<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/118797>118797</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [mlir] Inconsistent output when executing MLIR program with and without `-linalg-fuse-elementwise-ops`
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            mlir
      </td>
    </tr>

    <tr>
      <th>Assignees</th>
      <td>
      </td>
    </tr>

    <tr>
      <th>Reporter</th>
      <td>
          Emilyaxe
      </td>
    </tr>
</table>

<pre>
    git version: adf892d743d

system: `Ubuntu 18.04.6 LTS`

## Description:
I am experiencing an inconsistent result when executing the same MLIR program with and without  `-linalg-fuse-elementwise-ops`.
The outputs are also inconsistent when running multiple times with the `-linalg-fuse-elementwise-ops`

## Steps to Reproduce:


### 1. **MLIR Program (a.mlir)**:
a.mlir: 
``` 
module {
  func.func private @printMemrefI32(tensor<*xi32>)
  func.func private @printMemrefF32(tensor<*xf32>)
  func.func @main() {
    %0 = "tosa.const"() <{value = dense<1641> : tensor<1x15x2xi32>}> : () -> tensor<1x15x2xi32>
    %1 = "tosa.const"() <{value = dense<3819> : tensor<1x15x2xi32>}> : () -> tensor<1x15x2xi32>
    %2 = tosa.abs %0 : (tensor<1x15x2xi32>) -> tensor<1x15x2xi32>
    %3 = tosa.logical_right_shift %2, %1 : (tensor<1x15x2xi32>, tensor<1x15x2xi32>) -> tensor<1x15x2xi32>
 %4 = tosa.reduce_sum %3 {axis = 1 : i32} : (tensor<1x15x2xi32>) -> tensor<1x1x2xi32>
    %5 = tosa.cast %4 : (tensor<1x1x2xi32>) -> tensor<1x1x2xf32>
    %cast = tensor.cast %5 : tensor<1x1x2xf32> to tensor<*xf32>
    call @printMemrefF32(%cast) : (tensor<*xf32>) -> ()
 return
  }
}

``` 

 ### 2. **Command to Run without `-linalg-fuse-elementwise-ops`:**

``` 
/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt a.mlir -pass-pipeline="builtin.module(func.func(tosa-to-linalg))" | /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt  -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-parallel-loops   -convert-scf-to-cf -convert-arith-to-llvm  -expand-strided-metadata -finalize-memref-to-llvm -convert-func-to-llvm  -reconcile-unrealized-casts | /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void --shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so --shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so --shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_async_runtime.so

``` 

### 3. **Output without `-linalg-fuse-elementwise-ops`:**:

``` 
[[[0, 0]]]

``` 

### 4. **Command to Run With `-linalg-fuse-elementwise-ops`:**


``` 
/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt  a.mlir -pass-pipeline="builtin.module(func.func(tosa-to-linalg))" | /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt  -tosa-to-arith --linalg-fuse-elementwise-ops -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-parallel-loops   -convert-scf-to-cf -convert-arith-to-llvm  -expand-strided-metadata -finalize-memref-to-llvm -convert-func-to-llvm  -reconcile-unrealized-casts | /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void --shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so --shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so --shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_async_runtime.so
``` 

### 5. **Output with `-linalg-fuse-elementwise-ops`:**

``` 
[[[1.63388e+09,    109440]]]
``` 

</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJzsWNtu27oS_Rr6haBAkZJlPfjBiWsgQItz0PbgPAaUNLK5QZECL67Tr98g5UvSOE13LxsoUEBQIntmrTUjcpG0cE5uNcASlTeoXM9E8Dtjl28GqR7EAWaN6R6WW-nxHqyTRiO-wqLrFzXrqoJ3iK4QXbkH52GIX6E5_V8TtA84X2S0yOb47ccPaE6nOMQ4YhyvwbVWjn6CQ3R1h8WA4TCClaBbqbdYaCx1a7STzoP22IILyuNPO9AYDtAGH6P8DrATA-B3b-_e49GarRUD_iT9DgvdpX9M8DiqIkpqobakDw4IKBhA-0_SATGjQ3OaIbr6uANsgh-Dd1hYwEI581RFordB60g-BOXlqAB7OYCbWKOg18meNOODh9Fhb_B7GK3pQgtTTx4Hxbg8w4itEFulWv97rBWxhcgGJS1i9fT1lH38ML4RuoqU6YoPg-mCAoyqG0RXGPdBt1m84dHKvfCAUUFHK7V_B4OF_o4zxBYetDMW8VvEVgfJGeJvIuG3AGyeA_QvAKCCDkJqxBaI1WeFGCNWUoz4GiPGvHEii6_EI8ZOkfwWVTd7oQKksA60A8Rv83mRI_4GxzacBeSHvDywUxHV-hRwxCLx-YXgs5r8e9TwRV7_EjUs0SQponGnbiWMF1K_FZlfkJXZylaoeyu3O3_vdrL3iRux21NHvsp4-xLZ61oQK4uLEAtxjty7MBwVVjfiIF0KmFTEzGr9HS240oHyQtwK509aniG_Ctx_ATyhRewUd0Yvn42Pc3I0iWvzaMJshVJXp96RbBqaT4U_mYyT7GngRUwLPlid0OPApKvj_Qs_md7QyaXYyaVuzTBED47GFvTZir_BHPnqaGRXuNimE14gtnGfHxDbRCtEbKPUfiAWFAgHp8fRmr-g9YhtmiBVF_9GZ9lEWyRm9HhySExG4RwZ5QhKakB8jRiLGV7qbLJKxBZni4rNM04Qb45VJNutEWMYVXEi_Dx9mJyYhI1LCzEaiNsZT5rQ92Dl57Pa4yOJCuOiShoTdCesBBeVkdboPVh_arw3ZBRWKAWKKGNGh_ElxrV9DGj7y0eJP5Ws9gPGBA6j0B1x3soOOjKAF7FqTPqIH4UMafidU85IUeAjIAut0a1UQIK2kFI7Eseq-6ndbMdA4qINFhPAcYXBBLS3D2Q0Unsy7S0QX--N7DAhbicsdETJxqUO_4AKJZvpHoXct_eTjPvgpXKZM7-S7F-kEu5Bp9riTihz5rpLXEyCn0ziP2m39V3mwK_6Q9zDxovGBYeicn28XhFUvOBa_48T759b1i90rd_Ptr7Wuz-m9sfUfidTe9FAyiuO9qN7nZOX5dmc88UCELuhdfQ1jHFO66J46m-Pc2fdknc1r8UMlnnFeUGLoqpnu-W878oaaCn4vGtFU_R0Xjd0Xs-7fNHkQGdyySgrckbLnOcFn2eLsmJ9VQBvqgrmbYUKCoOQKotNzIzdzqRzAZZ5vqjqaqZEA8qlHxMYm86lDJXrmV2mpjdh61BBlXTeXRC89Cr9AJESyjW-e3zoNseWPj36f_3E_3rrZ8Gq5c770aWXsEFss5V-F5qsNcNxjDwfKqlWh9jmWO5-yf4OAAD__xC-YU0">