Re: [PATCH v2] c++: ICE with noexcept and canonical types [PR101715]

[Jason Merrill <jason@redhat.com>]

On 1/20/22 20:03, Marek Polacek wrote:
> On Thu, Jan 20, 2022 at 03:23:24PM -0500, Jason Merrill wrote:
>> On 1/18/22 11:05, Marek Polacek wrote:
>>> On Mon, Jan 17, 2022 at 01:48:48PM -0500, Jason Merrill wrote:
>>>> On 1/14/22 19:22, Marek Polacek wrote:
>>>>> This is a "canonical types differ for identical types" ICE, which started
>>>>> with r11-4682.  It's a bit tricky to explain.  Consider:
>>>>>
>>>>>      template <typename T> struct S {
>>>>>        S<T> bar() noexcept(T::value);  // #1
>>>>>        S<T> foo() noexcept(T::value);  // #2
>>>>>      };
>>>>>
>>>>>      template <typename T> S<T> S<T>::foo() noexcept(T::value) {}  // #3
>>>>>
>>>>> We ICE because #3 and #2 have the same type, but their canonical types
>>>>> differ: TYPE_CANONICAL (#3) == #2 but TYPE_CANONICAL (#2) == #1.
>>>>>
>>>>> The member functions #1 and #2 have the same type.  However, since their
>>>>> noexcept-specifier is deferred, when parsing them, we create a variant for
>>>>> both of them, because DEFERRED_PARSE cannot be compared.  In other words,
>>>>> build_cp_fntype_variant's
>>>>>
>>>>>      tree v = TYPE_MAIN_VARIANT (type);
>>>>>      for (; v; v = TYPE_NEXT_VARIANT (v))
>>>>>        if (cp_check_qualified_type (v, type, type_quals, rqual, raises, late))
>>>>>          return v;
>>>>>
>>>>> will *not* find an existing variant when creating a method_type for #2, so we
>>>>> have to create a new one.
>>>>>
>>>>> But then we perform delayed parsing and call fixup_deferred_exception_variants
>>>>> for #1 and #2.  f_d_e_v will replace TYPE_RAISES_EXCEPTIONS with the newly
>>>>> parsed noexcept-specifier.  It also sets TYPE_CANONICAL (#2) to #1.  Both
>>>>> noexcepts turned out to be the same, so now we have two equivalent variants in
>>>>> the list!  I.e.,
>>>>>
>>>>> +-----------------+      +-----------------+      +-----------------+
>>>>> |      main       |      |      #2         |      |      #1         |
>>>>> | S S::<T379>(S*) |----->| S S::<T37c>(S*) |----->| S S::<T37a>(S*) |----->NULL
>>>>> |    -            |      |  noex(T::value) |      |  noex(T::value) |
>>>>> +-----------------+      +-----------------+      +-----------------+
>>>>>
>>>>> Then we get to #3.  As for #1 and #2, grokdeclarator calls build_memfn_type,
>>>>> which ends up calling build_cp_fntype_variant, which will use the loop
>>>>> above to look for an existing variant.  The first one that matches
>>>>> cp_check_qualified_type will be used, so we use #2 rather than #1, and the
>>>>> TYPE_CANONICAL mismatch follows.  Hopefully that makes sense.
>>>>
>>>> Why doesn't the TYPE_CANONICAL (v) == v check prevent this?
>>>
>>> In other words, I think you're asking: why did fixup_deferred_exception_variants
>>> set TYPE_CANONICAL (#2) to #1 (which then differs from TYPE_CANONICAL (#3),
>>> which is #2)?
>>
>> I meant to ask why TYPE_CANONICAL (#3) got set to #2 instead of #1?
>>
>> And to answer my own question, it's because the check I mention is in
>> fixup_deferred_exception_variants, and #3 doesn't go through there at all;
>> the loop in build_cp_fntype_variant assumes no duplicate variants, which
>> your patch fixes.
> 
> Right, fixup_deferred_exception_variants is only called for fn decls in
> unparsed_noexcepts.
> 
>>> The method_type for #1 (I'll mark is as #1 here) is built with it being its own
>>> canonical type.
>>>
>>> The first call to fixup_deferred_exception_variants does not change it: in
>>> there, VARIANT is #1, the loop with 'TYPE_CANONICAL (v) == v' cannot find
>>> an existing variant that would match, so when we do
>>>
>>>       v = build_cp_fntype_variant (TYPE_CANONICAL (variant),
>>>                                    rqual, cr, false);
>>> we get #1 so
>>>       TYPE_CANONICAL (variant) = v;
>>> is just
>>>       TYPE_CANONICAL (#1) = #1;
>>> so no change.
>>>
>>> The second call to fixup_deferred_exception_variants: here we're working with
>>> VARIANT #2.  Now we again scan the list of variants {main, #2, #1} where we
>>> find a match for #2: #1.  #1's TYPE_CANONICAL is #1 as per above, so we set
>>>       TYPE_CANONICAL (#2) = #1;
>>> which I think is correct.
>>>
>>>
>>> I think TYPE_CANONICAL (#3) should also be #1, not #2, which my patch attempts
>>> to do.
>>>
>>>
>>> Hope this explanation makes some sense, please ask away if it doesn't!
>>>
>>>>> As for the fix, I didn't think I could rewrite the method_type #2 with #1
>>>>> because the type may have escaped via decltype.  So my approach is to
>>>>> elide #2 from the list, so when looking for a matching variant, we always
>>>>> find #1 (#2 remains live though, which admittedly sounds sort of dodgy).
>>>>>
>>>>> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk/11?
>>>>>
>>>>> 	PR c++/101715
>>>>>
>>>>> gcc/cp/ChangeLog:
>>>>>
>>>>> 	* tree.c (fixup_deferred_exception_variants): Remove duplicate
>>>>> 	variants after parsing the exception specifications.
>>>>>
>>>>> gcc/testsuite/ChangeLog:
>>>>>
>>>>> 	* g++.dg/cpp0x/noexcept72.C: New test.
>>>>> 	* g++.dg/cpp0x/noexcept73.C: New test.
>>>>> ---
>>>>>     gcc/cp/tree.c                           | 16 +++++++++++++++-
>>>>>     gcc/testsuite/g++.dg/cpp0x/noexcept72.C | 21 +++++++++++++++++++++
>>>>>     gcc/testsuite/g++.dg/cpp0x/noexcept73.C | 13 +++++++++++++
>>>>>     3 files changed, 49 insertions(+), 1 deletion(-)
>>>>>     create mode 100644 gcc/testsuite/g++.dg/cpp0x/noexcept72.C
>>>>>     create mode 100644 gcc/testsuite/g++.dg/cpp0x/noexcept73.C
>>>>>
>>>>> diff --git a/gcc/cp/tree.c b/gcc/cp/tree.c
>>>>> index 7f7de86b4e8..2efad49e7c1 100644
>>>>> --- a/gcc/cp/tree.c
>>>>> +++ b/gcc/cp/tree.c
>>>>> @@ -2804,8 +2804,9 @@ fixup_deferred_exception_variants (tree type, tree raises)
>>>>>       /* Though sucky, this walk will process the canonical variants
>>>>>          first.  */
>>>>> +  tree prev = NULL_TREE;
>>>>>       for (tree variant = TYPE_MAIN_VARIANT (type);
>>>>> -       variant; variant = TYPE_NEXT_VARIANT (variant))
>>>>> +       variant; prev = variant, variant = TYPE_NEXT_VARIANT (variant))
>>>>>         if (TYPE_RAISES_EXCEPTIONS (variant) == original)
>>>>>           {
>>>>>     	gcc_checking_assert (variant != TYPE_MAIN_VARIANT (type));
>>>>> @@ -2827,6 +2828,19 @@ fixup_deferred_exception_variants (tree type, tree raises)
>>>>>     	      v = build_cp_fntype_variant (TYPE_CANONICAL (variant),
>>>>>     					   rqual, cr, false);
>>>>>     	    TYPE_CANONICAL (variant) = v;
>>>>> +
>>>>> +	    /* If VARIANT became a duplicate (cp_check_qualified_type-wise)
>>>>> +	       of an existing variant in the variant list of TYPE after we
>>>>> +	       have parsed its exception specification, elide it.  Otherwise,
>>>>> +	       build_cp_fntype_variant would use it, leading to "canonical
>>>>> +	       types differ for identical types."  */
>>>>> +	    for (v = TYPE_MAIN_VARIANT (type); v; v = TYPE_NEXT_VARIANT (v))
>>>>> +	      if (v != variant
>>>>> +		  /* The main variant will not have TYPE_RAISES_EXCEPTIONS
>>>>> +		     so PREV should never be null.  */
>>>>> +		  && cp_check_qualified_type (v, variant, var_quals,
>>>>> +					      rqual, cr, false))
>>>>> +		TYPE_NEXT_VARIANT (prev) = TYPE_NEXT_VARIANT (variant);
>>
>> I think we don't two loops through the variants.  It ought to work to
>> replace the existing loop with yours; if we find v, we prune and use its
>> TYPE_CANONICAL.
> 
> Ah yes, good idea; I don't actually need to wait till TYPE_RAISES_EXCEPTIONS
> is set on variant!  The following seems to work just as well.
> 
> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
> 
> -- >8 --
> This is a "canonical types differ for identical types" ICE, which started
> with r11-4682.  It's a bit tricky to explain.  Consider:
> 
>    template <typename T> struct S {
>      S<T> bar() noexcept(T::value);  // #1
>      S<T> foo() noexcept(T::value);  // #2
>    };
> 
>    template <typename T> S<T> S<T>::foo() noexcept(T::value) {}  // #3
> 
> We ICE because #3 and #2 have the same type, but their canonical types
> differ: TYPE_CANONICAL (#3) == #2 but TYPE_CANONICAL (#2) == #1.
> 
> The member functions #1 and #2 have the same type.  However, since their
> noexcept-specifier is deferred, when parsing them, we create a variant for
> both of them, because DEFERRED_PARSE cannot be compared.  In other words,
> build_cp_fntype_variant's
> 
>    tree v = TYPE_MAIN_VARIANT (type);
>    for (; v; v = TYPE_NEXT_VARIANT (v))
>      if (cp_check_qualified_type (v, type, type_quals, rqual, raises, late))
>        return v;
> 
> will *not* find an existing variant when creating a method_type for #2, so we
> have to create a new one.
> 
> But then we perform delayed parsing and call fixup_deferred_exception_variants
> for #1 and #2.  f_d_e_v will replace TYPE_RAISES_EXCEPTIONS with the newly
> parsed noexcept-specifier.  It also sets TYPE_CANONICAL (#2) to #1.  Both
> noexcepts turned out to be the same, so now we have two equivalent variants in
> the list!  I.e.,
> 
> +-----------------+      +-----------------+      +-----------------+
> |      main       |      |      #2         |      |      #1         |
> | S S::<T379>(S*) |----->| S S::<T37c>(S*) |----->| S S::<T37a>(S*) |----->NULL
> |    -            |      |  noex(T::value) |      |  noex(T::value) |
> +-----------------+      +-----------------+      +-----------------+
> 
> Then we get to #3.  As for #1 and #2, grokdeclarator calls build_memfn_type,
> which ends up calling build_cp_fntype_variant, which will use the loop
> above to look for an existing variant.  The first one that matches
> cp_check_qualified_type will be used, so we use #2 rather than #1, and the
> TYPE_CANONICAL mismatch follows.  Hopefully that makes sense.
> 
> As for the fix, I didn't think I could rewrite the method_type #2 with #1
> because the type may have escaped via decltype.  So my approach is to
> elide #2 from the list, so when looking for a matching variant, we always
> find #1 (#2 remains live though, which admittedly sounds sort of dodgy).
> 
> 	PR c++/101715
> 
> gcc/cp/ChangeLog:
> 
> 	* tree.c (fixup_deferred_exception_variants): Remove duplicate
> 	variants after parsing the exception specifications.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp0x/noexcept72.C: New test.
> 	* g++.dg/cpp0x/noexcept73.C: New test.
> ---
>   gcc/cp/tree.cc                          | 16 ++++++++++++++--
>   gcc/testsuite/g++.dg/cpp0x/noexcept72.C | 21 +++++++++++++++++++++
>   gcc/testsuite/g++.dg/cpp0x/noexcept73.C | 13 +++++++++++++
>   3 files changed, 48 insertions(+), 2 deletions(-)
>   create mode 100644 gcc/testsuite/g++.dg/cpp0x/noexcept72.C
>   create mode 100644 gcc/testsuite/g++.dg/cpp0x/noexcept73.C
> 
> diff --git a/gcc/cp/tree.cc b/gcc/cp/tree.cc
> index bcd44e73921..17436f0512d 100644
> --- a/gcc/cp/tree.cc
> +++ b/gcc/cp/tree.cc
> @@ -2804,8 +2804,9 @@ fixup_deferred_exception_variants (tree type, tree raises)
>   
>     /* Though sucky, this walk will process the canonical variants
>        first.  */
> +  tree prev = NULL_TREE;
>     for (tree variant = TYPE_MAIN_VARIANT (type);
> -       variant; variant = TYPE_NEXT_VARIANT (variant))
> +       variant; prev = variant, variant = TYPE_NEXT_VARIANT (variant))
>       if (TYPE_RAISES_EXCEPTIONS (variant) == original)
>         {
>   	gcc_checking_assert (variant != TYPE_MAIN_VARIANT (type));
> @@ -2815,12 +2816,23 @@ fixup_deferred_exception_variants (tree type, tree raises)
>   	    cp_cv_quals var_quals = TYPE_QUALS (variant);
>   	    cp_ref_qualifier rqual = type_memfn_rqual (variant);
>   
> +	    /* If VARIANT would become a dup (cp_check_qualified_type-wise)
> +	       of an existing variant in the variant list of TYPE after its
> +	       exception specification has been parsed, elide it.  Otherwise,
> +	       build_cp_fntype_variant could use it, leading to "canonical
> +	       types differ for identical types."  */
>   	    tree v = TYPE_MAIN_VARIANT (type);
>   	    for (; v; v = TYPE_NEXT_VARIANT (v))
>   	      if (TYPE_CANONICAL (v) == v

I think we want to drop the TYPE_CANONICAL check here, and below change

TYPE_CANONICAL (variant) = v;

to

TYPE_CANONICAL (variant) = TYPE_CANONICAL (v);

so that this also works for e.g. signatures involving typedefs.

> +		  && v != variant

I think we don't need this check since we haven't changed 
TYPE_RAISES_EXCEPTIONS yet.

>   		  && cp_check_qualified_type (v, variant, var_quals,
>   					      rqual, cr, false))
> -		break;
> +		{
> +		  /* The main variant will not match V, so PREV will never
> +		     be null.  */
> +		  TYPE_NEXT_VARIANT (prev) = TYPE_NEXT_VARIANT (variant);
> +		  break;
> +		}
>   	    TYPE_RAISES_EXCEPTIONS (variant) = raises;
>   
>   	    if (!v)
> diff --git a/gcc/testsuite/g++.dg/cpp0x/noexcept72.C b/gcc/testsuite/g++.dg/cpp0x/noexcept72.C
> new file mode 100644
> index 00000000000..f1455b3b46b
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/cpp0x/noexcept72.C
> @@ -0,0 +1,21 @@
> +// PR c++/101715
> +// { dg-do compile { target c++11 } }
> +
> +template <typename T> struct S {
> +  S<T> bar() noexcept(T::value);  // #1
> +  S<T> foo() noexcept(T::value);  // #2
> +};
> +
> +template <typename T> S<T> S<T>::foo() noexcept(T::value) {}  // #3
> +
> +template <typename T> struct S2 {
> +  S2<T> bar1() noexcept(T::value);
> +  S2<T> bar2() noexcept(T::value);
> +  S2<T> bar3() noexcept(T::value);
> +  S2<T> bar4() noexcept(T::value);
> +  S2<T> bar5() noexcept(T::value);
> +  S2<T> baz() noexcept(T::value2);
> +  S2<T> foo() noexcept(T::value);
> +};
> +
> +template <typename T> S2<T> S2<T>::foo() noexcept(T::value) {}
> diff --git a/gcc/testsuite/g++.dg/cpp0x/noexcept73.C b/gcc/testsuite/g++.dg/cpp0x/noexcept73.C
> new file mode 100644
> index 00000000000..24524f3592a
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/cpp0x/noexcept73.C
> @@ -0,0 +1,13 @@
> +// PR c++/101715
> +// { dg-do compile { target c++11 } }
> +
> +template <typename T> struct S { };
> +
> +template<typename T>
> +struct A
> +{
> +    A& foo(A&&) noexcept((S<T>::value));
> +    A& assign(A&&) noexcept((S<T>::value));
> +};
> +template<typename T>
> +A<T>& A<T>::foo(A&&) noexcept((S<T>::value)) {}
> 
> base-commit: d2ad748eeef0dd260f3993b8dcbffbded3240a0a