[MLIR] Inconsistent output when executing MLIR program with and without `-convert-math-to-llvm`

[sweead]

test commit: 87e39c3

Description:

I am experiencing an inconsistent result when executing the same MLIR program with and without the -convert-arith-to-llvm.

Steps to Reproduce:

Minimal MLIR program (test.mlir):

#map = affine_map<(d0, d1) -> (d0, d1)>
module {
  func.func @main() -> () {
    %0 = "tosa.const"() <{values = dense<-170> : tensor<3x4xi64>}> : () -> tensor<3x4xi64>
    %1 = "tosa.const"() <{values = dense<0> : tensor<1x4xi64>}> : () -> tensor<1x4xi64>
    %2 = tosa.arithmetic_right_shift %0, %1 {round = true} : (tensor<3x4xi64>, tensor<1x4xi64>) -> tensor<3x4xi64>
    %cast = tensor.cast %2 : tensor<3x4xi64> to tensor<*xi64>
    call @printMemrefI64(%cast) : (tensor<*xi64>) -> ()
    return
  }
  func.func private @printMemrefI64(tensor<*xi64>)
}

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))"| \
/home/workdir/llvm-project/build/bin/mlir-opt   -sparsifier | \
/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:

[[-170,   -170,   -170,   -170], 
 [-170,   -170,   -170,   -170], 
 [-170,   -170,   -170,   -170]]

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))"| \
/home/workdir/llvm-project/build/bin/mlir-opt -convert-arith-to-llvm  -sparsifier | \
/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:

[[93942316279653,   0,   0,   0], 
 [0,   0,   0,   0], 
 [0,   0,   0,   0]]

[[94258813645071,   0,   140648166016064,   140648166015632], 
 [2,   0,   0,   0], 
 [3,   0,   0,   0]]

[[0,   0,   0,   0], 
 [0,   0,   0,   0], 
 [0,   0,   0,   0]]

[llvmbot]

@llvm/issue-subscribers-mlir

Author: None (sweead)

test commit: [87e39c3](https://github.com/llvm/llvm-project/commit/87e39c399c96128307a826fd9168679f7c8c8d10)

Description:

I am experiencing an inconsistent result when executing the same MLIR program with and without the -convert-arith-to-llvm.

Steps to Reproduce:

Minimal MLIR program (test.mlir):

#map = affine_map<(d0, d1) -> (d0, d1)>
module {
  func.func @<!-- -->main() -&gt; () {
    %0 = "tosa.const"() &lt;{values = dense&lt;-170&gt; : tensor&lt;3x4xi64&gt;}&gt; : () -&gt; tensor&lt;3x4xi64&gt;
    %1 = "tosa.const"() &lt;{values = dense&lt;0&gt; : tensor&lt;1x4xi64&gt;}&gt; : () -&gt; tensor&lt;1x4xi64&gt;
    %2 = tosa.arithmetic_right_shift %0, %1 {round = true} : (tensor&lt;3x4xi64&gt;, tensor&lt;1x4xi64&gt;) -&gt; tensor&lt;3x4xi64&gt;
    %cast = tensor.cast %2 : tensor&lt;3x4xi64&gt; to tensor&lt;*xi64&gt;
    call @<!-- -->printMemrefI64(%cast) : (tensor&lt;*xi64&gt;) -&gt; ()
    return
  }
  func.func private @<!-- -->printMemrefI64(tensor&lt;*xi64&gt;)
}

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))"| \
/home/workdir/llvm-project/build/bin/mlir-opt   -sparsifier | \
/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:

[[-170,   -170,   -170,   -170], 
 [-170,   -170,   -170,   -170], 
 [-170,   -170,   -170,   -170]]

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))"| \
/home/workdir/llvm-project/build/bin/mlir-opt -convert-arith-to-llvm  -sparsifier | \
/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:

[[93942316279653,   0,   0,   0], 
 [0,   0,   0,   0], 
 [0,   0,   0,   0]]

[[94258813645071,   0,   140648166016064,   140648166015632], 
 [2,   0,   0,   0], 
 [3,   0,   0,   0]]

[[0,   0,   0,   0], 
 [0,   0,   0,   0], 
 [0,   0,   0,   0]]