<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/148366>148366</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [MLIR] Inconsistent Results with `scf-for-loop-peeling`
        </td>
    </tr>

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

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

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

<pre>
    test commit: [f2e244f](https://github.com/llvm/llvm-project/commit/f2e244f443c99f814216bef37a82f1b8d3e5975d)

## Description:

I am experiencing an inconsistent result when running the same MLIR input through two pipelines that differ only in the peel-front flag of the scf-for-loop-peeling pass


## Steps to Reproduce:

### Minimal MLIR program (test.mlir):

```
module {
  func.func private @printMemrefI8(tensor<*xi8>) attributes {llvm.emit_c_interface}
  func.func @main() {
    %0 = "tosa.const"() <{values = dense<17> : tensor<2x3x4x2x1x3xi8>}> : () -> tensor<2x3x4x2x1x3xi8>
    %1 = "tosa.const"() <{values = dense<-2> : tensor<2x1x1x1x1x3xi8>}> : () -> tensor<2x1x1x1x1x3xi8>
    %3 = tosa.maximum %0, %1 : (tensor<2x3x4x2x1x3xi8>, tensor<2x1x1x1x1x3xi8>) -> tensor<2x3x4x2x1x3xi8>
    %cast_0 = tensor.cast %3 : tensor<2x3x4x2x1x3xi8> to tensor<*xi8>
    call @printMemrefI8(%cast_0) : (tensor<*xi8>) -> ()
    return
  }
}
```
### 1. Using -scf-for-loop-peeling=peel-front=`false`:

#### Command:
```
/home/workdir/llvm-project/build/bin/mlir-opt test.mlir --pass-pipeline="builtin.module(func.func(tosa-to-linalg))" | \
/home/workdir/llvm-project/build/bin/mlir-opt   -sparsification-and-bufferization  -tosa-to-arith -convert-linalg-to-loops  \
-scf-for-loop-peeling=peel-front=false -test-lower-to-llvm | \
/home/workdir/llvm-project/build/bin/mlir-runner -e main -entry-point-result=void -shared-libs=/home/workdir/llvm-project/build/lib/libmlir_runner_utils.so
```
#### Output:
```
[[[[[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]]], [[[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]]], [[[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]]]], [[[[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]]], [[[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]]], [[[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]]]]]

```
### 2. Using -scf-for-loop-peeling=peel-front=`true `:
```
/home/workdir/llvm-project/build/bin/mlir-opt test.mlir --pass-pipeline="builtin.module(func.func(tosa-to-linalg))" | \
/home/workdir/llvm-project/build/bin/mlir-opt   -sparsification-and-bufferization  -tosa-to-arith -convert-linalg-to-loops  \
-scf-for-loop-peeling=peel-front=true  -test-lower-to-llvm | \
/home/workdir/llvm-project/build/bin/mlir-runner -e main -entry-point-result=void -shared-libs=/home/workdir/llvm-project/build/lib/libmlir_runner_utils.so
```
#### Output:
```
[[[[[[, , &]], [[/, 0, V]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, �, �]]], [[[, L, 5]], [[V, , ]]]], [[[[, , ]], [[L, 5, V]]], [[[, , ]], [[�, l, L]]], [[[, , ]], [[, �, �]]], [[[, , ]], [[V, , ]]]], [[[[, , ]], [[L, 5, V]]], [[[, , ]], [[�, l, L]]], [[[, , ]], [[, �, �]]], [[[, , ]], [[V, , ]]]]], [[[[[, , ]], [[, , ]]], [[[, , P]], [[, l, L]]], [[[, V, ]], [[, �, �]]], [[[, L, 5]], [[, , ]]]], [[[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]]], [[[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]], [[[, , ]], [[, , ]]]]]]
```

It seems that the loop peeling process caused undefined values in memory.


</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJzsWFtv27gS_jX0y4CGROpiP_jBsWogQINzkIPT14CWRhZ3KVIgKcfZX7-gJDs3t5u2aFFsgxBSJJHffHPlmMI5udeIK5JekbSYid43xq7cPaKoZjtTPaw8Og-laVvpCV8DSa9qhixJapIWhC0a7ztH-JqwLWHbvfRNv5uXpiVsq9ThdKOdNX9g6QnbTlBsO8EkCS-Xy3oRJyzOdljzXCxYHe8WFcd0macVYUsSrcNgnDAOBbrSys5Lo4Pc4dM1iBbw2KGVqEup9yA0SF0a7aTzqD1YdL3ycN-gBttrHeb4BsGJFuHm4_UtSN31HnxjTb9vwN8b6GSHSmp04BvhoZJ1jRaMVg8g9bC6Q1S0tkZ7qJXYg6lH0LKmtbFUGdPRMCdI64RzkyJnXf7nsXPgDdxiZ03Vl3hWaZwRJt1ILVuhRpadNXsrWiBsETwzb5W0wULnZVk0jWjdmqpXCCS_ItEaoO51OQ8X6Kw8CI9AkqizUvsbbC3W14sBVDtjCd8Qtj7KBeEfCFuC8N7KXe_RBbTg0Tm20t-Vd1J7tLUokeTFCykkiVohNWGLAHFiAUBYGgHhBRDGvHFiHtzkCWOnmXxD8quDUH0QxwuoUDskfBPnhH-AEIVnluzIj8mRHeMjn9jmxWnSBEfD8xcWnEnF30SKsguk4unvzaReLjiT4oOwgVErjrLt28F-hG1OlNfw1G2v9WObLwn6GgOVwvm70XXj9Hl4c2L5Ja-EEL8QWSNyKZS6FIpniaP9n-v5LDwHDUbLTpgWfW_18DAG5nR9kh2PCRbP4f8upCi9lLiEF49pTnhBsqgWymFAeZWsAW5j2lboavr4TOK2MS0Str039s8qJO7L6rjrparCPeTNNiQ3NZ2Hc6oDpaGO0FNpCnwYC6u81PMx4QlbnJMwmMw4Qb2hSmqh9sFEYTAg-QZIuvkuXgDUdcI6WctShIpMha7org-FUv41vAGgJwbCSt8ALY0-oPUToYGaMZ2Dic2bnDB4AGiwC1XmHu0Aow7td6sVNge0QBFC8QKK2tsH2hmpPR03EcKLg5EVUNcIixVVcucGP7xZnJK78Rok3o0S73ovlZs787kwDaH1n953vb8QWWH7fj6GAhFMUQwbdfjvwvtnn37VVS8X_qo837W7qN0FBX9Vtu8e_A21G8er7vmx6rKv7Q-87REe-4P3FuBHtACDkd9bgDe0ACx7nQfb8DD8jPj00xLvw5ZcbclV8fzhszgfwyV9-fHT91SoCfIbtH7KXY3sfrgBLuG8q_9zOoz_Xlr2T5p_-rGh_xvuzf9u7c6dx9M6Hq2vPTjEdjp29A1C2AfhfJBoTYnOQSl6hxX0usJaaqxgOp2SGlpsjX2Yj3CzasWrJV-KGa7iPGWcJUnOZ81qV8Z5zbJFFSUiz5Y8YZGIOc_SEutFWS9ncsUilkZ5zGKW5CmfZ_EyyxfpMk_iLM6SJUkibIVU8-FI0Nj9TDrX4ypOFjzLZkrsULnhcJmx8aSSkbSY2dWwK-76vSNJpKTz7hHBS6-GA-mbj9e3JC3g-uk57u2w_zq4D30EyaKLnUIWzXqrVl99Nj2Qd4RtJ_6HFfs7AAD__2ys-gk">