1 : i32
   ! CHECK-NEXT: omp.wsloop private(@{{.*}} %{{.*}}#0 -> %[[A_PVT_REF:.*]], @{{.*}} %{{.*}}#0 -> %[[I_PVT_REF:.*]], @{{.*}} %{{.*}}#0 -> %[[J_PVT_REF:.*]] : !fir.ref<i32>, !fir.ref<i32>, !fir.ref<i32>) {
-  ! CHECK-NEXT: omp.loop_nest (%[[ARG1:.*]], %[[ARG2:.*]]) : i32 = (%[[LB1]], %[[LB2]]) to (%[[UB1]], %[[UB2]]) inclusive step (%[[STEP1]], %[[STEP2]]) {
+  ! CHECK-NEXT: omp.loop_nest (%[[ARG1:.*]], %[[ARG2:.*]]) : i32 = (%[[LB1]], %[[LB2]]) to (%[[UB1]], %[[UB2]]) inclusive step (%[[STEP1]], %[[STEP2]]) collapse(2) {
   ! CHECK:      %[[A_PVT_DECL:.*]]:2 = hlfir.declare %[[A_PVT_REF]] {uniq_name = "_QFomp_do_lastprivate_collapse2Ea"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
   ! CHECK:      %[[I_PVT_DECL:.*]]:2 = hlfir.declare %[[I_PVT_REF]] {uniq_name = "_QFomp_do_lastprivate_collapse2Ei"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
   ! CHECK:      %[[J_PVT_DECL:.*]]:2 = hlfir.declare %[[J_PVT_REF]] {uniq_name = "_QFomp_do_lastprivate_collapse2Ej"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
@@ -174,7 +174,7 @@ subroutine omp_do_lastprivate_collapse3(a)
   ! CHECK-NEXT: %[[UB3:.*]] = fir.load %[[ARG0_DECL]]#0 : !fir.ref<i32>
   ! CHECK-NEXT: %[[STEP3:.*]] = arith.constant 1 : i32
   ! CHECK-NEXT: omp.wsloop private(@{{.*}} %{{.*}}#0 -> %[[A_PVT_REF:.*]], @{{.*}} %{{.*}}#0 -> %[[I_PVT_REF:.*]], @{{.*}} %{{.*}}#0 -> %[[J_PVT_REF:.*]], @{{.*}} %{{.*}}#0 -> %[[K_PVT_REF:.*]] : !fir.ref<i32>, !fir.ref<i32>, !fir.ref<i32>, !fir.ref<i32>) {
-  ! CHECK-NEXT: omp.loop_nest (%[[ARG1:.*]], %[[ARG2:.*]], %[[ARG3:.*]]) : i32 = (%[[LB1]], %[[LB2]], %[[LB3]]) to (%[[UB1]], %[[UB2]], %[[UB3]]) inclusive step (%[[STEP1]], %[[STEP2]], %[[STEP3]]) {
+  ! CHECK-NEXT: omp.loop_nest (%[[ARG1:.*]], %[[ARG2:.*]], %[[ARG3:.*]]) : i32 = (%[[LB1]], %[[LB2]], %[[LB3]]) to (%[[UB1]], %[[UB2]], %[[UB3]]) inclusive step (%[[STEP1]], %[[STEP2]], %[[STEP3]]) collapse(3) {
   ! CHECK:      %[[A_PVT_DECL:.*]]:2 = hlfir.declare %[[A_PVT_REF]] {uniq_name = "_QFomp_do_lastprivate_collapse3Ea"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
   ! CHECK:      %[[I_PVT_DECL:.*]]:2 = hlfir.declare %[[I_PVT_REF]] {uniq_name = "_QFomp_do_lastprivate_collapse3Ei"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
   ! CHECK:      %[[J_PVT_DECL:.*]]:2 = hlfir.declare %[[J_PVT_REF]] {uniq_name = "_QFomp_do_lastprivate_collapse3Ej"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
diff --git a/flang/test/Lower/OpenMP/simd.f90 b/flang/test/Lower/OpenMP/simd.f90
index d815474b84b31..3572b9baff00b 100644
--- a/flang/test/Lower/OpenMP/simd.f90
+++ b/flang/test/Lower/OpenMP/simd.f90
@@ -175,7 +175,7 @@ subroutine simd_with_collapse_clause(n)
   ! CHECK-NEXT: omp.loop_nest (%[[ARG_0:.*]], %[[ARG_1:.*]]) : i32 = (
   ! CHECK-SAME:                %[[LOWER_I]], %[[LOWER_J]]) to (
   ! CHECK-SAME:                %[[UPPER_I]], %[[UPPER_J]]) inclusive step (
-  ! CHECK-SAME:                %[[STEP_I]], %[[STEP_J]]) {
+  ! CHECK-SAME:                %[[STEP_I]], %[[STEP_J]]) collapse(2) {
   !$OMP SIMD COLLAPSE(2)
   do i = 1, n
     do j = 1, n
diff --git a/flang/test/Lower/OpenMP/wsloop-variable.f90 b/flang/test/Lower/OpenMP/wsloop-variable.f90
index a7fb5fb8936e7..cceb77b974fee 100644
--- a/flang/test/Lower/OpenMP/wsloop-variable.f90
+++ b/flang/test/Lower/OpenMP/wsloop-variable.f90
@@ -23,7 +23,7 @@ program wsloop_variable
 !CHECK:      %[[TMP6:.*]] = fir.convert %[[TMP1]] : (i32) -> i64
 !CHECK:      %[[TMP7:.*]] = fir.convert %{{.*}} : (i32) -> i64
 !CHECK:      omp.wsloop private({{.*}}) {
-!CHECK-NEXT:   omp.loop_nest (%[[ARG0:.*]], %[[ARG1:.*]]) : i64 = (%[[TMP2]], %[[TMP5]]) to (%[[TMP3]], %[[TMP6]]) inclusive step (%[[TMP4]], %[[TMP7]]) {
+!CHECK-NEXT:   omp.loop_nest (%[[ARG0:.*]], %[[ARG1:.*]]) : i64 = (%[[TMP2]], %[[TMP5]]) to (%[[TMP3]], %[[TMP6]]) inclusive step (%[[TMP4]], %[[TMP7]]) collapse(2) {
 !CHECK:          %[[ARG0_I16:.*]] = fir.convert %[[ARG0]] : (i64) -> i16
 !CHECK:          hlfir.assign %[[ARG0_I16]] to %[[STORE_IV0:.*]]#0 : i16, !fir.ref<i16>
 !CHECK:          hlfir.assign %[[ARG1]] to %[[STORE_IV1:.*]]#0 : i64, !fir.ref<i64>
diff --git a/flang/test/Semantics/OpenMP/do-collapse.f90 b/flang/test/Semantics/OpenMP/do-collapse.f90
index 480bd45b79b83..ec6a3bdad3686 100644
--- a/flang/test/Semantics/OpenMP/do-collapse.f90
+++ b/flang/test/Semantics/OpenMP/do-collapse.f90
@@ -31,6 +31,7 @@ program omp_doCollapse
       end do
     end do
 
+  !ERROR: The value of the parameter in the COLLAPSE or ORDERED clause must not be larger than the number of nested loops following the construct.
   !ERROR: At most one COLLAPSE clause can appear on the SIMD directive
   !$omp simd collapse(2) collapse(1)
   do i = 1, 4
diff --git a/mlir/test/Target/LLVMIR/omptarget-wsloop-collapsed.mlir b/mlir/test/Target/LLVMIR/omptarget-wsloop-collapsed.mlir
index 0ebcec0e0ec31..189e3bc57db86 100644
--- a/mlir/test/Target/LLVMIR/omptarget-wsloop-collapsed.mlir
+++ b/mlir/test/Target/LLVMIR/omptarget-wsloop-collapsed.mlir
@@ -9,7 +9,7 @@ module attributes {dlti.dl_spec = #dlti.dl_spec<#dlti.dl_entry<"dlti.alloca_memo
     %loop_lb = llvm.mlir.constant(0 : i32) : i32
     %loop_step = llvm.mlir.constant(1 : index) : i32
     omp.wsloop {
-      omp.loop_nest (%arg1, %arg2) : i32 = (%loop_lb, %loop_lb) to (%loop_ub, %loop_ub) inclusive step (%loop_step, %loop_step) {
+      omp.loop_nest (%arg1, %arg2) : i32 = (%loop_lb, %loop_lb) to (%loop_ub, %loop_ub) inclusive step (%loop_step, %loop_step) collapse(2) {
         %1 = llvm.add %arg1, %arg2  : i32
         %2 = llvm.mul %arg2, %loop_ub overflow<nsw>  : i32
         %3 = llvm.add %arg1, %2 :i32
diff --git a/mlir/test/Target/LLVMIR/openmp-llvm.mlir b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
index 9ad9e93301239..c0e1ad05a1625 100644
--- a/mlir/test/Target/LLVMIR/openmp-llvm.mlir
+++ b/mlir/test/Target/LLVMIR/openmp-llvm.mlir
@@ -786,7 +786,7 @@ llvm.func @simd_simple(%lb : i64, %ub : i64, %step : i64, %arg0: !llvm.ptr) {
 // CHECK-LABEL: @simd_simple_multiple
 llvm.func @simd_simple_multiple(%lb1 : i64, %ub1 : i64, %step1 : i64, %lb2 : i64, %ub2 : i64, %step2 : i64, %arg0: !llvm.ptr, %arg1: !llvm.ptr) {
   omp.simd {
-    omp.loop_nest (%iv1, %iv2) : i64 = (%lb1, %lb2) to (%ub1, %ub2) inclusive step (%step1, %step2) {
+    omp.loop_nest (%iv1, %iv2) : i64 = (%lb1, %lb2) to (%ub1, %ub2) inclusive step (%step1, %step2) collapse(2) {
       %3 = llvm.mlir.constant(2.000000e+00 : f32) : f32
       // The form of the emitted IR is controlled by OpenMPIRBuilder and
       // tested there. Just check that the right metadata is added and collapsed
@@ -824,7 +824,7 @@ llvm.func @simd_simple_multiple(%lb1 : i64, %ub1 : i64, %step1 : i64, %lb2 : i64
 // CHECK-LABEL: @simd_simple_multiple_simdlen
 llvm.func @simd_simple_multiple_simdlen(%lb1 : i64, %ub1 : i64, %step1 : i64, %lb2 : i64, %ub2 : i64, %step2 : i64, %arg0: !llvm.ptr, %arg1: !llvm.ptr) {
   omp.simd simdlen(2) {
-    omp.loop_nest (%iv1, %iv2) : i64 = (%lb1, %lb2) to (%ub1, %ub2) step (%step1, %step2) {
+    omp.loop_nest (%iv1, %iv2) : i64 = (%lb1, %lb2) to (%ub1, %ub2) step (%step1, %step2) collapse(2) {
       %3 = llvm.mlir.constant(2.000000e+00 : f32) : f32
       // The form of the emitted IR is controlled by OpenMPIRBuilder and
       // tested there. Just check that the right metadata is added.
@@ -848,7 +848,7 @@ llvm.func @simd_simple_multiple_simdlen(%lb1 : i64, %ub1 : i64, %step1 : i64, %l
 // CHECK-LABEL: @simd_simple_multiple_safelen
 llvm.func @simd_simple_multiple_safelen(%lb1 : i64, %ub1 : i64, %step1 : i64, %lb2 : i64, %ub2 : i64, %step2 : i64, %arg0: !llvm.ptr, %arg1: !llvm.ptr) {
   omp.simd safelen(2) {
-    omp.loop_nest (%iv1, %iv2) : i64 = (%lb1, %lb2) to (%ub1, %ub2) step (%step1, %step2) {
+    omp.loop_nest (%iv1, %iv2) : i64 = (%lb1, %lb2) to (%ub1, %ub2) step (%step1, %step2) collapse(2) {
       %3 = llvm.mlir.constant(2.000000e+00 : f32) : f32
       %4 = llvm.getelementptr %arg0[%iv1] : (!llvm.ptr, i64) -> !llvm.ptr, f32
       %5 = llvm.getelementptr %arg1[%iv2] : (!llvm.ptr, i64) -> !llvm.ptr, f32
@@ -867,7 +867,7 @@ llvm.func @simd_simple_multiple_safelen(%lb1 : i64, %ub1 : i64, %step1 : i64, %l
 // CHECK-LABEL: @simd_simple_multiple_simdlen_safelen
 llvm.func @simd_simple_multiple_simdlen_safelen(%lb1 : i64, %ub1 : i64, %step1 : i64, %lb2 : i64, %ub2 : i64, %step2 : i64, %arg0: !llvm.ptr, %arg1: !llvm.ptr) {
   omp.simd simdlen(1) safelen(2) {
-    omp.loop_nest (%iv1, %iv2) : i64 = (%lb1, %lb2) to (%ub1, %ub2) step (%step1, %step2) {
+    omp.loop_nest (%iv1, %iv2) : i64 = (%lb1, %lb2) to (%ub1, %ub2) step (%step1, %step2) collapse(2) {
       %3 = llvm.mlir.constant(2.000000e+00 : f32) : f32
       %4 = llvm.getelementptr %arg0[%iv1] : (!llvm.ptr, i64) -> !llvm.ptr, f32
       %5 = llvm.getelementptr %arg1[%iv2] : (!llvm.ptr, i64) -> !llvm.ptr, f32
@@ -1267,7 +1267,7 @@ llvm.func @collapse_wsloop(
     // CHECK: store i32 %[[TOTAL_SUB_1]], ptr
     // CHECK: call void @__kmpc_for_static_init_4u
     omp.wsloop {
-      omp.loop_nest (%arg0, %arg1, %arg2) : i32 = (%0, %1, %2) to (%3, %4, %5) step (%6, %7, %8) {
+      omp.loop_nest (%arg0, %arg1, %arg2) : i32 = (%0, %1, %2) to (%3, %4, %5) step (%6, %7, %8) collapse(3) {
         %31 = llvm.load %20 : !llvm.ptr -> i32
         %32 = llvm.add %31, %arg0 : i32
         %33 = llvm.add %32, %arg1 : i32
@@ -1329,7 +1329,7 @@ llvm.func @collapse_wsloop_dynamic(
     // CHECK: store i32 %[[TOTAL]], ptr
     // CHECK: call void @__kmpc_dispatch_init_4u
     omp.wsloop schedule(dynamic) {
-      omp.loop_nest (%arg0, %arg1, %arg2) : i32 = (%0, %1, %2) to (%3, %4, %5) step (%6, %7, %8) {
+      omp.loop_nest (%arg0, %arg1, %arg2) : i32 = (%0, %1, %2) to (%3, %4, %5) step (%6, %7, %8) collapse(3) {
         %31 = llvm.load %20 : !llvm.ptr -> i32
         %32 = llvm.add %31, %arg0 : i32
         %33 = llvm.add %32, %arg1 : i32

>From dd11708798afab0ee6c1e9f2a9aad670c5fc68d8 Mon Sep 17 00:00:00 2001
From: Jan Leyonberg <jan_sjodin at yahoo.com>
Date: Tue, 10 Jun 2025 10:18:32 -0400
Subject: [PATCH 3/5] Enable stand-alone tiling, but it gives a warning and
 converting to simd.

---
 flang/lib/Lower/OpenMP/OpenMP.cpp           | 37 +++++++++++++++++++++
 flang/test/Lower/OpenMP/wsloop-collapse.f90 |  2 +-
 2 files changed, 38 insertions(+), 1 deletion(-)

diff --git a/flang/lib/Lower/OpenMP/OpenMP.cpp b/flang/lib/Lower/OpenMP/OpenMP.cpp
index d6f6ba70c9401..ae796048540b9 100644
--- a/flang/lib/Lower/OpenMP/OpenMP.cpp
+++ b/flang/lib/Lower/OpenMP/OpenMP.cpp
@@ -2153,6 +2153,40 @@ genLoopOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
   return loopOp;
 }
 
+static mlir::omp::LoopOp
+genTiledLoopOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
+          semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval,
+          mlir::Location loc, const ConstructQueue &queue,
+          ConstructQueue::const_iterator item) {
+  mlir::omp::LoopOperands loopClauseOps;
+  llvm::SmallVector<const semantics::Symbol *> loopReductionSyms;
+  genLoopClauses(converter, semaCtx, item->clauses, loc, loopClauseOps,
+                 loopReductionSyms);
+
+  DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval,
+                           /*shouldCollectPreDeterminedSymbols=*/true,
+                           /*useDelayedPrivatization=*/true, symTable);
+  dsp.processStep1(&loopClauseOps);
+
+  mlir::omp::LoopNestOperands loopNestClauseOps;
+  llvm::SmallVector<const semantics::Symbol *> iv;
+  genLoopNestClauses(converter, semaCtx, eval, item->clauses, loc,
+                     loopNestClauseOps, iv);
+
+  EntryBlockArgs loopArgs;
+  loopArgs.priv.syms = dsp.getDelayedPrivSymbols();
+  loopArgs.priv.vars = loopClauseOps.privateVars;
+  loopArgs.reduction.syms = loopReductionSyms;
+  loopArgs.reduction.vars = loopClauseOps.reductionVars;
+
+  auto loopOp =
+      genWrapperOp<mlir::omp::LoopOp>(converter, loc, loopClauseOps, loopArgs);
+  genLoopNestOp(converter, symTable, semaCtx, eval, loc, queue, item,
+                loopNestClauseOps, iv, {{loopOp, loopArgs}},
+                llvm::omp::Directive::OMPD_loop, dsp);
+  return loopOp;
+}
+
 static mlir::omp::MaskedOp
 genMaskedOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
             lower::StatementContext &stmtCtx,
@@ -3972,6 +4006,9 @@ static void genOMPDispatch(lower::AbstractConverter &converter,
                        item);
     break;
   case llvm::omp::Directive::OMPD_tile:
+    newOp =
+      genTiledLoopOp(converter, symTable, semaCtx, eval, loc, queue, item);
+    break;
   case llvm::omp::Directive::OMPD_unroll: {
     unsigned version = semaCtx.langOptions().OpenMPVersion;
     TODO(loc, "Unhandled loop directive (" +
diff --git a/flang/test/Lower/OpenMP/wsloop-collapse.f90 b/flang/test/Lower/OpenMP/wsloop-collapse.f90
index a4d5cbdc03d3e..0f27cd6f6cd7c 100644
--- a/flang/test/Lower/OpenMP/wsloop-collapse.f90
+++ b/flang/test/Lower/OpenMP/wsloop-collapse.f90
@@ -57,7 +57,7 @@ program wsloop_collapse
 !CHECK:           %[[VAL_31:.*]] = fir.load %[[VAL_11]]#0 : !fir.ref<i32>
 !CHECK:           %[[VAL_32:.*]] = arith.constant 1 : i32
 !CHECK:           omp.wsloop private(@{{.*}} %{{.*}}#0 -> %[[VAL_4:.*]], @{{.*}} %{{.*}}#0 -> %[[VAL_2:.*]], @{{.*}} %{{.*}}#0 -> %[[VAL_0:.*]] : !fir.ref<i32>, !fir.ref<i32>, !fir.ref<i32>) {
-!CHECK-NEXT:        omp.loop_nest (%[[VAL_33:.*]], %[[VAL_34:.*]], %[[VAL_35:.*]]) : i32 = (%[[VAL_24]], %[[VAL_27]], %[[VAL_30]]) to (%[[VAL_25]], %[[VAL_28]], %[[VAL_31]]) inclusive step (%[[VAL_26]], %[[VAL_29]], %[[VAL_32]]) {
+!CHECK-NEXT:        omp.loop_nest (%[[VAL_33:.*]], %[[VAL_34:.*]], %[[VAL_35:.*]]) : i32 = (%[[VAL_24]], %[[VAL_27]], %[[VAL_30]]) to (%[[VAL_25]], %[[VAL_28]], %[[VAL_31]]) inclusive step (%[[VAL_26]], %[[VAL_29]], %[[VAL_32]]) collapse(3) {
   !$omp do collapse(3)
   do i = 1, a
      do j= 1, b

>From 9393eccfab69facefe431133e6730c6a241ae533 Mon Sep 17 00:00:00 2001
From: Jan Leyonberg <jan_sjodin at yahoo.com>
Date: Wed, 11 Jun 2025 10:25:00 -0400
Subject: [PATCH 4/5] Add minimal test, remove debug print.

---
 flang/test/Lower/OpenMP/wsloop-tile.f90       | 30 +++++++++++++++++++
 .../Frontend/OpenMP/ConstructDecompositionT.h |  2 +-
 2 files changed, 31 insertions(+), 1 deletion(-)
 create mode 100644 flang/test/Lower/OpenMP/wsloop-tile.f90

diff --git a/flang/test/Lower/OpenMP/wsloop-tile.f90 b/flang/test/Lower/OpenMP/wsloop-tile.f90
new file mode 100644
index 0000000000000..f43b558ce46bb
--- /dev/null
+++ b/flang/test/Lower/OpenMP/wsloop-tile.f90
@@ -0,0 +1,30 @@
+! This test checks lowering of OpenMP DO Directive(Worksharing) with collapse.
+
+! RUN: bbc -fopenmp -fopenmp-version=51 -emit-hlfir %s -o - | FileCheck %s
+
+!CHECK-LABEL:   func.func @_QQmain() attributes {fir.bindc_name = "wsloop_tile"} {
+program wsloop_tile
+  integer :: i, j, k
+  integer :: a, b, c
+  integer :: x
+
+  a=30
+  b=20
+  c=50
+  x=0
+
+  !CHECK: omp.loop_nest
+  !CHECK-SAME: tiles(2, 5, 10)
+
+  !$omp do
+  !$omp tile sizes(2,5,10)
+  do i = 1, a
+     do j= 1, b
+        do k = 1, c
+           x = x + i + j + k
+        end do
+     end do
+  end do
+  !$omp end tile
+  !$omp end do
+end program wsloop_tile
diff --git a/llvm/include/llvm/Frontend/OpenMP/ConstructDecompositionT.h b/llvm/include/llvm/Frontend/OpenMP/ConstructDecompositionT.h
index dea8cf0bada10..6d41d41b98183 100644
--- a/llvm/include/llvm/Frontend/OpenMP/ConstructDecompositionT.h
+++ b/llvm/include/llvm/Frontend/OpenMP/ConstructDecompositionT.h
@@ -498,7 +498,7 @@ bool ConstructDecompositionT<C, H>::applyClause(
       last.clauses.push_back(node);
       return true;
     } else {
-      llvm::errs() << "** OVERRIDING isAllowedClauseForDirective **\n";
+      //      llvm::errs() << "** OVERRIDING isAllowedClauseForDirective **\n";
       last.clauses.push_back(node);
       return true;
     }

>From b52269cd2283985be89a0e3203efc54b6f921a09 Mon Sep 17 00:00:00 2001
From: Jan Leyonberg <jan_sjodin at yahoo.com>
Date: Fri, 13 Jun 2025 09:53:10 -0400
Subject: [PATCH 5/5] Fix formatting

---
 flang/include/flang/Parser/parse-tree.h       |  6 ++---
 flang/lib/Lower/OpenMP/OpenMP.cpp             | 27 +++++++++++--------
 flang/lib/Lower/OpenMP/Utils.cpp              |  3 +--
 flang/lib/Semantics/canonicalize-omp.cpp      |  4 +--
 flang/lib/Semantics/resolve-directives.cpp    |  3 ++-
 mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp  |  3 +--
 .../OpenMP/OpenMPToLLVMIRTranslation.cpp      | 12 ++++-----
 7 files changed, 31 insertions(+), 27 deletions(-)

diff --git a/flang/include/flang/Parser/parse-tree.h b/flang/include/flang/Parser/parse-tree.h
index 4ae1c354d504a..8922e6421e150 100644
--- a/flang/include/flang/Parser/parse-tree.h
+++ b/flang/include/flang/Parser/parse-tree.h
@@ -5075,10 +5075,10 @@ struct OpenMPBlockConstruct {
 struct OpenMPLoopConstruct {
   TUPLE_CLASS_BOILERPLATE(OpenMPLoopConstruct);
   OpenMPLoopConstruct(OmpBeginLoopDirective &&a)
-    : t({std::move(a), std::nullopt, std::nullopt, std::nullopt}) {}
+      : t({std::move(a), std::nullopt, std::nullopt, std::nullopt}) {}
   std::tuple<OmpBeginLoopDirective, std::optional<DoConstruct>,
-             std::optional<common::Indirection<OpenMPLoopConstruct>>,
-             std::optional<OmpEndLoopDirective>>
+      std::optional<common::Indirection<OpenMPLoopConstruct>>,
+      std::optional<OmpEndLoopDirective>>
       t;
 };
 
diff --git a/flang/lib/Lower/OpenMP/OpenMP.cpp b/flang/lib/Lower/OpenMP/OpenMP.cpp
index ae796048540b9..d01e34b852369 100644
--- a/flang/lib/Lower/OpenMP/OpenMP.cpp
+++ b/flang/lib/Lower/OpenMP/OpenMP.cpp
@@ -477,9 +477,10 @@ static void processHostEvalClauses(lower::AbstractConverter &converter,
               if (innerOptional.has_value()) {
                 const auto &innerLoopDirective = innerOptional.value().value();
                 const auto &innerBegin =
-                  std::get<parser::OmpBeginLoopDirective>(innerLoopDirective.t);
+                    std::get<parser::OmpBeginLoopDirective>(
+                        innerLoopDirective.t);
                 const auto &innerDirective =
-                  std::get<parser::OmpLoopDirective>(innerBegin.t);
+                    std::get<parser::OmpLoopDirective>(innerBegin.t);
                 if (innerDirective.v == llvm::omp::Directive::OMPD_tile) {
                   middleClauseList =
                       &std::get<parser::OmpClauseList>(innerBegin.t);
@@ -2153,11 +2154,13 @@ genLoopOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
   return loopOp;
 }
 
-static mlir::omp::LoopOp
-genTiledLoopOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
-          semantics::SemanticsContext &semaCtx, lower::pft::Evaluation &eval,
-          mlir::Location loc, const ConstructQueue &queue,
-          ConstructQueue::const_iterator item) {
+static mlir::omp::LoopOp genTiledLoopOp(lower::AbstractConverter &converter,
+                                        lower::SymMap &symTable,
+                                        semantics::SemanticsContext &semaCtx,
+                                        lower::pft::Evaluation &eval,
+                                        mlir::Location loc,
+                                        const ConstructQueue &queue,
+                                        ConstructQueue::const_iterator item) {
   mlir::omp::LoopOperands loopClauseOps;
   llvm::SmallVector<const semantics::Symbol *> loopReductionSyms;
   genLoopClauses(converter, semaCtx, item->clauses, loc, loopClauseOps,
@@ -4007,7 +4010,7 @@ static void genOMPDispatch(lower::AbstractConverter &converter,
     break;
   case llvm::omp::Directive::OMPD_tile:
     newOp =
-      genTiledLoopOp(converter, symTable, semaCtx, eval, loc, queue, item);
+        genTiledLoopOp(converter, symTable, semaCtx, eval, loc, queue, item);
     break;
   case llvm::omp::Directive::OMPD_unroll: {
     unsigned version = semaCtx.langOptions().OpenMPVersion;
@@ -4461,13 +4464,15 @@ static void genOMP(lower::AbstractConverter &converter, lower::SymMap &symTable,
   List<Clause> clauses = makeClauses(
       std::get<parser::OmpClauseList>(beginLoopDirective.t), semaCtx);
 
-  const auto &innerOptional = std::get<std::optional<common::Indirection<parser::OpenMPLoopConstruct>>>(loopConstruct.t);
+  const auto &innerOptional =
+      std::get<std::optional<common::Indirection<parser::OpenMPLoopConstruct>>>(
+          loopConstruct.t);
   if (innerOptional.has_value()) {
     const auto &innerLoopDirective = innerOptional.value().value();
     const auto &innerBegin =
-      std::get<parser::OmpBeginLoopDirective>(innerLoopDirective.t);
+        std::get<parser::OmpBeginLoopDirective>(innerLoopDirective.t);
     const auto &innerDirective =
-      std::get<parser::OmpLoopDirective>(innerBegin.t);
+        std::get<parser::OmpLoopDirective>(innerBegin.t);
     if (innerDirective.v == llvm::omp::Directive::OMPD_tile) {
       clauses.append(
           makeClauses(std::get<parser::OmpClauseList>(innerBegin.t), semaCtx));
diff --git a/flang/lib/Lower/OpenMP/Utils.cpp b/flang/lib/Lower/OpenMP/Utils.cpp
index 453dec9fbf3d7..69b5672a22aa8 100644
--- a/flang/lib/Lower/OpenMP/Utils.cpp
+++ b/flang/lib/Lower/OpenMP/Utils.cpp
@@ -51,8 +51,7 @@ int64_t getCollapseValue(const List<Clause> &clauses) {
   }
 
   collapseValue = collapseValue - numTileSizes;
-  int64_t result =
-    collapseValue > numTileSizes ? collapseValue : numTileSizes;
+  int64_t result = collapseValue > numTileSizes ? collapseValue : numTileSizes;
   return result;
 }
 
diff --git a/flang/lib/Semantics/canonicalize-omp.cpp b/flang/lib/Semantics/canonicalize-omp.cpp
index 67d2bbf31d0e2..8601c7c0eab1b 100644
--- a/flang/lib/Semantics/canonicalize-omp.cpp
+++ b/flang/lib/Semantics/canonicalize-omp.cpp
@@ -8,7 +8,7 @@
 
 #include "canonicalize-omp.h"
 #include "flang/Parser/parse-tree-visitor.h"
-# include <stack>
+#include <stack>
 // After Loop Canonicalization, rewrite OpenMP parse tree to make OpenMP
 // Constructs more structured which provide explicit scopes for later
 // structural checks and semantic analysis.
@@ -137,7 +137,7 @@ class CanonicalizationOfOmp {
       std::stack<parser::OpenMPLoopConstruct *> loops;
       loops.push(&x);
       if (auto *innerConstruct{
-          GetConstructIf<parser::OpenMPConstruct>(*nextIt)}) {
+              GetConstructIf<parser::OpenMPConstruct>(*nextIt)}) {
         if (auto *innerOmpLoop{
                 std::get_if<parser::OpenMPLoopConstruct>(&innerConstruct->u)}) {
           auto &innerBeginDir{
diff --git a/flang/lib/Semantics/resolve-directives.cpp b/flang/lib/Semantics/resolve-directives.cpp
index 2cfa6a2867e64..44e08422a3fd6 100644
--- a/flang/lib/Semantics/resolve-directives.cpp
+++ b/flang/lib/Semantics/resolve-directives.cpp
@@ -2061,7 +2061,8 @@ void OmpAttributeVisitor::PrivatizeAssociatedLoopIndexAndCheckLoopLevel(
     innerMostLoop = &innerLoop.value().value();
   }
 
-  const auto &outer{std::get<std::optional<parser::DoConstruct>>(innerMostLoop->t)};
+  const auto &outer{
+      std::get<std::optional<parser::DoConstruct>>(innerMostLoop->t)};
   if (outer.has_value()) {
     for (const parser::DoConstruct *loop{&*outer}; loop && level > 0; --level) {
       // go through all the nested do-loops and resolve index variables
diff --git a/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp b/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
index 53e3bf9793569..d6eb4e334bcda 100644
--- a/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
+++ b/mlir/lib/Dialect/OpenMP/IR/OpenMPDialect.cpp
@@ -2978,8 +2978,7 @@ ParseResult LoopNestOp::parse(OpAsmParser &parser, OperationState &result) {
   };
 
   if (!parser.parseOptionalKeyword("tiles") &&
-      (parser.parseLParen() ||
-      parser.parseCommaSeparatedList(parseTiles) ||
+      (parser.parseLParen() || parser.parseCommaSeparatedList(parseTiles) ||
        parser.parseRParen()))
     return failure();
 
diff --git a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
index 0ded962434938..4f570e2d863e0 100644
--- a/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
+++ b/mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp
@@ -2923,7 +2923,7 @@ convertOmpSimd(Operation &opInst, llvm::IRBuilderBase &builder,
 /// Converts an OpenMP loop nest into LLVM IR using OpenMPIRBuilder.
 static LogicalResult
 convertOmpLoopNest(Operation &opInst, llvm::IRBuilderBase &builder,
-                        LLVM::ModuleTranslation &moduleTranslation) {
+                   LLVM::ModuleTranslation &moduleTranslation) {
   llvm::OpenMPIRBuilder *ompBuilder = moduleTranslation.getOpenMPBuilder();
   auto loopOp = cast<omp::LoopNestOp>(opInst);
   // Set up the source location value for OpenMP runtime.
@@ -2993,7 +2993,7 @@ convertOmpLoopNest(Operation &opInst, llvm::IRBuilderBase &builder,
 
   //  llvm::OpenMPIRBuilder::InsertPointTy afterIP = builder.saveIP();
   llvm::OpenMPIRBuilder::InsertPointTy afterIP =
-    loopInfos.front()->getAfterIP();
+      loopInfos.front()->getAfterIP();
 
   // Initialize the new loop info to the current one, in case there
   // are no loop transformations done.
@@ -3005,12 +3005,12 @@ convertOmpLoopNest(Operation &opInst, llvm::IRBuilderBase &builder,
     SmallVector<llvm::Value *> TileSizes;
 
     for (auto tile : tiles.value()) {
-      llvm::Value *TileVal =  llvm::ConstantInt::get(IVType, tile);
+      llvm::Value *TileVal = llvm::ConstantInt::get(IVType, tile);
       TileSizes.push_back(TileVal);
     }
 
-    std::vector<llvm::CanonicalLoopInfo*> NewLoops =
-      ompBuilder->tileLoops(ompLoc.DL, loopInfos, TileSizes);
+    std::vector<llvm::CanonicalLoopInfo *> NewLoops =
+        ompBuilder->tileLoops(ompLoc.DL, loopInfos, TileSizes);
 
     // Collapse loops. Store the insertion point because LoopInfos may get
     // invalidated.
@@ -3021,7 +3021,7 @@ convertOmpLoopNest(Operation &opInst, llvm::IRBuilderBase &builder,
 
     // Update the loop infos
     loopInfos.clear();
-    for (const auto& newLoop : NewLoops) {
+    for (const auto &newLoop : NewLoops) {
       loopInfos.push_back(newLoop);
     }
   } // Tiling done



More information about the flang-commits mailing list