[flang-commits] [flang] [flang][OpenMP] Rewrite standalone `loop` (without `bind`) directives to `simd` (PR #122632)

[Sergio Afonso via flang-commits]

================
@@ -100,6 +117,62 @@ class GenericLoopConversionPattern
     return result;
   }
 
+  /// Rewrites standalone `loop` directives to equivalent `simd` constructs.
+  /// The reasoning behind this decision is that according to the spec (version
+  /// 5.2, section 11.7.1):
+  ///
+  /// "If the bind clause is not specified on a construct for which it may be
+  /// specified and the construct is closely nested inside a teams or parallel
+  /// construct, the effect is as if binding is teams or parallel. If none of
+  /// those conditions hold, the binding region is not defined."
+  ///
+  /// which means that standalone `loop` directives have undefined binding
+  /// region. Moreover, the spec says (in the next paragraph):
+  ///
+  /// "The specified binding region determines the binding thread set.
+  /// Specifically, if the binding region is a teams region, then the binding
+  /// thread set is the set of initial threads that are executing that region
+  /// while if the binding region is a parallel region, then the binding thread
+  /// set is the team of threads that are executing that region. If the binding
+  /// region is not defined, then the binding thread set is the encountering
+  /// thread."
+  ///
+  /// which means that the binding thread set for a standalone `loop` directive
+  /// is only the encountering thread.
+  ///
+  /// Since the encountering thread is the binding thread (set) for a
+  /// standalone `loop` directive, the best we can do in such case is to "simd"
+  /// the directive.
+  void rewriteToSimdLoop(mlir::omp::LoopOp loopOp,
+                         mlir::ConversionPatternRewriter &rewriter) const {
+    loopOp.emitWarning("Detected standalone OpenMP `loop` directive, the "
+                       "associated loop will be rewritten to `simd`.");
----------------
skatrak wrote:

Got it, thanks for the explanation.

https://github.com/llvm/llvm-project/pull/122632