Re: [PATCH] c++: ICE with -fno-elide-constructors and trivial fn [PR101073]

[Jason Merrill <jason@redhat.com>]

On 2/15/23 13:37, Marek Polacek wrote:
> On Wed, Feb 15, 2023 at 02:39:16PM -0500, Jason Merrill wrote:
>> On 2/9/23 09:39, Marek Polacek wrote:
>>> In constexpr-nsdmi3.C, with -fno-elide-constructors, we don't elide
>>> the Y::Y(const Y&) call used to initialize o.c.  So store_init_value
>>> -> cxx_constant_init must constexpr-evaluate the call to Y::Y(const Y&)
>>> in cxx_eval_call_expression.  It's a trivial function, so we do the
>>> "Shortcut trivial constructor/op=" code and rather than evaluating
>>> the function, we just create an assignment
>>>
>>>     o.c = *(const struct Y &) (const struct Y *) &(&<PLACEHOLDER_EXPR struct X>)->b
>>>
>>> which is a MODIFY_EXPR, so the preeval code in cxx_eval_store_expression
>>> clears .ctor and .object, therefore we can't replace the PLACEHOLDER_EXPR
>>> whereupon we crash at
>>>
>>>         /* A placeholder without a referent.  We can get here when
>>>            checking whether NSDMIs are noexcept, or in massage_init_elt;
>>>            just say it's non-constant for now.  */
>>>         gcc_assert (ctx->quiet);
>>>
>>> The PLACEHOLDER_EXPR can also be on the LHS as in constexpr-nsdmi10.C.
>>> I don't think we can do much here, but I noticed that the whole
>>> trivial_fn_p (fun) block is only entered when -fno-elide-constructors.
>>> This is true since GCC 9; it wasn't easy to bisect what changes made it
>>> so, but r240845 is probably one of them.  -fno-elide-constructors is an
>>> option for experiments only so it's not clear to me why we'd still want
>>> to shortcut trivial constructor/op=.  I propose to remove the code and
>>> add a checking assert to make sure we're not getting a trivial_fn_p
>>> unless -fno-elide-constructors.
>>
>> Hmm, trivial op= doesn't ever hit this code?
> 
> With -fno-elide-constructors we hit the trivial_fn_p block twice in
> constexpr-nsdmi9.C, once for "constexpr Y::Y(const Y&)" and then for
> "constexpr Y& Y::operator=(Y&&)".  So it does hit the code, but only
> with -fno-elide-constructors.

Odd, I'm not sure why that would make a difference for assignment.

>>> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?  I don't
>>> think I want to backport this.
>>>
>>> 	PR c++/101073
>>>
>>> gcc/cp/ChangeLog:
>>>
>>> 	* constexpr.cc (cxx_eval_call_expression): Replace shortcutting trivial
>>> 	constructor/op= with a checking assert.
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>> 	* g++.dg/cpp0x/constexpr-nsdmi3.C: New test.
>>> 	* g++.dg/cpp1y/constexpr-nsdmi10.C: New test.
>>> ---
>>>    gcc/cp/constexpr.cc                           | 25 +++----------------
>>>    gcc/testsuite/g++.dg/cpp0x/constexpr-nsdmi3.C | 17 +++++++++++++
>>>    .../g++.dg/cpp1y/constexpr-nsdmi10.C          | 18 +++++++++++++
>>>    3 files changed, 38 insertions(+), 22 deletions(-)
>>>    create mode 100644 gcc/testsuite/g++.dg/cpp0x/constexpr-nsdmi3.C
>>>    create mode 100644 gcc/testsuite/g++.dg/cpp1y/constexpr-nsdmi10.C
>>>
>>> diff --git a/gcc/cp/constexpr.cc b/gcc/cp/constexpr.cc
>>> index 564766c8a00..1d53dcf0f20 100644
>>> --- a/gcc/cp/constexpr.cc
>>> +++ b/gcc/cp/constexpr.cc
>>> @@ -2865,28 +2865,9 @@ cxx_eval_call_expression (const constexpr_ctx *ctx, tree t,
>>>          ctx = &new_ctx;
>>>        }
>>> -  /* Shortcut trivial constructor/op=.  */
>>> -  if (trivial_fn_p (fun))
>>> -    {
>>> -      tree init = NULL_TREE;
>>> -      if (call_expr_nargs (t) == 2)
>>> -	init = convert_from_reference (get_nth_callarg (t, 1));
>>> -      else if (TREE_CODE (t) == AGGR_INIT_EXPR
>>> -	       && AGGR_INIT_ZERO_FIRST (t))
>>> -	init = build_zero_init (DECL_CONTEXT (fun), NULL_TREE, false);
>>> -      if (init)
>>> -	{
>>> -	  tree op = get_nth_callarg (t, 0);
>>> -	  if (is_dummy_object (op))
>>> -	    op = ctx->object;
>>> -	  else
>>> -	    op = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (op)), op);
>>> -	  tree set = build2 (MODIFY_EXPR, TREE_TYPE (op), op, init);
>>
>> I think the problem is using MODIFY_EXPR instead of INIT_EXPR to represent a
>> constructor; that's why cxx_eval_store_expression thinks it's OK to
>> preevaluate.  This should properly use those two tree codes for op= and
>> ctor, respectively.
> 
> Maybe it was so that the RHS in SET could refer to the op in the LHS?

I think it was just an oversight.  You need INIT_EXPR for the rhs to 
refer to the lhs.

>>> -	  new_ctx.call = &new_call;
>>> -	  return cxx_eval_constant_expression (&new_ctx, set, lval,
>>> -					       non_constant_p, overflow_p);
>>> -	}
>>> -    }
>>> +  /* We used to shortcut trivial constructor/op= here, but nowadays
>>> +     we can only get a trivial function here with -fno-elide-constructors.  */
>>> +  gcc_checking_assert (!trivial_fn_p (fun) || !flag_elide_constructors);
>>
>> ...but if this optimization is so rarely triggered, this simplification is
>> OK too.
> 
> I'd say that's better so that we don't have to update the code (like
> r234345 did).

Indeed, the patch is OK.

Jason