Re: [PATCH] c++: ahead-of-time overload set pruning for non-dep calls

[Jason Merrill <jason@redhat.com>]

On 12/16/21 14:53, Patrick Palka wrote:
> On Thu, 16 Dec 2021, Patrick Palka wrote:
> 
>> On Wed, 15 Dec 2021, Jason Merrill wrote:
>>
>>> On 12/15/21 12:49, Patrick Palka wrote:
>>>> This patch makes us remember the function selected by overload
>>>> resolution during ahead of time processing of a non-dependent call
>>>> expression, so that we avoid repeating most of the work of overload
>>>> resolution at instantiation time.  This mirrors what we already do for
>>>> non-dependent operator expressions via build_min_non_dep_op_overload.
>>>>
>>>> Some caveats:
>>>>
>>>>    * When processing ahead of time a non-dependent call to a member
>>>>      function template inside a class template (as in
>>>>      g++.dg/template/deduce4.C), we end up generating an "inverted" partial
>>>>      instantiation such as S<T>::foo<int, int>(), the kinds of which we're
>>>>      apparently not prepared to fully instantiate (e.g. tsubst_baselink
>>>>      mishandles it).  So this patch disables this optimization for such
>>>>      functions and adds a FIXME.
>>>
>>> I wonder if it would be worthwhile to build a TEMPLATE_ID_EXPR to remember the
>>> deduced template args, even if we are failing to remember the actual function?
>>
>> Hmm, that transformation could have observable effects, since overload
>> resolution for f<int>(0) might end up instantiating more things than for
>> f(0) due to the explicit-args substitution step:
>>
>>    template<class T> struct A { using type = typename T::type; };
>>
>>    template<class T> void f(T);
>>    template<class T, class U = typename T::type> typename A<T>::type f(T);
>>
>> Here overload resolution for f(0) succeeds and selects the first
>> overload but for f<int>(0) induces a hard error.  Also I worry that such
>> a transformation might affect declaration matching in weird ways due
>> to conflating f(0) with f<int>(0).
>>
>> We could at least though prune the overload set to the corresponding
>> selected function template rather than the "inside-out" specialization;
>> I'll try to implement that.
> 
> Here's a bootstrapped and regtested patch which implements that:

OK.

> -- >8 --
> 
> This patch makes us remember the function selected by overload resolution
> during ahead of time processing of a non-dependent call expression, so
> that we avoid repeating most of the work of overload resolution for the
> call at instantiation time.  Note that we already do this for
> non-dependent operator expressions via build_min_non_dep_op_overload.
> 
> Some caveats:
> 
>   * When processing ahead of time a non-dependent call to a member
>     function template inside a common class template (as in
>     g++.dg/template/deduce4.C), we end up generating an "inverted" partial
>     instantiation such as S<T>::foo<int, int>(), the likes of which we're
>     apparently not prepared to fully instantiate (e.g. tsubst_baselink
>     mishandles it).  So this patch makes us prune to the selected
>     template instead of the specialization in this case.
> 
>   * This change triggered a latent FUNCTION_DECL pretty printing issue
>     in cpp0x/error2.C -- since we now resolve the call to foo<0> ahead
>     of time, the error now looks like:
> 
>       error: expansion pattern ‘foo()()=0’ contains no parameter pack
> 
>     where the FUNCTION_DECL for foo<0> is clearly misprinted.  But this
>     pretty-printing issue could be reproduced without this patch if
>     we define foo as a non-template function.  Since this testcase was
>     added to verify pretty printing of TEMPLATE_ID_EXPR, I work around
>     this test failure by making the call to foo type-dependent and thus
>     immune to this ahead of time pruning.
> 
>   * We now reject parts of cpp0x/fntmp-equiv1.C because we notice that
>     the call d(f, b) in
> 
>       template <unsigned long f, unsigned b, typename> e<d(f, b)> d();
> 
>     is always non-constexpr because the selected d isn't.  I tried fixing
>     this by making it constexpr, but then the call to d from main becomes
>     ambiguous.  So I settled with removing this part of the testcase.
> 
> gcc/cp/ChangeLog:
> 
> 	* call.c (build_new_method_call): For a non-dependent call
> 	expression inside a template, returning a templated tree
> 	whose overload set contains just the selected function.
> 	* semantics.c (finish_call_expr): Likewise.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp0x/error2.C: Make the call to foo type-dependent in
> 	order to avoid latent pretty-printing issue for FUNCTION_DECL
> 	inside MODOP_EXPR.
> 	* g++.dg/cpp0x/fntmp-equiv1.C: Remove ill-formed parts of
> 	testcase that we now diagnose.
> 	* g++.dg/template/non-dependent16.C: New test.
> 	* g++.dg/template/non-dependent16a.C: New test.
> 	* g++.dg/template/non-dependent17.C: New test.
> ---
>   gcc/cp/call.c                                 | 27 ++++++++++++++
>   gcc/cp/semantics.c                            | 15 ++++++++
>   gcc/testsuite/g++.dg/cpp0x/error2.C           |  4 +-
>   gcc/testsuite/g++.dg/cpp0x/fntmp-equiv1.C     |  4 --
>   .../g++.dg/template/non-dependent16.C         | 37 +++++++++++++++++++
>   .../g++.dg/template/non-dependent16a.C        | 36 ++++++++++++++++++
>   .../g++.dg/template/non-dependent17.C         | 21 +++++++++++
>   7 files changed, 138 insertions(+), 6 deletions(-)
>   create mode 100644 gcc/testsuite/g++.dg/template/non-dependent16.C
>   create mode 100644 gcc/testsuite/g++.dg/template/non-dependent16a.C
>   create mode 100644 gcc/testsuite/g++.dg/template/non-dependent17.C
> 
> diff --git a/gcc/cp/call.c b/gcc/cp/call.c
> index 495dcdd77b3..1fbfc580a1e 100644
> --- a/gcc/cp/call.c
> +++ b/gcc/cp/call.c
> @@ -11163,6 +11163,33 @@ build_new_method_call (tree instance, tree fns, vec<tree, va_gc> **args,
>   	}
>         if (INDIRECT_REF_P (call))
>   	call = TREE_OPERAND (call, 0);
> +
> +      /* Prune all but the selected function from the original overload
> +	 set so that we can avoid some duplicate work at instantiation time.  */
> +      if (really_overloaded_fn (fns))
> +	{
> +	  if (DECL_TEMPLATE_INFO (fn)
> +	      && DECL_MEMBER_TEMPLATE_P (DECL_TI_TEMPLATE (fn))
> +	      && dependent_type_p (DECL_CONTEXT (fn)))
> +	    {
> +	      /* FIXME: We're not prepared to fully instantiate "inside-out"
> +		 partial instantiations such as A<T>::f<int>().  So instead
> +		 use the selected template, not the specialization.  */
> +
> +	      if (OVL_SINGLE_P (fns))
> +		/* If the original overload set consists of a single function
> +		   template, this isn't beneficial.  */
> +		goto skip_prune;
> +
> +	      fn = ovl_make (DECL_TI_TEMPLATE (fn));
> +	      if (template_only)
> +		fn = lookup_template_function (fn, explicit_targs);
> +	    }
> +	  orig_fns = copy_node (orig_fns);
> +	  BASELINK_FUNCTIONS (orig_fns) = fn;
> +	}
> +
> +skip_prune:
>         call = (build_min_non_dep_call_vec
>   	      (call,
>   	       build_min (COMPONENT_REF, TREE_TYPE (CALL_EXPR_FN (call)),
> diff --git a/gcc/cp/semantics.c b/gcc/cp/semantics.c
> index 6603066c620..6ffd82cb0ec 100644
> --- a/gcc/cp/semantics.c
> +++ b/gcc/cp/semantics.c
> @@ -2893,6 +2893,21 @@ finish_call_expr (tree fn, vec<tree, va_gc> **args, bool disallow_virtual,
>       {
>         if (INDIRECT_REF_P (result))
>   	result = TREE_OPERAND (result, 0);
> +
> +      /* Prune all but the selected function from the original overload
> +	 set so that we can avoid some duplicate work at instantiation time.  */
> +      if (TREE_CODE (result) == CALL_EXPR
> +	  && really_overloaded_fn (orig_fn))
> +	{
> +	  orig_fn = CALL_EXPR_FN (result);
> +	  if (TREE_CODE (orig_fn) == COMPONENT_REF)
> +	    {
> +	      /* The non-dependent result of build_new_method_call.  */
> +	      orig_fn = TREE_OPERAND (orig_fn, 1);
> +	      gcc_assert (BASELINK_P (orig_fn));
> +	    }
> +	}
> +
>         result = build_call_vec (TREE_TYPE (result), orig_fn, orig_args);
>         SET_EXPR_LOCATION (result, input_location);
>         KOENIG_LOOKUP_P (result) = koenig_p;
> diff --git a/gcc/testsuite/g++.dg/cpp0x/error2.C b/gcc/testsuite/g++.dg/cpp0x/error2.C
> index e6af294c180..eb966362ccb 100644
> --- a/gcc/testsuite/g++.dg/cpp0x/error2.C
> +++ b/gcc/testsuite/g++.dg/cpp0x/error2.C
> @@ -3,7 +3,7 @@
>   
>   template<int> int foo();
>   
> -template<typename F> void bar(F f)
> +template<typename F, int N> void bar(F f)
>   {
> -  f((foo<0>()=0)...); // { dg-error "pattern '\\(foo\\<0\\>\\)\\(\\)=0'" }
> +  f((foo<N>()=0)...); // { dg-error "pattern '\\(foo\\<N\\>\\)\\(\\)=0'" }
>   }
> diff --git a/gcc/testsuite/g++.dg/cpp0x/fntmp-equiv1.C b/gcc/testsuite/g++.dg/cpp0x/fntmp-equiv1.C
> index 833ae6fc85c..60ebad8d1d3 100644
> --- a/gcc/testsuite/g++.dg/cpp0x/fntmp-equiv1.C
> +++ b/gcc/testsuite/g++.dg/cpp0x/fntmp-equiv1.C
> @@ -1,10 +1,7 @@
>   // PR c++/86946, DR 1321
>   // { dg-do compile { target c++11 } }
>   
> -int d(int, int);
>   template <long> class e {};
> -template <unsigned long f, unsigned b, typename> e<sizeof(d(f, b))> d();
> -template <unsigned long f, unsigned b, typename> e<d(f, b)> d();
>   
>   template <class T, class U> constexpr T d2(T, U) { return 42; }
>   template <unsigned long f, unsigned b, typename> e<d2(f, b)> d2();
> @@ -17,7 +14,6 @@ template <unsigned long f, unsigned b, typename> e<sizeof(d3(f, b))> d3();
>   
>   int main()
>   {
> -  d<1,2,int>();
>     d2<1,2,int>();
>     d3<1,2,int>();
>   }
> diff --git a/gcc/testsuite/g++.dg/template/non-dependent16.C b/gcc/testsuite/g++.dg/template/non-dependent16.C
> new file mode 100644
> index 00000000000..ee8ef902529
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/template/non-dependent16.C
> @@ -0,0 +1,37 @@
> +// This test verifies that after resolving a non-dependent call expression
> +// ahead of time, we prune all but the selected candidate from the overload
> +// set.  Without this optimization, overload resolution for the final call to
> +// f<void>() would be exponential in the size of the overload set.
> +
> +// { dg-do compile { target c++11 } }
> +
> +template<class T> void f();
> +template<class T> auto f() -> decltype(f<void>(), 1, *T());
> +template<class T> auto f() -> decltype(f<void>(), 2, *T());
> +template<class T> auto f() -> decltype(f<void>(), 3, *T());
> +template<class T> auto f() -> decltype(f<void>(), 4, *T());
> +template<class T> auto f() -> decltype(f<void>(), 5, *T());
> +template<class T> auto f() -> decltype(f<void>(), 6, *T());
> +template<class T> auto f() -> decltype(f<void>(), 7, *T());
> +template<class T> auto f() -> decltype(f<void>(), 8, *T());
> +template<class T> auto f() -> decltype(f<void>(), 9, *T());
> +template<class T> auto f() -> decltype(f<void>(), 10, *T());
> +template<class T> auto f() -> decltype(f<void>(), 11, *T());
> +template<class T> auto f() -> decltype(f<void>(), 12, *T());
> +template<class T> auto f() -> decltype(f<void>(), 13, *T());
> +template<class T> auto f() -> decltype(f<void>(), 14, *T());
> +template<class T> auto f() -> decltype(f<void>(), 15, *T());
> +template<class T> auto f() -> decltype(f<void>(), 16, *T());
> +template<class T> auto f() -> decltype(f<void>(), 17, *T());
> +template<class T> auto f() -> decltype(f<void>(), 18, *T());
> +template<class T> auto f() -> decltype(f<void>(), 19, *T());
> +template<class T> auto f() -> decltype(f<void>(), 20, *T());
> +template<class T> auto f() -> decltype(f<void>(), 21, *T());
> +template<class T> auto f() -> decltype(f<void>(), 22, *T());
> +template<class T> auto f() -> decltype(f<void>(), 23, *T());
> +template<class T> auto f() -> decltype(f<void>(), 24, *T());
> +template<class T> auto f() -> decltype(f<void>(), 25, *T());
> +
> +int main() {
> +  f<void>();
> +}
> diff --git a/gcc/testsuite/g++.dg/template/non-dependent16a.C b/gcc/testsuite/g++.dg/template/non-dependent16a.C
> new file mode 100644
> index 00000000000..0e04d646c0b
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/template/non-dependent16a.C
> @@ -0,0 +1,36 @@
> +// Like non-dependent16.C, but using member functions.
> +
> +// { dg-do compile { target c++11 } }
> +
> +struct A {
> +  template<class T> static void f();
> +  template<class T> static auto f() -> decltype(f<void>(), 1, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 2, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 3, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 4, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 5, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 6, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 7, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 8, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 9, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 10, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 11, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 12, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 13, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 14, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 15, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 16, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 17, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 18, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 19, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 20, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 21, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 22, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 23, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 24, *T());
> +  template<class T> static auto f() -> decltype(f<void>(), 25, *T());
> +};
> +
> +int main() {
> +  A::f<void>();
> +}
> diff --git a/gcc/testsuite/g++.dg/template/non-dependent17.C b/gcc/testsuite/g++.dg/template/non-dependent17.C
> new file mode 100644
> index 00000000000..bc664999e84
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/template/non-dependent17.C
> @@ -0,0 +1,21 @@
> +// A variant of deduce4.C with multiple overloads of foo.  Verify we don't
> +// crash when after ahead-of-time pruning of the overload set for the
> +// non-dependent call to foo.
> +// { dg-do compile }
> +
> +template <typename T>
> +struct S {
> +  template <typename U, typename V>
> +  static void foo(V) { }
> +  template <typename U>
> +  static void foo(...) { }
> +
> +  void bar () { foo<int>(10); }
> +};
> +
> +void
> +test ()
> +{
> +  S<int> s;
> +  s.bar ();
> +}