[MLIR] Fixes arith.sub folder crash on dynamically shaped tensors

[joker-eph]

We can't create a constant for a value with dynamic shape.

Fixes #118772

[llvmbot]

@llvm/pr-subscribers-mlir-arith

@llvm/pr-subscribers-mlir

Author: Mehdi Amini (joker-eph)

Changes

We can't create a constant for a value with dynamic shape.

Fixes #118772

---

Full diff: https://github.com/llvm/llvm-project/pull/118908.diff

2 Files Affected:

• (modified) mlir/lib/Dialect/Arith/IR/ArithOps.cpp (+6-2)
• (modified) mlir/test/Dialect/Arith/canonicalize.mlir (+11)

diff --git a/mlir/lib/Dialect/Arith/IR/ArithOps.cpp b/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
index 5f445231b80fdf..5a79add16569f4 100644
--- a/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
+++ b/mlir/lib/Dialect/Arith/IR/ArithOps.cpp
@@ -393,8 +393,12 @@ void arith::AddUIExtendedOp::getCanonicalizationPatterns(
 
 OpFoldResult arith::SubIOp::fold(FoldAdaptor adaptor) {
   // subi(x,x) -> 0
-  if (getOperand(0) == getOperand(1))
-    return Builder(getContext()).getZeroAttr(getType());
+  if (getOperand(0) == getOperand(1)) {
+    auto tensorType = dyn_cast<TensorType>(getType());
+    // We can't generate a constant with a dynamic shaped tensor.
+    if (!tensorType || tensorType.hasStaticShape())
+      return Builder(getContext()).getZeroAttr(getType());
+  }
   // subi(x,0) -> x
   if (matchPattern(adaptor.getRhs(), m_Zero()))
     return getLhs();
diff --git a/mlir/test/Dialect/Arith/canonicalize.mlir b/mlir/test/Dialect/Arith/canonicalize.mlir
index 69df83d42f543e..8c22506b9afcf1 100644
--- a/mlir/test/Dialect/Arith/canonicalize.mlir
+++ b/mlir/test/Dialect/Arith/canonicalize.mlir
@@ -869,6 +869,17 @@ func.func @tripleAddAddOvf2(%arg0: index) -> index {
   return %add2 : index
 }
 
+
+// CHECK-LABEL: @foldSubXX
+//       CHECK:   %[[c0:.+]] = arith.constant dense<0> : tensor<10xi32> 
+//       CHECK:   %[[sub:.+]] = arith.subi
+//       CHECK:   return %[[c0]], %[[sub]]
+func.func @foldSubXX(%dyn : tensor<?x?xi32>, %static : tensor<10xi32>) -> (tensor<10xi32>, tensor<?x?xi32>) {
+  %static_sub = arith.subi %static, %static : tensor<10xi32>
+  %dyn_sub = arith.subi %dyn, %dyn : tensor<?x?xi32>
+  return %static_sub, %dyn_sub : tensor<10xi32>, tensor<?x?xi32>
+}
+
 // CHECK-LABEL: @tripleAddSub0
 //       CHECK:   %[[cres:.+]] = arith.constant 59 : index
 //       CHECK:   %[[add:.+]] = arith.subi %[[cres]], %arg0 : index

[Hardcode84]

Maybe check for ShapedType instead? In case we decide to introduce dynamically-sized vectors one day.

[Hardcode84]

Actually, we already kinda have them with scalabale vectors. Can you add test for them too?

[joker-eph]

Good point for using ShapedType.

The scalable vectors aren't an issue I believe since you can create a constant. They act as if they are statically shaped from this point of view.

[Hardcode84]

Thanks