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

[Michael Kruse via llvm-commits]

================
@@ -38,14 +59,21 @@ 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();
+    }
----------------
Meinersbur wrote:

This mixes the directive that collapse goes on (DO, DISTRIBUTE, ..) and those sizes goes on (TILE). During resolving merge conflicts of #145917 and #144785 the approach of handling tile like a compound construct already caused problems. E.g. it could be `tile` twice with different `sizes`, or `unroll partial` followed by `tile` vs the other way around, and there is no way to disambiguate them.

So it will need to be tangled apart eventually, and I would appreciate if we could avoid the assumption that there is just one `tile` which is always first.

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