[MLIR] Inconsistent output when executing MLIR program with and without `-tosa-layerwise-constant-fold`

[Lambor24]

My git version is 55fdecc.

Description:

I am experiencing an inconsistent result when executing the same MLIR program with and without the -tosa-layerwise-constant-fold.

Steps to Reproduce:

1. MLIR Program (test.mlir):

test.mlir:

module {
  func.func private @printMemrefI16(%ptr : tensor<*xi16>) attributes { llvm.emit_c_interface }
  func.func @main() -> () {
    %arg0 = "tosa.const"() {value = dense<0> : tensor<1x1xi64>} : () -> tensor<1x1xi64>
    %5 = tosa.equal %arg0, %arg0 : (tensor<1x1xi64>, tensor<1x1xi64>) -> tensor<1x1xi1>
    %6 = tosa.cast %5 : (tensor<1x1xi1>) -> tensor<1x1xi16>
    %rtn1 = tensor.cast %6 : tensor<1x1xi16> to tensor<*xi16>
    call @printMemrefI16(%rtn1) : (tensor<*xi16>) -> ()
    return
  }
}

2. Command to Run Without -tosa-layerwise-constant-fold:

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

3. Output Without -tosa-layerwise-constant-fold:

[[1]]

4. Command to Run With -tosa-layerwise-constant-fold:

/path/llvm-project/build/bin/mlir-opt test.mlir -tosa-layerwise-constant-fold | \
/path/llvm-project/build/bin/mlir-opt -pass-pipeline='builtin.module(func.func(tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-parallel-loops -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-runner -e main -entry-point-result=void \
-shared-libs=/path/llvm-project/build/lib/libmlir_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_c_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_async_runtime.so

5. Output With -tosa-layerwise-constant-fold:

[[-1]]

I'm not sure if there is any bug in my program or if the wrong usage of the above passes caused this result.

[llvmbot]

@llvm/issue-subscribers-mlir

Author: None (Lambor24)

My git version is [55fdecc](https://github.com/llvm/llvm-project/commit/55fdeccc4567bcd4e3f8df0d177195880a194a6a).

Description:

I am experiencing an inconsistent result when executing the same MLIR program with and without the -tosa-layerwise-constant-fold.

Steps to Reproduce:

1. MLIR Program (test.mlir):

test.mlir:

module {
  func.func private @<!-- -->printMemrefI16(%ptr : tensor&lt;*xi16&gt;) attributes { llvm.emit_c_interface }
  func.func @<!-- -->main() -&gt; () {
    %arg0 = "tosa.const"() {value = dense&lt;0&gt; : tensor&lt;1x1xi64&gt;} : () -&gt; tensor&lt;1x1xi64&gt;
    %5 = tosa.equal %arg0, %arg0 : (tensor&lt;1x1xi64&gt;, tensor&lt;1x1xi64&gt;) -&gt; tensor&lt;1x1xi1&gt;
    %6 = tosa.cast %5 : (tensor&lt;1x1xi1&gt;) -&gt; tensor&lt;1x1xi16&gt;
    %rtn1 = tensor.cast %6 : tensor&lt;1x1xi16&gt; to tensor&lt;*xi16&gt;
    call @<!-- -->printMemrefI16(%rtn1) : (tensor&lt;*xi16&gt;) -&gt; ()
    return
  }
}

2. Command to Run Without -tosa-layerwise-constant-fold:

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

3. Output Without -tosa-layerwise-constant-fold:

[[1]]

4. Command to Run With -tosa-layerwise-constant-fold:

/path/llvm-project/build/bin/mlir-opt test.mlir -tosa-layerwise-constant-fold | \
/path/llvm-project/build/bin/mlir-opt -pass-pipeline='builtin.module(func.func(tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-parallel-loops -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-runner -e main -entry-point-result=void \
-shared-libs=/path/llvm-project/build/lib/libmlir_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_c_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_async_runtime.so

5. Output With -tosa-layerwise-constant-fold:

[[-1]]

I'm not sure if there is any bug in my program or if the wrong usage of the above passes caused this result.

[CoTinker]

Fixed by #147078.