[PATCH] c++: mark TARGET_EXPRs for function arguments eliding [PR114707]

[PATCH] c++: mark TARGET_EXPRs for function arguments eliding [PR114707]

[Marek Polacek]

Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?

-- >8 --
Coming back to our discussion in
<https://gcc.gnu.org/pipermail/gcc-patches/2024-April/649426.html>:
TARGET_EXPRs that initialize a function argument are not marked
TARGET_EXPR_ELIDING_P even though gimplify_arg drops such TARGET_EXPRs
on the floor.  To work around it, I added a pset to
replace_placeholders_for_class_temp_r, but it would be best to just rely
on TARGET_EXPR_ELIDING_P.

This patch changes the diagnostic we emit in constexpr-diag1.C: instead
of:

constexpr-diag1.C:20:21: error: temporary of non-literal type 'B' in a constant expression

we say:

constexpr-diag1.C:20:23: error: call to non-'constexpr' function 'B::B()'

	PR c++/114707

gcc/cp/ChangeLog:

	* call.cc (convert_for_arg_passing): Call set_target_expr_eliding.
	* typeck2.cc (replace_placeholders_for_class_temp_r): Don't use pset.
	(digest_nsdmi_init): Call cp_walk_tree_without_duplicates instead of
	cp_walk_tree.

gcc/testsuite/ChangeLog:

	* g++.dg/cpp0x/constexpr-diag1.C: Adjust dg-error.
---
 gcc/cp/call.cc                               |  3 +++
 gcc/cp/typeck2.cc                            | 20 ++++----------------
 gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C |  2 +-
 3 files changed, 8 insertions(+), 17 deletions(-)

diff --git a/gcc/cp/call.cc b/gcc/cp/call.cc
index ed68eb3c568..750ecf60fd9 100644
--- a/gcc/cp/call.cc
+++ b/gcc/cp/call.cc
@@ -9437,6 +9437,9 @@ convert_for_arg_passing (tree type, tree val, tsubst_flags_t complain)
   if (complain & tf_warning)
     warn_for_address_of_packed_member (type, val);
 
+  /* gimplify_arg elides TARGET_EXPRs that initialize a function argument.  */
+  set_target_expr_eliding (val);
+
   return val;
 }
 
diff --git a/gcc/cp/typeck2.cc b/gcc/cp/typeck2.cc
index 06bad4d3303..7782f38da43 100644
--- a/gcc/cp/typeck2.cc
+++ b/gcc/cp/typeck2.cc
@@ -1409,16 +1409,14 @@ digest_init_flags (tree type, tree init, int flags, tsubst_flags_t complain)
    in the context of guaranteed copy elision).  */
 
 static tree
-replace_placeholders_for_class_temp_r (tree *tp, int *, void *data)
+replace_placeholders_for_class_temp_r (tree *tp, int *, void *)
 {
   tree t = *tp;
-  auto pset = static_cast<hash_set<tree> *>(data);
 
   /* We're looking for a TARGET_EXPR nested in the whole expression.  */
   if (TREE_CODE (t) == TARGET_EXPR
       /* That serves as temporary materialization, not an initializer.  */
-      && !TARGET_EXPR_ELIDING_P (t)
-      && !pset->add (t))
+      && !TARGET_EXPR_ELIDING_P (t))
     {
       tree init = TARGET_EXPR_INITIAL (t);
       while (TREE_CODE (init) == COMPOUND_EXPR)
@@ -1433,16 +1431,6 @@ replace_placeholders_for_class_temp_r (tree *tp, int *, void *data)
 	  gcc_checking_assert (!find_placeholders (init));
 	}
     }
-  /* TARGET_EXPRs initializing function arguments are not marked as eliding,
-     even though gimplify_arg drops them on the floor.  Don't go replacing
-     placeholders in them.  */
-  else if (TREE_CODE (t) == CALL_EXPR || TREE_CODE (t) == AGGR_INIT_EXPR)
-    for (int i = 0; i < call_expr_nargs (t); ++i)
-      {
-	tree arg = get_nth_callarg (t, i);
-	if (TREE_CODE (arg) == TARGET_EXPR && !TARGET_EXPR_ELIDING_P (arg))
-	  pset->add (arg);
-      }
 
   return NULL_TREE;
 }
@@ -1490,8 +1478,8 @@ digest_nsdmi_init (tree decl, tree init, tsubst_flags_t complain)
      temporary materialization does not occur when initializing an object
      from a prvalue of the same type, therefore we must not replace the
      placeholder with a temporary object so that it can be elided.  */
-  hash_set<tree> pset;
-  cp_walk_tree (&init, replace_placeholders_for_class_temp_r, &pset, nullptr);
+  cp_walk_tree_without_duplicates (&init, replace_placeholders_for_class_temp_r,
+				   nullptr);
 
   return init;
 }
diff --git a/gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C b/gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C
index ccb8d81adca..da6fc2030bc 100644
--- a/gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C
+++ b/gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C
@@ -17,4 +17,4 @@ constexpr int b = A<B>().f();	// { dg-error "" }
 
 template <class T>
 constexpr int f (T t) { return 42; }
-constexpr int x = f(B());	     // { dg-error "non-literal" }
+constexpr int x = f(B());	     // { dg-error "call to non-.constexpr." }

base-commit: 2b2476d4d18c92b8aba3567ebccd2100c2f7c258
-- 
2.45.1

Re: [PATCH] c++: mark TARGET_EXPRs for function arguments eliding [PR114707]

[Jason Merrill]

On 5/23/24 10:41, Marek Polacek wrote:
> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
> 
> -- >8 --
> Coming back to our discussion in
> <https://gcc.gnu.org/pipermail/gcc-patches/2024-April/649426.html>:
> TARGET_EXPRs that initialize a function argument are not marked
> TARGET_EXPR_ELIDING_P even though gimplify_arg drops such TARGET_EXPRs
> on the floor.

But only if TREE_TYPE (TARGET_EXPR_INITIAL is non-void, I think we 
should check that here too to be parallel.

Perhaps most/all affected TARGET_EXPRs will have been handled earlier in 
the function under the TREE_ADDRESSABLE check, but I wouldn't rely on 
that without an assert.

> To work around it, I added a pset to
> replace_placeholders_for_class_temp_r, but it would be best to just rely
> on TARGET_EXPR_ELIDING_P.
> 
> This patch changes the diagnostic we emit in constexpr-diag1.C: instead
> of:
> 
> constexpr-diag1.C:20:21: error: temporary of non-literal type 'B' in a constant expression
> 
> we say:
> 
> constexpr-diag1.C:20:23: error: call to non-'constexpr' function 'B::B()'
> 
> 	PR c++/114707
> 
> gcc/cp/ChangeLog:
> 
> 	* call.cc (convert_for_arg_passing): Call set_target_expr_eliding.
> 	* typeck2.cc (replace_placeholders_for_class_temp_r): Don't use pset.
> 	(digest_nsdmi_init): Call cp_walk_tree_without_duplicates instead of
> 	cp_walk_tree.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp0x/constexpr-diag1.C: Adjust dg-error.
> ---
>   gcc/cp/call.cc                               |  3 +++
>   gcc/cp/typeck2.cc                            | 20 ++++----------------
>   gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C |  2 +-
>   3 files changed, 8 insertions(+), 17 deletions(-)
> 
> diff --git a/gcc/cp/call.cc b/gcc/cp/call.cc
> index ed68eb3c568..750ecf60fd9 100644
> --- a/gcc/cp/call.cc
> +++ b/gcc/cp/call.cc
> @@ -9437,6 +9437,9 @@ convert_for_arg_passing (tree type, tree val, tsubst_flags_t complain)
>     if (complain & tf_warning)
>       warn_for_address_of_packed_member (type, val);
>   
> +  /* gimplify_arg elides TARGET_EXPRs that initialize a function argument.  */
> +  set_target_expr_eliding (val);
> +
>     return val;
>   }
>   
> diff --git a/gcc/cp/typeck2.cc b/gcc/cp/typeck2.cc
> index 06bad4d3303..7782f38da43 100644
> --- a/gcc/cp/typeck2.cc
> +++ b/gcc/cp/typeck2.cc
> @@ -1409,16 +1409,14 @@ digest_init_flags (tree type, tree init, int flags, tsubst_flags_t complain)
>      in the context of guaranteed copy elision).  */
>   
>   static tree
> -replace_placeholders_for_class_temp_r (tree *tp, int *, void *data)
> +replace_placeholders_for_class_temp_r (tree *tp, int *, void *)
>   {
>     tree t = *tp;
> -  auto pset = static_cast<hash_set<tree> *>(data);
>   
>     /* We're looking for a TARGET_EXPR nested in the whole expression.  */
>     if (TREE_CODE (t) == TARGET_EXPR
>         /* That serves as temporary materialization, not an initializer.  */
> -      && !TARGET_EXPR_ELIDING_P (t)
> -      && !pset->add (t))
> +      && !TARGET_EXPR_ELIDING_P (t))
>       {
>         tree init = TARGET_EXPR_INITIAL (t);
>         while (TREE_CODE (init) == COMPOUND_EXPR)
> @@ -1433,16 +1431,6 @@ replace_placeholders_for_class_temp_r (tree *tp, int *, void *data)
>   	  gcc_checking_assert (!find_placeholders (init));
>   	}
>       }
> -  /* TARGET_EXPRs initializing function arguments are not marked as eliding,
> -     even though gimplify_arg drops them on the floor.  Don't go replacing
> -     placeholders in them.  */
> -  else if (TREE_CODE (t) == CALL_EXPR || TREE_CODE (t) == AGGR_INIT_EXPR)
> -    for (int i = 0; i < call_expr_nargs (t); ++i)
> -      {
> -	tree arg = get_nth_callarg (t, i);
> -	if (TREE_CODE (arg) == TARGET_EXPR && !TARGET_EXPR_ELIDING_P (arg))
> -	  pset->add (arg);
> -      }
>   
>     return NULL_TREE;
>   }
> @@ -1490,8 +1478,8 @@ digest_nsdmi_init (tree decl, tree init, tsubst_flags_t complain)
>        temporary materialization does not occur when initializing an object
>        from a prvalue of the same type, therefore we must not replace the
>        placeholder with a temporary object so that it can be elided.  */
> -  hash_set<tree> pset;
> -  cp_walk_tree (&init, replace_placeholders_for_class_temp_r, &pset, nullptr);
> +  cp_walk_tree_without_duplicates (&init, replace_placeholders_for_class_temp_r,
> +				   nullptr);
>   
>     return init;
>   }
> diff --git a/gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C b/gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C
> index ccb8d81adca..da6fc2030bc 100644
> --- a/gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C
> +++ b/gcc/testsuite/g++.dg/cpp0x/constexpr-diag1.C
> @@ -17,4 +17,4 @@ constexpr int b = A<B>().f();	// { dg-error "" }
>   
>   template <class T>
>   constexpr int f (T t) { return 42; }
> -constexpr int x = f(B());	     // { dg-error "non-literal" }
> +constexpr int x = f(B());	     // { dg-error "call to non-.constexpr." }
> 
> base-commit: 2b2476d4d18c92b8aba3567ebccd2100c2f7c258

[PATCH v2] c++: mark TARGET_EXPRs for function arguments eliding [PR114707]

[Marek Polacek]

On Thu, May 23, 2024 at 04:04:13PM -0400, Jason Merrill wrote:
> On 5/23/24 10:41, Marek Polacek wrote:
> > Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
> > 
> > -- >8 --
> > Coming back to our discussion in
> > <https://gcc.gnu.org/pipermail/gcc-patches/2024-April/649426.html>:
> > TARGET_EXPRs that initialize a function argument are not marked
> > TARGET_EXPR_ELIDING_P even though gimplify_arg drops such TARGET_EXPRs
> > on the floor.
> 
> But only if TREE_TYPE (TARGET_EXPR_INITIAL is non-void, I think we should
> check that here too to be parallel.

Ah yes, definitely.
 
> Perhaps most/all affected TARGET_EXPRs will have been handled earlier in the
> function under the TREE_ADDRESSABLE check, but I wouldn't rely on that
> without an assert.

So like this or you want an assert somewhere too?  dg.exp passed.

-- >8 --
Coming back to our discussion in
<https://gcc.gnu.org/pipermail/gcc-patches/2024-April/649426.html>:
TARGET_EXPRs that initialize a function argument are not marked
TARGET_EXPR_ELIDING_P even though gimplify_arg drops such TARGET_EXPRs
on the floor.  To work around it, I added a pset to
replace_placeholders_for_class_temp_r, but it would be best to just rely
on TARGET_EXPR_ELIDING_P.

	PR c++/114707

gcc/cp/ChangeLog:

	* call.cc (convert_for_arg_passing): Call set_target_expr_eliding.
	* typeck2.cc (replace_placeholders_for_class_temp_r): Don't use pset.
	(digest_nsdmi_init): Call cp_walk_tree_without_duplicates instead of
	cp_walk_tree.
---
 gcc/cp/call.cc    |  6 ++++++
 gcc/cp/typeck2.cc | 20 ++++----------------
 2 files changed, 10 insertions(+), 16 deletions(-)

diff --git a/gcc/cp/call.cc b/gcc/cp/call.cc
index ed68eb3c568..35c024f2c7c 100644
--- a/gcc/cp/call.cc
+++ b/gcc/cp/call.cc
@@ -9437,6 +9437,12 @@ convert_for_arg_passing (tree type, tree val, tsubst_flags_t complain)
   if (complain & tf_warning)
     warn_for_address_of_packed_member (type, val);
 
+  /* gimplify_arg elides TARGET_EXPRs that initialize a function argument.  */
+  if (TREE_CODE (val) == TARGET_EXPR)
+    if (tree init = TARGET_EXPR_INITIAL (val))
+      if (!VOID_TYPE_P (TREE_TYPE (init)))
+	set_target_expr_eliding (val);
+
   return val;
 }
 
diff --git a/gcc/cp/typeck2.cc b/gcc/cp/typeck2.cc
index 06bad4d3303..7782f38da43 100644
--- a/gcc/cp/typeck2.cc
+++ b/gcc/cp/typeck2.cc
@@ -1409,16 +1409,14 @@ digest_init_flags (tree type, tree init, int flags, tsubst_flags_t complain)
    in the context of guaranteed copy elision).  */
 
 static tree
-replace_placeholders_for_class_temp_r (tree *tp, int *, void *data)
+replace_placeholders_for_class_temp_r (tree *tp, int *, void *)
 {
   tree t = *tp;
-  auto pset = static_cast<hash_set<tree> *>(data);
 
   /* We're looking for a TARGET_EXPR nested in the whole expression.  */
   if (TREE_CODE (t) == TARGET_EXPR
       /* That serves as temporary materialization, not an initializer.  */
-      && !TARGET_EXPR_ELIDING_P (t)
-      && !pset->add (t))
+      && !TARGET_EXPR_ELIDING_P (t))
     {
       tree init = TARGET_EXPR_INITIAL (t);
       while (TREE_CODE (init) == COMPOUND_EXPR)
@@ -1433,16 +1431,6 @@ replace_placeholders_for_class_temp_r (tree *tp, int *, void *data)
 	  gcc_checking_assert (!find_placeholders (init));
 	}
     }
-  /* TARGET_EXPRs initializing function arguments are not marked as eliding,
-     even though gimplify_arg drops them on the floor.  Don't go replacing
-     placeholders in them.  */
-  else if (TREE_CODE (t) == CALL_EXPR || TREE_CODE (t) == AGGR_INIT_EXPR)
-    for (int i = 0; i < call_expr_nargs (t); ++i)
-      {
-	tree arg = get_nth_callarg (t, i);
-	if (TREE_CODE (arg) == TARGET_EXPR && !TARGET_EXPR_ELIDING_P (arg))
-	  pset->add (arg);
-      }
 
   return NULL_TREE;
 }
@@ -1490,8 +1478,8 @@ digest_nsdmi_init (tree decl, tree init, tsubst_flags_t complain)
      temporary materialization does not occur when initializing an object
      from a prvalue of the same type, therefore we must not replace the
      placeholder with a temporary object so that it can be elided.  */
-  hash_set<tree> pset;
-  cp_walk_tree (&init, replace_placeholders_for_class_temp_r, &pset, nullptr);
+  cp_walk_tree_without_duplicates (&init, replace_placeholders_for_class_temp_r,
+				   nullptr);
 
   return init;
 }

base-commit: ee492101c2e51b58e926307448d35f539aec0b2c
-- 
2.45.1

Re: [PATCH v2] c++: mark TARGET_EXPRs for function arguments eliding [PR114707]

[Jason Merrill]

On 5/23/24 20:32, Marek Polacek wrote:
> On Thu, May 23, 2024 at 04:04:13PM -0400, Jason Merrill wrote:
>> On 5/23/24 10:41, Marek Polacek wrote:
>>> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
>>>
>>> -- >8 --
>>> Coming back to our discussion in
>>> <https://gcc.gnu.org/pipermail/gcc-patches/2024-April/649426.html>:
>>> TARGET_EXPRs that initialize a function argument are not marked
>>> TARGET_EXPR_ELIDING_P even though gimplify_arg drops such TARGET_EXPRs
>>> on the floor.
>>
>> But only if TREE_TYPE (TARGET_EXPR_INITIAL is non-void, I think we should
>> check that here too to be parallel.
> 
> Ah yes, definitely.
>   
>> Perhaps most/all affected TARGET_EXPRs will have been handled earlier in the
>> function under the TREE_ADDRESSABLE check, but I wouldn't rely on that
>> without an assert.
> 
> So like this or you want an assert somewhere too?  dg.exp passed.
> 
> -- >8 --
> Coming back to our discussion in
> <https://gcc.gnu.org/pipermail/gcc-patches/2024-April/649426.html>:
> TARGET_EXPRs that initialize a function argument are not marked
> TARGET_EXPR_ELIDING_P even though gimplify_arg drops such TARGET_EXPRs
> on the floor.  To work around it, I added a pset to
> replace_placeholders_for_class_temp_r, but it would be best to just rely
> on TARGET_EXPR_ELIDING_P.
> 
> 	PR c++/114707
> 
> gcc/cp/ChangeLog:
> 
> 	* call.cc (convert_for_arg_passing): Call set_target_expr_eliding.
> 	* typeck2.cc (replace_placeholders_for_class_temp_r): Don't use pset.
> 	(digest_nsdmi_init): Call cp_walk_tree_without_duplicates instead of
> 	cp_walk_tree.
> ---
>   gcc/cp/call.cc    |  6 ++++++
>   gcc/cp/typeck2.cc | 20 ++++----------------
>   2 files changed, 10 insertions(+), 16 deletions(-)
> 
> diff --git a/gcc/cp/call.cc b/gcc/cp/call.cc
> index ed68eb3c568..35c024f2c7c 100644
> --- a/gcc/cp/call.cc
> +++ b/gcc/cp/call.cc
> @@ -9437,6 +9437,12 @@ convert_for_arg_passing (tree type, tree val, tsubst_flags_t complain)
>     if (complain & tf_warning)
>       warn_for_address_of_packed_member (type, val);
>   
> +  /* gimplify_arg elides TARGET_EXPRs that initialize a function argument.  */
> +  if (TREE_CODE (val) == TARGET_EXPR)
> +    if (tree init = TARGET_EXPR_INITIAL (val))
> +      if (!VOID_TYPE_P (TREE_TYPE (init)))

You can simplify this test to 'if (SIMPLE_TARGET_EXPR_P ...'.  OK with 
that change.

Jason