[mlir][linalg]-Fix wrong assertion in the getMatchingYieldValue inter…

[amirBish]

…face

In order to have a consistent implementation for getMatchingYieldValue for linalg generic with buffer/tensor semantics, we should assert the opOperand index based on the numDpsInits and not numOfResults which may be zero in the buffer semantics.

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-linalg

Author: Amir Bishara (amirBish)

Changes

…face

In order to have a consistent implementation for getMatchingYieldValue for linalg generic with buffer/tensor semantics, we should assert the opOperand index based on the numDpsInits and not numOfResults which may be zero in the buffer semantics.

---

Full diff: https://github.com/llvm/llvm-project/pull/89590.diff

1 Files Affected:

• (modified) mlir/include/mlir/Dialect/Linalg/IR/LinalgInterfaces.td (+1-1)

diff --git a/mlir/include/mlir/Dialect/Linalg/IR/LinalgInterfaces.td b/mlir/include/mlir/Dialect/Linalg/IR/LinalgInterfaces.td
index fbf3f19cde0e9b..d20925da778e0a 100644
--- a/mlir/include/mlir/Dialect/Linalg/IR/LinalgInterfaces.td
+++ b/mlir/include/mlir/Dialect/Linalg/IR/LinalgInterfaces.td
@@ -486,7 +486,7 @@ def LinalgStructuredInterface
         int64_t resultIndex =
             opOperand->getOperandNumber() - $_op.getNumDpsInputs();
         assert(resultIndex >= 0 &&
-               resultIndex < this->getOperation()->getNumResults());
+               resultIndex < $_op.getNumDpsInits());
         Operation *yieldOp = getBlock()->getTerminator();
         return &yieldOp->getOpOperand(resultIndex);
       }]

[akuegel]

Could you please add a test that would trigger the wrong assert?

[amirBish]

Could you please add a test that would trigger the wrong assert?

Sure, should it be added as a unit test under mlir/unittests/Dialect or any other location?

[ftynse]

If it's possible to trigger this with on some input IR with mlir-opt and a pass, we should rather have that.

[amirBish]

If it's possible to trigger this with on some input IR with mlir-opt and a pass, we should rather have that.

Thanks for the suggestion, just checked/searched there is no usage for this method in mlir sources. any insights about how to test the change?

[amirBish]

If it's possible to trigger this with on some input IR with mlir-opt and a pass, we should rather have that.

Thanks for the suggestion, just checked/searched there is no usage for this method in mlir sources. any insights about how to test the change?

@ftynse ping :)