[Mlir-commits] [mlir] ac370b8 - [MLIR][Linalg] Fix LinalgSpecializeGenericOpsPass for scalar inputs (#208466)

[llvmlistbot at llvm.org]

Author: Hugo Trachino
Date: 2026-07-13T15:02:40+01:00
New Revision: ac370b8e838ada3dd482560c4b8a568e760f95b5

URL: https://github.com/llvm/llvm-project/commit/ac370b8e838ada3dd482560c4b8a568e760f95b5
DIFF: https://github.com/llvm/llvm-project/commit/ac370b8e838ada3dd482560c4b8a568e760f95b5.diff

LOG: [MLIR][Linalg] Fix LinalgSpecializeGenericOpsPass for scalar inputs (#208466)

The test case added would crash because the Ranked Tensor-ness of
operands was assumed but never checked. This crash only happened when there
were multiple inputs as it must bail out somewhere else when there is only
1 input.

Added: 
    

Modified: 
    mlir/lib/Dialect/Linalg/Transforms/DecomposeGenericByUnfoldingPermutation.cpp
    mlir/test/Dialect/Linalg/specialize-generic-ops-fail.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/lib/Dialect/Linalg/Transforms/DecomposeGenericByUnfoldingPermutation.cpp b/mlir/lib/Dialect/Linalg/Transforms/DecomposeGenericByUnfoldingPermutation.cpp
index 9015cbb096f88..8ab0f67d9b7de 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/DecomposeGenericByUnfoldingPermutation.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/DecomposeGenericByUnfoldingPermutation.cpp
@@ -164,8 +164,8 @@ LogicalResult DecomposeProjectedPermutation::matchAndRewrite(
   // out which operand can supply that runtime-value (tensor.dim).
   // Leaving it as a future TODO.
   if (llvm::any_of(op->getOpOperands(), [](OpOperand &oper) {
-        auto opType = cast<RankedTensorType>(oper.get().getType());
-        return ShapedType::isDynamicShape(opType.getShape());
+        auto opType = dyn_cast<RankedTensorType>(oper.get().getType());
+        return !opType || ShapedType::isDynamicShape(opType.getShape());
       }))
     return failure();
 
@@ -209,8 +209,8 @@ LogicalResult DecomposeProjectedPermutation::matchAndRewrite(
     // Does it require broadcast?
     if (!broadcastedDims.empty()) {
       assert(!broadcastedDims.empty() && "should have non size broadcast");
-      Value emptyTensor = tensor::EmptyOp::create(rewriter, loc, outputShape,
-                                                  inputRTType.getElementType());
+      Value emptyTensor =
+          tensor::EmptyOp::create(rewriter, loc, outputShape, elType);
 
       auto broadcastOp = linalg::BroadcastOp::create(
           rewriter, loc, newInitValues[i], emptyTensor, broadcastedDims);

diff  --git a/mlir/test/Dialect/Linalg/specialize-generic-ops-fail.mlir b/mlir/test/Dialect/Linalg/specialize-generic-ops-fail.mlir
index 5d66837fca510..f0792d7a696de 100644
--- a/mlir/test/Dialect/Linalg/specialize-generic-ops-fail.mlir
+++ b/mlir/test/Dialect/Linalg/specialize-generic-ops-fail.mlir
@@ -46,3 +46,19 @@ func.func @not_copy(%input: tensor<8xi32>, %init: tensor<8xi32>) -> tensor<8xi32
   } -> tensor<8xi32>
   return %res : tensor<8xi32>
 }
+
+
+// -----
+
+// CHECK-LABEL: @scalar_input
+// CHECK: linalg.generic
+#map = affine_map<(d0) -> (d0)>
+#map1 = affine_map<(d0) -> ()>
+func.func @scalar_input(%arg0: tensor<128xf32>, %arg1: f32) -> tensor<128xf32> {
+  %1 = linalg.generic {indexing_maps = [#map, #map1, #map], iterator_types = ["parallel"]} ins(%arg0, %arg1 : tensor<128xf32>, f32) outs(%arg0 : tensor<128xf32>) {
+  ^bb0(%in: f32, %in_1: f32, %out: f32):
+    %2 = arith.addf %in_1, %in_1 : f32
+    linalg.yield %2 : f32
+  } -> tensor<128xf32>
+  return %1 : tensor<128xf32>
+}