[Bug ipa/114985] [15 regression] internal compiler error: in discriminator_fail during stage2

["jamborm at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114985

--- Comment #20 from Martin Jambor <jamborm at gcc dot gnu.org> ---
The IL we generate the jump function from is:
  <bb 2>
  _1 = cclauses_2(D) != 0B;
  c_parser_omp_all_clauses (_1);

Which translates to the expected jump function:
  callsite  void c_parser_omp_teams(int**)/3 -> int*
c_parser_omp_all_clauses(bool)/1 :
     param 0: PASS THROUGH: 0, op ne_expr 0B

so IPA looks like it's doing what it should.

(In reply to Aldy Hernandez from comment #6)
> I wonder if something like this would work.
> 
> diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
> index 5781f50..ea8a685 100644
> --- a/gcc/ipa-cp.cc
> +++ b/gcc/ipa-cp.cc
> @@ -1730,6 +1730,8 @@ ipa_value_range_from_jfunc (vrange &vr,
>         }
>        else
>         {
> +         if (TREE_CODE_CLASS (operation) == tcc_comparison)
> +           vr_type = boolean_type_node;
>           Value_Range op_res (vr_type);
>           Value_Range res (vr_type);
>           tree op = ipa_get_jf_pass_through_operand (jfunc);

This looks OKish and we also do a similar thing in
ipa_get_jf_arith_result.

Also note that the ipa_value_range_from_jfunc already has a parameter
that tells it what type the result should be.  It is called parm_type,
which is boolean_type in the case that ICEs.  So we can even bail out
if we really encounter jump function created from bad IL.

I was thinking of using use parm_type from the beginning, to
initialize op_res with it, but there are jump functions representing
an operation followed by a truncation, for example for:

  _2 = complain_6(D) & 1;
  _3 = (int) std_alignof_7(D);
  cxx_sizeof_or_alignof_type (_3, _2);

where _r is in fact bool (has smaller size and precision) and trying
to make ranger do the bit_and_expr directly to bool leads to a failed
assert in fold_range (the test of m_operator->operand_check_p).

So doing the operation in the original type - unless it is a
comparison - and then using ipa_vr_operation_and_type_effects seems to
be the right thing to do.

But I am really curious why propagate_vr_across_jump_function does not
need the same check for tcc_comparison operators and generally why is
it so different (in the non-scc case)?  Why is ipa_supports_p (this
predicate has a really really really bad name BTW and I am completely
at loss as to what it does and how or why) used there and not in
ipa_value_range_from_jfunc?

(I also cannot prevent myself from ranting a little that it would
really help if all the ranger (helper) classes and functions were
better documented.)