[Mlir-commits] [clang] [llvm] [mlir] [openmp] [Clang][OpenMP] Canonicalize Intra-tiles in Loop Tiling (PR #191114)

[Michael Kruse]

================
@@ -9695,8 +9744,19 @@ static bool checkOpenMPIterationSpace(
   std::tie(ResultIterSpaces[CurrentNestedLoopCount].MinValue,
            ResultIterSpaces[CurrentNestedLoopCount].MaxValue) =
       ISC.buildMinMaxValues(DSA.getCurScope(), Captures);
-  ResultIterSpaces[CurrentNestedLoopCount].FinalCondition =
-      ISC.buildFinalCondition(DSA.getCurScope());
+  if (TileRectCond) {
+    // The body guard is the overshoot predicate, which is absent when the tile
+    // can never be partial. Lower bound is the live .floor.iv; wire as a floor
+    // dependence for collapse
+    // (see design note in ActOnOpenMPTileDirective).
+    ResultIterSpaces[CurrentNestedLoopCount].FinalCondition = TileBodyPredicate;
+    ResultIterSpaces[CurrentNestedLoopCount].ReinterpretTileLB = true;
+    ResultIterSpaces[CurrentNestedLoopCount].ReinterpretTileLoc =
+        ISC.getInitSrcRange().getBegin();
+  } else {
+    ResultIterSpaces[CurrentNestedLoopCount].FinalCondition =
+        ISC.buildFinalCondition(DSA.getCurScope());
----------------
Meinersbur wrote:

The `ISC.buildFinalCondition` is not executed in a intratile loop anymore. Can it be either combined with `TileBodyPredicate` or `assert`-ed to be `nullptr`?

https://github.com/llvm/llvm-project/pull/191114