<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/140259>140259</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [mlir] Inconsistent output when executing MLIR program with `--scf-parallel-loop-fusion`
        </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: ba631508ae7f

system: `Ubuntu 18.04.6 LTS`

## Description:
I am experiencing an inconsistent result when executing the same MLIR program with and without `--scf-parallel-loop-fusion`. 

## 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() {
    %idx0 = index.constant 0
    %1 = "tosa.const"() <{values = dense<-5857> : tensor<1x4x4xi32>}> : () -> tensor<1x4x4xi32>
    %10 = tosa.clz %1 : (tensor<1x4x4xi32>) -> tensor<1x4x4xi32>
    %28 = tosa.cast %10 : (tensor<1x4x4xi32>) -> tensor<1x4x4xf32>
    %48 = tosa.floor %28 : (tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
    %55 = tosa.reverse %48 {axis = 1 : i32} : (tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
    %cast_5 = tensor.cast %55 : tensor<1x4x4xf32> to tensor<*xf32>
    call @printMemrefF32(%cast_5) : (tensor<*xf32>) -> ()
    return
  }
}

``` 

 ### 2. **Command to Run without  `--scf-parallel-loop-fusion` :**

``` 
/pathto/mlir-opt a.mlir -tosa-to-scf  -tosa-to-arith   \
| /pathto/mlir-opt -pass-pipeline="builtin.module(func.func(tosa-to-linalg))"  \
| /pathto/mlir-opt  -convert-scf-to-cf   --cse \
-one-shot-bufferize="bufferize-function-boundaries" -finalize-memref-to-llvm  -convert-linalg-to-parallel-loops  \
 -convert-scf-to-cf -finalize-memref-to-llvm    -convert-math-to-llvm -convert-arith-to-llvm  \
-convert-index-to-llvm   -convert-cf-to-llvm -convert-func-to-llvm  -reconcile-unrealized-casts \
| timeout 10 /pathto/mlir-cpu-runner -e main -entry-point-result=void \
--shared-libs=/pathto/libmlir_c_runner_utils.so \
--shared-libs=/pathto/libmlir_runner_utils.so \
--shared-libs=/pathto/libmlir_async_runtime.so

``` 

### 3. **Output without  `--scf-parallel-loop-fusion` :**:

``` 
Unranked Memref base@ = 0x55a554495b40 rank = 3 offset = 0 sizes = [1, 4, 4] strides = [16, 4, 1] data = 
[[[0,    0,    0,    0], 
  [0,    0,    0,    0], 
  [0,    0, 0,    0], 
  [0,    0,    0,    0]]]

``` 

### 4. **Command to Run with `--scf-parallel-loop-fusion` :**


``` 
/pathto/mlir-opt a.mlir -tosa-to-scf  -tosa-to-arith   \
| /pathto/mlir-opt -pass-pipeline="builtin.module(func.func(tosa-to-linalg))"  \
| /pathto/mlir-opt  -convert-scf-to-cf   --cse \
-one-shot-bufferize="bufferize-function-boundaries" -finalize-memref-to-llvm  -convert-linalg-to-parallel-loops \
--scf-parallel-loop-fusion -convert-scf-to-cf -finalize-memref-to-llvm -convert-math-to-llvm  -convert-arith-to-llvm  \
-convert-index-to-llvm -convert-cf-to-llvm   -convert-func-to-llvm  -reconcile-unrealized-casts \
| timeout 10 /pathto/mlir-cpu-runner -e main -entry-point-result=void \
--shared-libs=/pathto/libmlir_c_runner_utils.so \
--shared-libs=/pathto/libmlir_runner_utils.so \
--shared-libs=/pathto/libmlir_async_runtime.so
``` 

### 5. **Output with `--scf-parallel-loop-fusion` :**

``` 
Unranked Memref base@ = 0x5625e8d1c780 rank = 3 offset = 0 sizes = [1, 4, 4] strides = [16, 4, 1] data = 
[[[2.68938e+20,    7.20647e+31,    2.89575e+32,    2.48354e+27], 
 [5.12554e-11,    2.63753e+23,    7.00624e+22,    5.61567e+13], 
  [0,    0, 0,    0], 
  [0,    0,    0,    0]]]
``` 

### 6. **Analysis for this case :**

This issue can be consistently reproduced by enabling the `--scf-parallel-loop-fusion` option, which fuses the last two `scf.parallel` loops.

I'm not sure whether the root cause lies in the `--scf-parallel-loop-fusion` transformation itself or in the subsequent `-convert-cf-to-llvm` lowering.

For reference:

 
[input.txt](https://github.com/user-attachments/files/20254061/input.txt)

contains the IR before applying `--scf-parallel-loop-fusion`.

[output.txt](https://github.com/user-attachments/files/20254107/output.txt)

 contains the IR after applying `--scf-parallel-loop-fusion`.


</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJzsWEFv6zYS_jX0ZSBBokTJPvjgxM9AgBa7eG3PD5Q0irlLkVqSSuz36xckLdt5cZom2-2pgCFbnOE3H2eGnyxya8WjQlwTdkfYdsEnt9dm_WUQ8sgPuGh0d1w_CgdPaKzQihQbaHhV5Cxbcqx7km1ItrFH63DwNlJlvzWTchPkyzQr0wp--vUXUmXRj9CC0AK2aFsjRhfxSLZ5AD4AHkY0AlUr1CNwBUK1WllhHSoHBu0kHTzvUQEesJ2c93J7BMsHhJ9_evgKo9GPhg_wLNweuOrCDz05TypJbNsnIzdcSpSJ1HpM-imsqMpSeEHvF4ejBafhK45Gd1OLkeXZwzvlKRC6IXQTQv_zFJrQJU8HKQyhq2iOU0-DPkHZxqcjfPzNoLtJIpD6jmQb6CfVpv4CoxFP3CGQMhuNUO5nHAz2DwUldOlQWW1IcU_o5iAKSoovPl62gXcBdq8B-jcASJkNXChCl4SuZoIAQCgT3SEDUmxBqA4Pqa-T48pBdnHJg51Q6rTl0YNQOoMV96S-e-JyQhv8OlQWSXGfsCWrSfEFfKrOLPNDeSjnhdbb2X4CS_z9bd8Lm0g3cpHfZ4IB4_bUPwhMl1fA3LpzsI9D9z9Al1fQvdTanOPdwu4_hM3YBdug39s4h6zv-EHEqsQUedb19k8J6zP07RQ6eJ5zFgi9Knmc77firY6NsC2X8maTn-PFjnvJ_kXjR-6xnU6oBt1kVLjxHZdtTtcfdq93uGgCnTXhXg-DFyCvIZM669C7QuRpnmTjRiy6G7nbO03ozqtJokcHUVgg8ZVMnPbgcLnjxkshAGH3YQn3cBMkGbm1yShGlEIhKbaE0mYS0gmVRn0idHkWBp_GE74UisvHoHUrQum7gSBptXpC40ISnE48W0iS1rdfmJpohYnda5c0U9-jEd_PfE63iefgHx1JoyfVcSPQ-thJ78l4hyE0QeAnn4aroJGuN7zIv51536L3NuwV8MDd_mw4j4bsX_xPC5ytQTqv0M6Gtn8N5dd8tSCDrVatkJhMymCg1yW-2e1VBZwY0Hedl6Mfi9GOU2ImpdBAguBlHhJUzhyTUQvlkvi4JcX2SYtuZp7YPTfYJVI0NhTlAipF43G_td8i6rfJCWlTqz8w9_MzuT2qENkvObX69ka97NNi3qf_mNw4uU_tz-LWFv1NGa7-jR1EHYKGWyRlfPZkB8Y4Y2W5Yk2ZgXcM4wXovrfoohNY8f30TCTsLif0Hsp4YVuwzojuylqdzbk3d9zxaMs24e-c_2TeDACvv9nW_4wS90m_T4HFzzslKn9HSj8uo38r6V-lpOdN-0Z1PiSxt_X1cwJ7S17hb4H9vMC-uXXZDXX93_75vCOrFWW47PK2Xv51skrTarkqlkjoHZ3VrU5pVpW1Hyvy0xhNlytWszBGz2PlsmBlmFtfKSdhdyzNKWMlJvkFoCpqVgTn4hwoyyoaAWZQllY5q0LwvPj_afabVa_mqm8Ul0crLPTagNsLCy33ovSyvr96g7B2Qmi5ggbh8rIvj2DmN-8OmiOg4o2c3_ff7SQdDxboPTzvRbuHfrJow1TpXzbcs_YYtu3TGcHPCvKVRnIPhNYDKO3ATgbheY9ujyZAGK0dtHyyCFKgBaH-GClnuLK9NgP35EA4i7IHbWYAOzUW_zOhiqcVr8UqcnxGI9TjieZOGzDYo0F1dURxalOhxsml7uBCmZd750YbarAjdPco3H5q0lYPhO4miybhzvF2P6ByltBdL6R_FOxoRlmZVTmhuwteeEci2abVynGhYmofvkKDvTYIfBzl0Rfr3VOXUwuxOx2E4k8hm2c1obsrwJkt_EiX9w7Nh9kuunXRrYoVX-A6r8u6zKpqVS726yVtmoJh1dVlX1Y1r4uio3XN27pqs6JrF2LtCWYsr_KyLFmZ1nmzLFnb87Jry2XbkjLDgQuZ-nKn2jwuwv5Y52VG2WoheYPShnM6SuMREyVsuzBr758006MlZSaFdfaC4IST4WwvTGBbeLg-VNMngX55qPb6LO293CwmI9e_U7TQv_ErGY3-F7bOd5RfnS_caYFPa_rfAAAA__-urivF">