Re: [PATCH] c++: ICE w/ maybe_warn_nodiscard on non-dep call [PR123495, PR126860]

On 8/14/26 12:38 PM, Patrick Palka wrote:
> On Fri, 14 Aug 2026, Marek Polacek wrote:
> 
>> On Fri, Aug 14, 2026 at 12:22:55PM -0400, Patrick Palka wrote:
>>> Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK
>>> for trunk? It's a checking-only ICE so no big need to backport,
>>> maybe to 16 at most.
>>>
>>> -- >8 --
>>>
>>> Another latent issue uncovered by the NON_DEPENDENT_EXPR removal.
>>> cp_get_fndecl_from_callee, called from maybe_warn_nodiscard, tries to
>>> fold the templated callee via constant evaluation, but constant
>>> evaluation expects non-templated trees.  So use fold_non_dependent_expr
>>> which instantiates templated trees beforehand.
>>>
>>> 	PR c++/123495
>>> 	PR c++/126860
>>>
>>> gcc/cp/ChangeLog:
>>>
>>> 	* cvt.cc (cp_get_fndecl_from_callee): Use fold_non_dependent_expr
>>> 	instead of maybe_constant_init.
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>> 	* g++.dg/template/non-dependent36.C: New test.
>>> 	* g++.dg/template/non-dependent37.C: New test.
>>> ---
>>>   gcc/cp/cvt.cc                                   | 2 +-
>>>   gcc/testsuite/g++.dg/template/non-dependent36.C | 4 ++++
>>>   gcc/testsuite/g++.dg/template/non-dependent37.C | 9 +++++++++
>>>   3 files changed, 14 insertions(+), 1 deletion(-)
>>>   create mode 100644 gcc/testsuite/g++.dg/template/non-dependent36.C
>>>   create mode 100644 gcc/testsuite/g++.dg/template/non-dependent37.C
>>>
>>> diff --git a/gcc/cp/cvt.cc b/gcc/cp/cvt.cc
>>> index a378b51a2bed..33d528b3c324 100644
>>> --- a/gcc/cp/cvt.cc
>>> +++ b/gcc/cp/cvt.cc
>>> @@ -1069,7 +1069,7 @@ cp_get_fndecl_from_callee (tree fn, bool fold /* = true */)
>>>     if (type == NULL_TREE || !INDIRECT_TYPE_P (type))
>>>       return NULL_TREE;
>>>     if (fold)
>>> -    fn = maybe_constant_init (fn);
>>> +    fn = fold_non_dependent_expr (fn);
>>
>> Since this is _init, I would expect fold_non_dependent_init.
>> Was _expr a deliberate choice?
> 
> Ah I forgot about fold_non_dependent_init, so I defaulted to _expr,
> but if I knew about it I would use _init for sake of consistency.
> Testing that now.  I'm not sure why we use _init here in the
> first place though, maybe Jason knows?
Hmm, no, I don't remember and can't think of any reason for that; there 
should be no difference for a result that doesn't end up as NULL_TREE 
anyway.  The _expr variant is OK.

Jason