[Openmp-commits] [flang] [llvm] [mlir] [openmp] [MLIR][OpenMP] Add omp.fuse operation (PR #168898)

[Michael Kruse via Openmp-commits]

================
@@ -2216,6 +2218,64 @@ static void genTileOp(Fortran::lower::AbstractConverter &converter,
                             sizesClause.sizes);
 }
 
+static void genFuseOp(Fortran::lower::AbstractConverter &converter,
+                      Fortran::lower::SymMap &symTable,
+                      lower::StatementContext &stmtCtx,
+                      Fortran::semantics::SemanticsContext &semaCtx,
+                      Fortran::lower::pft::Evaluation &eval, mlir::Location loc,
+                      const ConstructQueue &queue,
+                      ConstructQueue::const_iterator item) {
+  fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
+
+  int32_t first = 0;
+  int32_t count = 0;
+  auto iter = llvm::find_if(item->clauses, [](const Clause &clause) {
+    return clause.id == llvm::omp::Clause::OMPC_looprange;
+  });
+  if (iter != item->clauses.end()) {
+    const auto &looprange = std::get<clause::Looprange>(iter->u);
+    first = evaluate::ToInt64(std::get<0>(looprange.t)).value();
+    count = evaluate::ToInt64(std::get<1>(looprange.t)).value();
+  }
+
+  llvm::SmallVector<mlir::Value> applyees;
+  for (auto &child : eval.getNestedEvaluations()) {
+    // Skip OmpEndLoopDirective
+    if (&child == &eval.getLastNestedEvaluation())
+      break;
+    // Skip Compiler Directive
+    if (child.getIf<parser::CompilerDirective>())
+      continue;
+    // Skip the
+    // if (nested.getIf<parser::NonLabelDoStmt>())
+    //  continue;
+
+    // Emit the associated loop
+    llvm::SmallVector<mlir::omp::CanonicalLoopOp> canonLoops;
+    genCanonicalLoopNest(converter, symTable, semaCtx, eval, &child, loc, queue,
+                         item, 1, canonLoops);
+
+    auto cli = llvm::getSingleElement(canonLoops).getCli();
+    applyees.push_back(cli);
+  }
+  // One generated loop + one for each loop not inside the specified looprange
+  // if present
+  llvm::SmallVector<mlir::Value> generatees;
+  int64_t numGeneratees = count == 0 ? 1 : applyees.size() - count + 1;
----------------
Meinersbur wrote:

Could you add to the comment that `count==0` represents no looprange clause present.

I assume semantics checks that you cannot use `fuse(1,0)`? 

[not a change request] IMHO it would have been a good idea to consider missing `looprange` somethine different than `count==0`. The latter logically would degernate to "fuse nothing". However, the OpenMP 6.0 spec says count must be positive.

https://github.com/llvm/llvm-project/pull/168898