[MLIR] Canonicalisation of unreachable `arith.addi` hangs

[Sirraide]

Consider (https://godbolt.org/z/a4r55zdEd):

module @test {
  func.func @main() {
    cf.br ^bb4
  ^bb1(%2: i64):
    cf.br ^bb4
  ^bb3:
    %c1_i64_3 = arith.constant 1 : i64
    %9 = arith.addi %2, %c1_i64_3 : i64
    cf.br ^bb1(%9 : i64)
  ^bb4:
    return
  }
}

Attempting to canonicalise this with mlir-opt --canonicalize loops for ever; when I was running this in the debugger earlier before reducing it and dumped the add, it printed %7 = arith.addi %7, %c1_i64_3 : i64; so something doesn’t seem right here...

[llvmbot]

@llvm/issue-subscribers-mlir

Author: None (Sirraide)

Consider (https://godbolt.org/z/a4r55zdEd): ```mlir module @test { func.func @main() { cf.br ^bb4 ^bb1(%2: i64): cf.br ^bb4 ^bb3: %c1_i64_3 = arith.constant 1 : i64 %9 = arith.addi %2, %c1_i64_3 : i64 cf.br ^bb1(%9 : i64) ^bb4: return } } ``` Attempting to canonicalise this with `mlir-opt --canonicalize` loops for ever; when I was running this in the debugger earlier before reducing it and dumped the add, it printed `%7 = arith.addi %7, %c1_i64_3 : i64`; so something doesn’t seem right here...