[Openmp-commits] [clang] [llvm] [openmp] [Clang][OpenMP] Add interchange directive (PR #93022)

[Michael Kruse via Openmp-commits]

================
@@ -14853,6 +14861,158 @@ StmtResult SemaOpenMP::ActOnOpenMPReverseDirective(Stmt *AStmt,
                                      buildPreInits(Context, PreInits));
 }
 
+StmtResult SemaOpenMP::ActOnOpenMPInterchangeDirective(
+    ArrayRef<OMPClause *> Clauses, Stmt *AStmt, SourceLocation StartLoc,
+    SourceLocation EndLoc) {
+  ASTContext &Context = getASTContext();
+  DeclContext *CurContext = SemaRef.CurContext;
+  Scope *CurScope = SemaRef.getCurScope();
+
+  // Empty statement should only be possible if there already was an error.
+  if (!AStmt)
+    return StmtError();
+
+  // interchange without permutation clause swaps two loops.
+  constexpr size_t NumLoops = 2;
+
+  // Verify and diagnose loop nest.
+  SmallVector<OMPLoopBasedDirective::HelperExprs, 4> LoopHelpers(NumLoops);
+  Stmt *Body = nullptr;
+  SmallVector<SmallVector<Stmt *, 0>, 2> OriginalInits;
+  if (!checkTransformableLoopNest(OMPD_interchange, AStmt, NumLoops,
+                                  LoopHelpers, Body, OriginalInits))
+    return StmtError();
+
+  // Delay interchange to when template is completely instantiated.
+  if (CurContext->isDependentContext())
+    return OMPInterchangeDirective::Create(Context, StartLoc, EndLoc, Clauses,
+                                           NumLoops, AStmt, nullptr, nullptr);
+
+  assert(LoopHelpers.size() == NumLoops &&
+         "Expecting loop iteration space dimensionaly to match number of "
+         "affected loops");
+  assert(OriginalInits.size() == NumLoops &&
+         "Expecting loop iteration space dimensionaly to match number of "
+         "affected loops");
+
+  // Decode the permutation clause.
+  constexpr uint64_t Permutation[] = {1, 0};
+
+  // Find the affected loops.
+  SmallVector<Stmt *> LoopStmts(NumLoops, nullptr);
+  collectLoopStmts(AStmt, LoopStmts);
+
+  // Collect pre-init statements on the order before the permuation.
+  SmallVector<Stmt *> PreInits;
+  for (auto I : llvm::seq<int>(NumLoops)) {
+    OMPLoopBasedDirective::HelperExprs &LoopHelper = LoopHelpers[I];
+
+    assert(LoopHelper.Counters.size() == 1 &&
+           "Single-dimensional loop iteration space expected");
+    auto *OrigCntVar = cast<DeclRefExpr>(LoopHelper.Counters.front());
+
+    std::string OrigVarName = OrigCntVar->getNameInfo().getAsString();
----------------
Meinersbur wrote:

I don't get this. `getAsString` returns an `std::string` r-value. Assigning it to a `SmallString` will unpack that `std::string` and copy it over to `SmallString`'s buffer, compared to RVO where no additional copy happens.

Furthermore, contemporary implementations of `std::string` are [already using a small-size buffer](https://devblogs.microsoft.com/oldnewthing/20240510-00/?p=109742).

I guess I could call `OrigCntVar->getNameInfo().operator<<`, and pass a `raw_svector_ostream` instantiated using a `SmallString`. This would enable us to choose a larger small buffer size ourselves. Is that what you are asking for? If so, what should the small buffer size be?

https://github.com/llvm/llvm-project/pull/93022