[mlir][inlining] Infinite inlining loop

[vzakhari]

The MLIR inliner pass seems to go into an infinite loop on this test:

subroutine s3
call s1
call s2
end subroutine s3
subroutine s2
call s1
call s3
end subroutine s2
subroutine s1
call s2
call s3
end subroutine s1

Compile with: flang-new -O2 inline.f90 -mllvm -inline-all

The -mllvm -inline-all option just instructs the FIRInlinerInterface that all fir.call ops are legal to inline. It seems to be the MLIR inliner pass' job to prevent the infinite inlining.

[joker-eph]

Can you provide the MLIR IR and an opt command?

[vzakhari]

Here is a reproducer with just the func dialect operations:

module {
  func.func @_QPs3() {
    func.call @_QPs1() : () -> ()
    func.call @_QPs2() : () -> ()
    return
  }
  func.func @_QPs2() {
    func.call @_QPs1() : () -> ()
    func.call @_QPs3() : () -> ()
    return
  }
  func.func @_QPs1() {
    func.call @_QPs2() : () -> ()
    func.call @_QPs3() : () -> ()
    return
  }
}

Compile with: mlir-opt --inline repro.mlir

[MbjYjbpivj]

Can I perform inline to a specified depth?

[jeanPerier]

Here is another MLIR reproducer that goes into an infinite loop with mlir-opt -inline:

module {
  func.func @parent1(%arg0: i32) -> i32 {
    %0 = call @child(%arg0) : (i32) -> i32
    return %0 : i32
  }
  func.func @parent2(%arg0: i32) -> i32 {
    %0 = call @child(%arg0) : (i32) -> i32
    return %0 : i32
  }
  func.func @child(%arg0: i32) -> i32 {
    %c10_i32 = arith.constant 10 : i32
    %c1_i32 = arith.constant 1 : i32
    %0 = arith.cmpi sge, %arg0, %c10_i32 : i32
    %1 = scf.if %0 -> (i32) {
      scf.yield %arg0 : i32
    } else {
      %2 = arith.addi %arg0, %c1_i32 : i32
      %3 = func.call @parent1(%2) : (i32) -> i32
      %4 = func.call @parent2(%2) : (i32) -> i32
      %5 = arith.addi %3, %4 : i32
      scf.yield %5 : i32
    }
    return %1 : i32
  }
}

It is for the C-like input:

int child(int);
int parent1(int num) {return child(num);}
int parent2(int num) {return child(num);}
int child(int num) {
    if (num >= 10) return num;
    return parent1(num+1)+parent2(num+1);
}

It looks like this check is not working:

llvm-project/mlir/lib/Transforms/Utils/Inliner.cpp

Line 722 in 56ab56c

// Don't allow inlining if the target is an ancestor of the call. This