Re: [PATCH v2] c++: ICE with constexpr lambda [PR107280]

[Jason Merrill <jason@redhat.com>]

On 3/15/23 10:37, Marek Polacek wrote:
> On Fri, Mar 10, 2023 at 01:47:46PM -0500, Jason Merrill wrote:
>> On 3/10/23 11:17, Marek Polacek wrote:
>>> We crash here since r10-3661, the store_init_value hunk in particular.
>>> Before, we called cp_fully_fold_init, so e.g.
>>>
>>>     {.str=VIEW_CONVERT_EXPR<char[8]>("")}
>>>
>>> was folded into
>>>
>>>     {.str=""}
>>>
>>> but now we don't fold and keep the VCE around, and it causes trouble in
>>> cxx_eval_store_expression: in the !refs->is_empty () loop we descend on
>>> .str's initializer but since it's wrapped in a VCE, we skip the STRING_CST
>>> check and then crash on the CONSTRUCTOR_NO_CLEARING.
>>>
>>> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk/12?
>>>
>>> 	PR c++/107280
>>>
>>> gcc/cp/ChangeLog:
>>>
>>> 	* constexpr.cc (cxx_eval_store_expression): Strip location wrappers.
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>> 	* g++.dg/cpp1z/constexpr-lambda28.C: New test.
>>> ---
>>>    gcc/cp/constexpr.cc                             |  3 ++-
>>>    gcc/testsuite/g++.dg/cpp1z/constexpr-lambda28.C | 15 +++++++++++++++
>>>    2 files changed, 17 insertions(+), 1 deletion(-)
>>>    create mode 100644 gcc/testsuite/g++.dg/cpp1z/constexpr-lambda28.C
>>>
>>> diff --git a/gcc/cp/constexpr.cc b/gcc/cp/constexpr.cc
>>> index 8683c00596a..abf6ee560c5 100644
>>> --- a/gcc/cp/constexpr.cc
>>> +++ b/gcc/cp/constexpr.cc
>>> @@ -6033,7 +6033,8 @@ cxx_eval_store_expression (const constexpr_ctx *ctx, tree t,
>>>    	  *valp = build_constructor (type, NULL);
>>>    	  CONSTRUCTOR_NO_CLEARING (*valp) = no_zero_init;
>>>    	}
>>> -      else if (TREE_CODE (*valp) == STRING_CST)
>>> +      else if (STRIP_ANY_LOCATION_WRAPPER (*valp),
>>> +	       TREE_CODE (*valp) == STRING_CST)
>>
>> Seems like this is stripping the location wrapper when we try to modify the
>> string; I think we want to strip it earlier, when we first initialize the
>> array member.
> 
> Hmm, I suppose we don't want to do the stripping too early.  I could
> have put it in get_nsdmi, for instance, but maybe here is good as well?

I guess I was thinking more around the time when the value is imported 
into constant evaluation, i.e. this chunk of cxx_eval_constant_expression:

>         if (tree init = DECL_INITIAL (r))
>           {
>             init = cxx_eval_constant_expression (ctx, init, vc_prvalue,
>                                                  non_constant_p, overflow_p);
>             /* Don't share a CONSTRUCTOR that might be changed.  */
>             init = unshare_constructor (init);
>             /* Remember that a constant object's constructor has already                                                            
>                run.  */
>             if (CLASS_TYPE_P (TREE_TYPE (r))
>                 && CP_TYPE_CONST_P (TREE_TYPE (r)))
>               TREE_READONLY (init) = true;
>             ctx->global->put_value (r, init);
>           }

Feel free to pursue that approach or go ahead and push your first patch, 
whichever you prefer.

Jason