[MLIR] Inconsistent output with `-convert-arith-to-llvm`

[sweead]

test commit: 98e8f01

Description:

Since the shift amount is 0, the tosa.arithmetic_right_shift operation does not change the value. However, when using the -convert-arith-to-llvm pass, I am encountering inconsistent results.

Steps to Reproduce:

Minimal MLIR program (test.mlir):

module {
  func.func private @printMemrefI64(tensor<*xi64>)
  func.func @main() {
    %cst0 = "tosa.const"() {values = dense<125> : tensor<i64>} : () -> tensor<i64>
    %cst1 = "tosa.const"() {values = dense<0> : tensor<i64>} : () -> tensor<i64>  // Modified to ensure a non-negative shift amount
    %2 = tosa.arithmetic_right_shift %cst0, %cst1 {round = true} : (tensor<i64>, tensor<i64>) -> tensor<i64>
    %cast = tensor.cast %2 : tensor<i64> to tensor<*xi64>
    call @printMemrefI64(%cast) : (tensor<*xi64>) -> ()
    return
  }
}

1. Without -convert-arith-to-llvm:

Command:

/home/workdir/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline="builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))" |\
/home/workdir/llvm-project-latest/build/bin/mlir-opt -linalg-fuse-elementwise-ops  -one-shot-bufferize="bufferize-function-boundaries=1"   -convert-linalg-to-loops  -convert-scf-to-cf  -gpu-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:

[125]

2. With -convert-arith-to-llvm:

/home/workdir/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline="builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))" |\
/home/workdir/llvm-project-latest/build/bin/mlir-opt -convert-arith-to-llvm -linalg-fuse-elementwise-ops  -one-shot-bufferize="bufferize-function-boundaries=1"   -convert-linalg-to-loops  -convert-scf-to-cf  -gpu-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:

The results are inconsistent. Possible output include:

[94195214772498]

[0]

[10]

[llvmbot]

@llvm/issue-subscribers-mlir

Author: Anonymous_reporter (sweead)

test commit: [98e8f01](https://github.com/llvm/llvm-project/commit/98e8f01d183177a4f54187c23183da50a7cf6daf)

Description:

Since the shift amount is 0, the tosa.arithmetic_right_shift operation does not change the value. However, when using the -convert-arith-to-llvm pass, I am encountering inconsistent results.

Steps to Reproduce:

Minimal MLIR program (test.mlir):

module {
  func.func private @<!-- -->printMemrefI64(tensor&lt;*xi64&gt;)
  func.func @<!-- -->main() {
    %cst0 = "tosa.const"() {values = dense&lt;125&gt; : tensor&lt;i64&gt;} : () -&gt; tensor&lt;i64&gt;
    %cst1 = "tosa.const"() {values = dense&lt;0&gt; : tensor&lt;i64&gt;} : () -&gt; tensor&lt;i64&gt;  // Modified to ensure a non-negative shift amount
    %2 = tosa.arithmetic_right_shift %cst0, %cst1 {round = true} : (tensor&lt;i64&gt;, tensor&lt;i64&gt;) -&gt; tensor&lt;i64&gt;
    %cast = tensor.cast %2 : tensor&lt;i64&gt; to tensor&lt;*xi64&gt;
    call @<!-- -->printMemrefI64(%cast) : (tensor&lt;*xi64&gt;) -&gt; ()
    return
  }
}

1. Without -convert-arith-to-llvm:

Command:

/home/workdir/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline="builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))" |\
/home/workdir/llvm-project-latest/build/bin/mlir-opt -linalg-fuse-elementwise-ops  -one-shot-bufferize="bufferize-function-boundaries=1"   -convert-linalg-to-loops  -convert-scf-to-cf  -gpu-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:

[125]

2. With -convert-arith-to-llvm:

/home/workdir/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline="builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))" |\
/home/workdir/llvm-project-latest/build/bin/mlir-opt -convert-arith-to-llvm -linalg-fuse-elementwise-ops  -one-shot-bufferize="bufferize-function-boundaries=1"   -convert-linalg-to-loops  -convert-scf-to-cf  -gpu-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:

The results are inconsistent. Possible output include:

[94195214772498]

[0]

[10]

[sweead]

@CoTinker Hi, I previously reported a similar issue #148812. It seems there are still some problems here. Could you help me take a look?

[CoTinker]

I think you use the wrong pipeline, for that we should not put the convert-arith-to-llvm pass before bufferization.

[sweead]

I think you use the wrong pipeline, for that we should not put the convert-arith-to-llvm pass before bufferization.

I got it, thank you. I encountered this issue while testing on MLIR, and users might run into the same problem as well. Is it possible to improve the documentation or implement other measures to help prevent this situation?

[joker-eph]

I think you use the wrong pipeline, for that we should not put the convert-arith-to-llvm pass before bufferization.

How does it explain a numerical difference though?