[erlc] internal consistency check: unassigned / bad_typed_register

[RobinMorisset]

Describe the bug
Running erlc on the following code:

f() ->
    <<
        0
     || register(
            _V5 =
                ((_V5 = is_list(
                    try [0 || <<_>> <= <<>>] of
                        _ ->
                            "a"
                    catch
                        _ ->
                            ok
                    end
                )) andalso _V5),
            ok
        ),
        _V5
    >>.

results in the following error message:

unassigned:1: function f/0+29:
  Internal consistency check failed - please report this bug.
  Instruction: {test,is_eq_exact,
                     {f,14},
                     [{tr,{x,0},{t_atom,[false,true]}},
                      {tr,{y,1},{t_atom,[false,true]}}]}
  Error:       {unassigned,{y,1}}:

It is possible to minimize the testcase further, but it changes the error message. The following:

f() ->
    _V5 =
        ((_V5 = is_list(
            try now() of
                _ ->
                    "a"
            catch
                _ ->
                    ok
            end
        )) andalso _V5).

leads to the following error:

test961435:1: function f/0+23:
  Internal consistency check failed - please report this bug.
  Instruction: {test,is_eq_exact,
                     {f,14},
                     [{tr,{x,1},{t_atom,[false,true]}},
                      {tr,{x,0},{t_atom,[false,true]}}]}
  Error:       {bad_typed_register,
                   {t_union,
                       {t_atom,[ok]},
                       {t_cons,{t_integer,{52,120}},nil},
                       none,none,none},
                   {t_atom,[false,true]}}:

Affected versions
Both testcases fail with the following:

• master
• master + beam_ssa_type: Fix bugs relating to subtraction in type inference #7489

Only the top testcase (the one with the "unassigned" error message) fails with an old version of OTP 25.

Additional context
This might be a duplicate of #7504 as the two seem quite similar, despite the error message being slightly different.

[bjorng]

Yes, it is the same bug.