[MLIR] RemoveDeadValues pass errors when run on `private` function

[Sirraide]

Consider (https://godbolt.org/z/W95n8jEvM):

func.func private @test(%arg0: i64) -> (i64) {
    %c0_i64 = arith.constant 0 : i64
    %2 = arith.cmpi eq, %arg0, %c0_i64 : i64
    cf.cond_br %2, ^bb1, ^bb2
  ^bb1:  // pred: ^bb0
    %c1_i64 = arith.constant 1 : i64
    return %c1_i64 : i64
  ^bb2:  // pred: ^bb0
    %c3_i64 = arith.constant 3 : i64
    return %c3_i64 : i64
}

Running mlir-opt --remove-dead-values results in the following error:

<source>:4:5: error: 'cf.cond_br' op expected 1 or more operands, but found 0
    cf.cond_br %2, ^bb1, ^bb2
    ^
<source>:4:5: note: see current operation: "cf.cond_br"()[^bb1, ^bb2] <{operandSegmentSizes = array<i32: 1, 0, 0>}> : () -> ()

The input I had before reduction instead causes it to complain about a ‘null operand’, but I presume the bug is the same (https://godbolt.org/z/cM14hjqjh):

func.func private @test(%arg0: i64) {
    %0 = "custom.slot"() <{alignment = 8 : i64, bytes = 8 : i64}> : () -> !llvm.ptr
    "custom.store"(%0, %arg0) <{alignment = 8 : i64}> : (!llvm.ptr, i64) -> ()
    %1 = "custom.load"(%0) <{alignment = 8 : i64}> : (!llvm.ptr) -> i64
    %c0_i64 = arith.constant 0 : i64
    %2 = arith.cmpi eq, %1, %c0_i64 : i64
    cf.cond_br %2, ^bb1, ^bb2
  ^bb1:  // pred: ^bb0
    %c1_i64 = arith.constant 1 : i64
    "custom.ret"(%c1_i64) : (i64) -> ()
  ^bb2:  // pred: ^bb0
    %c3_i64 = arith.constant 3 : i64
    "custom.ret"(%c3_i64) : (i64) -> ()
}

Error message:

<source>:7:5: error: null operand found
    cf.cond_br %2, ^bb1, ^bb2
    ^
<source>:7:5: note: see current operation: "cf.cond_br"(<<NULL VALUE>>)[^bb1, ^bb2] <{operandSegmentSizes = array<i32: 1, 0, 0>}> : (<<NULL TYPE>>) -> ()

In both cases, removing private causes the error to disappear.

[llvmbot]

@llvm/issue-subscribers-mlir

Author: None (Sirraide)

Consider (https://godbolt.org/z/W95n8jEvM): ```mlir func.func private @test(%arg0: i64) -> (i64) { %c0_i64 = arith.constant 0 : i64 %2 = arith.cmpi eq, %arg0, %c0_i64 : i64 cf.cond_br %2, ^bb1, ^bb2 ^bb1: // pred: ^bb0 %c1_i64 = arith.constant 1 : i64 return %c1_i64 : i64 ^bb2: // pred: ^bb0 %c3_i64 = arith.constant 3 : i64 return %c3_i64 : i64 } ```

Running mlir-opt --remove-dead-values results in the following error:

<source>:4:5: error: 'cf.cond_br' op expected 1 or more operands, but found 0
    cf.cond_br %2, ^bb1, ^bb2
    ^
<source>:4:5: note: see current operation: "cf.cond_br"()[^bb1, ^bb2] <{operandSegmentSizes = array<i32: 1, 0, 0>}> : () -> ()

The input I had before reduction instead causes it to complain about a ‘null operand’, but I presume the bug is the same (https://godbolt.org/z/cM14hjqjh):

func.func private @<!-- -->test(%arg0: i64) {
    %0 = "custom.slot"() &lt;{alignment = 8 : i64, bytes = 8 : i64}&gt; : () -&gt; !llvm.ptr
    "custom.store"(%0, %arg0) &lt;{alignment = 8 : i64}&gt; : (!llvm.ptr, i64) -&gt; ()
    %1 = "custom.load"(%0) &lt;{alignment = 8 : i64}&gt; : (!llvm.ptr) -&gt; i64
    %c0_i64 = arith.constant 0 : i64
    %2 = arith.cmpi eq, %1, %c0_i64 : i64
    cf.cond_br %2, ^bb1, ^bb2
  ^bb1:  // pred: ^bb0
    %c1_i64 = arith.constant 1 : i64
    "custom.ret"(%c1_i64) : (i64) -&gt; ()
  ^bb2:  // pred: ^bb0
    %c3_i64 = arith.constant 3 : i64
    "custom.ret"(%c3_i64) : (i64) -&gt; ()
}

Error message:

<source>:7:5: error: null operand found
    cf.cond_br %2, ^bb1, ^bb2
    ^
<source>:7:5: note: see current operation: "cf.cond_br"(<<NULL VALUE>>)[^bb1, ^bb2] <{operandSegmentSizes = array<i32: 1, 0, 0>}> : (<<NULL TYPE>>) -> ()

In both cases, removing private causes the error to disappear.