[mlir] Inconsistent behavior when executing MLIR program with and without --scf-for-to-while and --test-scf-while-op-builder optimizations

[Emilyaxe]

git version: 0a44b24

system: Ubuntu 18.04.6 LTS

Description:

I am experiencing an inconsistent result when executing the same MLIR program with and without the --scf-for-to-while and --test-scf-while-op-builder optimizations. When these optimizations are applied, the program enters a dead loop.

Steps to Reproduce:

1. MLIR Program (a.mlir):

module {
  func.func private @printMemrefI32(tensor<*xi32>)
  func.func private @printMemrefF32(tensor<*xf32>)
  func.func @main() {
    %2 = "tosa.const"() <{value = dense<42> : tensor<1x3x7xi32>}> : () -> tensor<1x3x7xi32>
    %4 = "tosa.const"() <{value = dense<35> : tensor<1x3x7xi32>}> : () -> tensor<1x3x7xi32>
    %12 = tosa.minimum %2, %4 : (tensor<1x3x7xi32>, tensor<1x3x7xi32>) -> tensor<1x3x7xi32>
    %cast = tensor.cast %12 : tensor<1x3x7xi32> to tensor<*xi32>
    call @printMemrefI32(%cast) : (tensor<*xi32>) -> ()
    return
  }
}

2. Command to Run Without Optimizations::

/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt a.mlir  -pass-pipeline="builtin.module(func.func(tosa-to-linalg))"  \
| /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt  -tosa-to-arith    -convert-linalg-to-parallel-loops  \
-one-shot-bufferize="bufferize-function-boundaries"  -convert-linalg-to-loops  -finalize-memref-to-llvm \
-convert-arith-to-llvm   -convert-scf-to-cf  -convert-func-to-llvm  -reconcile-unrealized-casts  \
| timeout 10  /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so 

3. Output WithOut Optimizations::

Unranked Memref base@ = 0x561bcc0853c0 rank = 3 offset = 0 sizes = [1, 3, 7] strides = [21, 7, 1] data = 
[[[35,    35,    35,    35,    35,    35,    35], 
  [35,    35,    35,    35,    35,    35,    35], 
  [35,    35,    35,    35,    35,    35,    35]]]

4. Command to Run With --scf-for-to-while and --test-scf-while-op-builder Optimizations::

/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt a.mlir  \
-pass-pipeline="builtin.module(func.func(tosa-to-linalg))" | /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt  \
-tosa-to-arith    -convert-linalg-to-parallel-loops   -one-shot-bufferize="bufferize-function-boundaries"  \
-convert-linalg-to-loops    --scf-for-to-while   --test-scf-while-op-builder -finalize-memref-to-llvm  \
-convert-arith-to-llvm   -convert-scf-to-cf  -convert-func-to-llvm  -reconcile-unrealized-casts \
| timeout 10  /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so

5. Behavior with Optimizations::

When I apply the --scf-for-to-while and --test-scf-while-op-builder optimizations, the program enters a dead loop and doesn't produce any output.

[Emilyaxe]

These are the results after running with --scf-for-to-while and --scf-for-to-while --test-scf-while-op-builder respectively.

I suspect that the dead loop might be caused by the three scf.condition(%true) loops. When I try to annotate these three loops as following, It can normally output the results

I don't really understand the semantics of pass --test-scf-while-op-builder. So I try to run the original regression test case for --test-scf-while-op-builder in mlir/test/Dialect/SCF/while-op-builder.mlir, and find it also can correctly output the results with the command

/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt b.mlir  -test-scf-while-op-builder -convert-arith-to-llvm  -convert-vector-to-llvm  -convert-scf-to-cf  -convert-func-to-llvm  -reconcile-unrealized-casts \
| timeout 10  /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so

b.mlir

func.func @main() {
  %c1 = arith.constant 1 : i32
  %c10 = arith.constant 5 : i32
  %0:2 = scf.while (%arg1 = %c10) : (i32) -> (i32, i32) {
    %1 = arith.cmpi slt, %arg1, %c10 : i32
    scf.condition(%1) %arg1, %c1 : i32, i32
  } do {
  ^bb0(%arg1: i32, %arg2: i32):
    %1 = arith.addi %arg1, %arg2 : i32
    scf.yield %1 : i32
  }
  vector.print %0#0 : i32
  return
}

Hi, @ingomueller-net I am sorry to disturb you, but could you please take a moment to review this problem? Thank you!

[ingomueller-net]

Why do you want to run -test-scf-while-op-builder? That pass (see implementation), like all -test-* passes, is meant for testing a specific feature of some internal API and isn't meant to produce any meaningful change to the input IR. In other words: I believe that the behavior that you see is expected and you are probably looking for a different pass.

[Emilyaxe]

@ingomueller-net,
Thank you for your quick reply. I am currently conducting research on testing MLIR infrastructure, and from a user's perspective, all passes provided by MLIR seem to serve the same purpose.

Additionally, I noticed that some developers believe the -test-pass should ensure correctness, as mentioned in #119876. Of course, if the -test-pass discussed in issue #119876 differs from the -test-scf-while-op-builder pass, I would be happy to close this issue.

Thank you again for your kind explanation.

[ingomueller-net]

OK, I see there are people who argue that even -test-* passes should preserve the semantics if the input IR. In this case, this pass would need changing. Maybe it should even be a unittest instead of a LIT test. I am happy to leave the issue open with the goal of tracking that change.