Re: [PATCH] c++: Fix ICE in tsubst_default_argument [PR92010]

[Jason Merrill <jason@redhat.com>]

On 3/31/20 3:50 PM, Patrick Palka wrote:
> On Tue, 31 Mar 2020, Jason Merrill wrote:
> 
>> On 3/30/20 6:46 PM, Patrick Palka wrote:
>>> On Mon, 30 Mar 2020, Jason Merrill wrote:
>>>> On 3/30/20 3:58 PM, Patrick Palka wrote:
>>>>> On Thu, 26 Mar 2020, Jason Merrill wrote:
>>>>>
>>>>>> On 3/22/20 9:21 PM, Patrick Palka wrote:
>>>>>>> This patch relaxes an assertion in tsubst_default_argument that
>>>>>>> exposes
>>>>>>> a
>>>>>>> latent
>>>>>>> bug in how we substitute an array type into a cv-qualified wildcard
>>>>>>> function
>>>>>>> parameter type.  Concretely, the latent bug is that given the
>>>>>>> function
>>>>>>> template
>>>>>>>
>>>>>>>       template<typename T> void foo(const T t);
>>>>>>>
>>>>>>> one would expect the type of foo<int[]> to be void(const int*), but
>>>>>>> we
>>>>>>> (seemingly prematurely) strip function parameter types of their
>>>>>>> top-level
>>>>>>> cv-qualifiers when building the function's TYPE_ARG_TYPES, and
>>>>>>> instead
>>>>>>> end
>>>>>>> up
>>>>>>> obtaining void(int*) as the type of foo<int[]> after substitution
>>>>>>> and
>>>>>>> decaying.
>>>>>>>
>>>>>>> We still however correctly substitute into and decay the formal
>>>>>>> parameter
>>>>>>> type,
>>>>>>> obtaining const int* as the type of t after substitution.  But this
>>>>>>> then
>>>>>>> leads
>>>>>>> to us tripping over the assert in tsubst_default_argument that
>>>>>>> verifies
>>>>>>> the
>>>>>>> formal parameter type and the function type are consistent.
>>>>>>>
>>>>>>> Assuming it's too late at this stage to fix the substitution bug, we
>>>>>>> can
>>>>>>> still
>>>>>>> relax the assertion like so.  Tested on x86_64-pc-linux-gnu, does
>>>>>>> this
>>>>>>> look
>>>>>>> OK?
>>>>>>
>>>>>> This is core issues 1001/1322, which have not been resolved.  Clang
>>>>>> does
>>>>>> the
>>>>>> substitution the way you suggest; EDG rejects the testcase because the
>>>>>> two
>>>>>> substitutions produce different results.  I think it would make sense
>>>>>> to
>>>>>> follow the EDG behavior until this issue is actually resolved.
>>>>>
>>>>> Here is what I have so far towards that end.  When substituting into the
>>>>> PARM_DECLs of a function decl, we now additionally check if the
>>>>> aforementioned Core issues are relevant and issue a (fatal) diagnostic
>>>>> if so.  This patch checks this in tsubst_decl <case PARM_DECL> rather
>>>>> than in tsubst_function_decl for efficiency reasons, so that we don't
>>>>> have to perform another traversal over the DECL_ARGUMENTS /
>>>>> TYPE_ARG_TYPES just to implement this check.
>>>>
>>>> Hmm, this seems like writing more complicated code for a very marginal
>>>> optimization; how many function templates have so many parameters that
>>>> walking
>>>> over them once to compare types will have any effect on compile time?
>>>
>>> Good point... though I just tried implementing this check in
>>> tsubst_function_decl, and it seems it might be just as complicated to
>>> implement it there instead, at least if we want to handle function
>>> parameter packs correctly.
>>>
>>> If we were to implement this check in tsubst_function_decl, then since
>>> we have access to the instantiated function, it would presumably suffice
>>> to compare its substituted DECL_ARGUMENTS with its substituted
>>> TYPE_ARG_TYPES to see if they're consistent.  Doing so would certainly
>>> catch the original testcase, i.e.
>>>
>>>     template<typename T>
>>>       void foo(const T);
>>>     int main() { foo<int[]>(0); }
>>>
>>> because the DECL_ARGUMENTS of foo<int[]> would be {const int*} and its
>>> TYPE_ARG_TYPES would be {int*}.  But apparently it doesn't catch the
>>> corresponding testcase that uses a function parameter pack, i.e.
>>>
>>>     template<typename... Ts>
>>>       void foo(const Ts...);
>>>     int main() { foo<int[]>(0); }
>>>
>>> because it turns out we don't strip top-level cv-qualifiers from
>>> function parameter packs from TYPE_ARG_TYPES at declaration time, as we
>>> do with regular function parameters.  So in this second testcase both
>>> DECL_ARGUMENTS and TYPE_ARG_TYPES of foo<int[]> would be {const int*},
>>> and yet we would (presumably) want to reject this instantiation too.
>>>
>>> So it seems comparing TYPE_ARG_TYPES and DECL_ARGUMENTS from
>>> tsubst_function_decl would not suffice, and we would still need to do a
>>> variant of the trick that's done in this patch, i.e. substitute into
>>> each dependent parameter type stripped of its top-level cv-qualifiers,
>>> to see if these cv-qualifiers make a material difference in the
>>> resulting function type.  Or maybe there's yet another way to detect
>>> this?
>>
>> I think let's go ahead with comparing TYPE_ARG_TYPES and DECL_ARGUMENTS; the
>> problem comes when they disagree.  If we're handling pack expansions wrong,
>> that's a separate issue.
> 
> Hm, comparing TYPE_ARG_TYPES and DECL_ARGUMENTS for compatibility seems
> to be exposing a latent bug with how we handle lambdas that appear in
> function parameter types.  Take g++.dg/cpp2a/lambda-uneval3.C for
> example:
> 
>      template <class T> void spam(decltype([]{}) (*s)[sizeof(T)]) {}
>      int main() { spam<char>(nullptr); }
> 
> According to tsubst_function_decl in current trunk, the type of the
> function paremeter 's' of spam<char> according to its TYPE_ARG_TYPES is
>      struct ._anon_4[1] *
> and according to its DECL_ARGUMENTS the type of 's' is
>      struct ._anon_5[1] *
> 
> The disagreement happens because we call tsubst_lambda_expr twice during
> substitution and thereby generate two distinct lambda types, one when
> substituting into the TYPE_ARG_TYPES and another when substituting into
> the DECL_ARGUMENTS.  I'm not sure how to work around this
> bug/false-positive..

Oof.

I think probably the right answer is to rebuild TYPE_ARG_TYPES from 
DECL_ARGUMENTS if they don't match.

Jason