[MLIR] [Vector] Error when canonicalizing `vector.scatter` with tensor output

[sakupan102]

I tried to compile the code below with mlir-opt sample.mlir --pass-pipeline="builtin.module(func.func(canonicalize))"

func.func @scatter_tensor(%base: tensor<16xf32>, %mask: vector<16xi1>, %value: vector<16xf32>) -> tensor<16xf32> {
  %c0 = arith.constant 0 : index
  %index = vector.step : vector<16xindex>
  %0 = vector.scatter %base[%c0] [%index], %mask, %value
      : tensor<16xf32>, vector<16xindex>, vector<16xi1>, vector<16xf32> -> tensor<16xf32>
  return %0 : tensor<16xf32>
}

output

Assertion failed: (op->getNumResults() == newOp->getNumResults() && "ops have different number of results"), function replaceOp, file PatternMatch.cpp, line 144.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace and instructions to reproduce the bug.

Version

915e9ad

[llvmbot]

@llvm/issue-subscribers-mlir

Author: Ryutaro Okada (sakupan102)

I tried to compile the code below with `mlir-opt sample.mlir --pass-pipeline="builtin.module(func.func(canonicalize))"` ```mlir func.func @scatter_tensor(%base: tensor<16xf32>, %mask: vector<16xi1>, %value: vector<16xf32>) -> tensor<16xf32> { %c0 = arith.constant 0 : index %index = vector.step : vector<16xindex> %0 = vector.scatter %base[%c0] [%index], %mask, %value : tensor<16xf32>, vector<16xindex>, vector<16xi1>, vector<16xf32> -> tensor<16xf32> return %0 : tensor<16xf32> } ```

output

Assertion failed: (op->getNumResults() == newOp->getNumResults() && "ops have different number of results"), function replaceOp, file PatternMatch.cpp, line 144.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace and instructions to reproduce the bug.

Version

915e9ad

[sakupan102]

This error is due to 7e7ea9c (which I authored).
I'll work on this issue.