[Mlir-commits] [mlir] [MLIR][SCF] Fix normalizeForallOp helper function (PR #138615)

[Colin De Vlieghere]

================
@@ -1482,30 +1482,31 @@ FailureOr<scf::ForallOp> mlir::normalizeForallOp(RewriterBase &rewriter,
   SmallVector<OpFoldResult> ubs = forallOp.getMixedUpperBound();
   SmallVector<OpFoldResult> steps = forallOp.getMixedStep();
 
-  if (llvm::all_of(
-          lbs, [](OpFoldResult ofr) { return isConstantIntValue(ofr, 0); }) &&
-      llvm::all_of(
-          steps, [](OpFoldResult ofr) { return isConstantIntValue(ofr, 1); })) {
+  if (forallOp.isNormalized())
     return forallOp;
-  }
 
-  SmallVector<OpFoldResult> newLbs, newUbs, newSteps;
+  OpBuilder::InsertionGuard g(rewriter);
+  rewriter.setInsertionPoint(forallOp);
+  SmallVector<OpFoldResult> newUbs;
   for (auto [lb, ub, step] : llvm::zip_equal(lbs, ubs, steps)) {
     Range normalizedLoopParams =
         emitNormalizedLoopBounds(rewriter, forallOp.getLoc(), lb, ub, step);
-    newLbs.push_back(normalizedLoopParams.offset);
     newUbs.push_back(normalizedLoopParams.size);
-    newSteps.push_back(normalizedLoopParams.stride);
   }
+  (void)foldDynamicIndexList(newUbs);
----------------
Cubevoid wrote:

`foldDynamicIndexList` returns success/failure based on if anything was folded, which I don't think we care about here. I just wanted to make sure that the upper bounds are always defined by a constant attribute, not an `arith::ConstantOp`.

https://github.com/llvm/llvm-project/pull/138615