RFA (make_dispatcher_decl): PATCH for c++/83911, ICE with multiversioned constructor

RFA (make_dispatcher_decl): PATCH for c++/83911, ICE with multiversioned constructor

[Jason Merrill]

[-- Attachment #1: Type: text/plain, Size: 467 bytes --]

As I mentioned in the PR, the problem here is that we're replacing a
constructor with a dispatcher function which doesn't look much like a
constructor.  This patch adjusts make_dispatcher_decl to make it look
more like the functions it dispatches to, but other things are certain
to break for similar reasons down the road.  A proper solution should
be more transparent, like thunks.

Tested x86_64-pc-linux-gnu.  Does this seem worth applying to fix the
regression?

[-- Attachment #2: 83911.diff --]
[-- Type: text/plain, Size: 3632 bytes --]

commit d3b4453834bb2e3cc1c875e76fe593c16f263f77
Author: Jason Merrill <jason@redhat.com>
Date:   Fri Feb 16 16:53:47 2018 -0500

            PR c++/83911 - ICE with multiversioned constructor.
    
    gcc/
            * attribs.c (make_dispatcher_decl): Copy METHOD_TYPE,
            DECL_CXX_{CON,DE}STRUCTOR_P.
    gcc/cp/
            * cp-tree.h (DECL_CONSTRUCTOR_P): Use DECL_CXX_CONSTRUCTOR_P.
            (DECL_DESTRUCTOR_P): Use DECL_CXX_DESTRUCTOR_P.
            * mangle.c (write_unqualified_name): Check IDENTIFIER_[CD]TOR_P.

diff --git a/gcc/attribs.c b/gcc/attribs.c
index bfadf124dcb..2e9e69904f0 100644
--- a/gcc/attribs.c
+++ b/gcc/attribs.c
@@ -1063,9 +1063,14 @@ make_dispatcher_decl (const tree decl)
   func_name = xstrdup (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl)));
 
   fn_type = TREE_TYPE (decl);
-  func_type = build_function_type (TREE_TYPE (fn_type),
-				   TYPE_ARG_TYPES (fn_type));
-  
+  if (TREE_CODE (fn_type) == METHOD_TYPE)
+    func_type = build_method_type_directly (TYPE_METHOD_BASETYPE (fn_type),
+					    TREE_TYPE (fn_type),
+					    TYPE_ARG_TYPES (fn_type));
+  else
+    func_type = build_function_type (TREE_TYPE (fn_type),
+				     TYPE_ARG_TYPES (fn_type));
+
   func_decl = build_fn_decl (func_name, func_type);
   XDELETEVEC (func_name);
   TREE_USED (func_decl) = 1;
@@ -1078,6 +1083,9 @@ make_dispatcher_decl (const tree decl)
   /* This will be of type IFUNCs have to be externally visible.  */
   TREE_PUBLIC (func_decl) = 1;
 
+  DECL_CXX_CONSTRUCTOR_P (func_decl) = DECL_CXX_CONSTRUCTOR_P (decl);
+  DECL_CXX_DESTRUCTOR_P (func_decl) = DECL_CXX_DESTRUCTOR_P (decl);
+
   return func_decl;  
 }
 
diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h
index 17d8c6d2650..268be0fd543 100644
--- a/gcc/cp/cp-tree.h
+++ b/gcc/cp/cp-tree.h
@@ -2695,7 +2695,7 @@ struct GTY(()) lang_decl {
 /* For FUNCTION_DECLs and TEMPLATE_DECLs: nonzero means that this function
    is a constructor.  */
 #define DECL_CONSTRUCTOR_P(NODE) \
-  IDENTIFIER_CTOR_P (DECL_NAME (NODE))
+  DECL_CXX_CONSTRUCTOR_P (STRIP_TEMPLATE (NODE))
 
 /* Nonzero if NODE (a FUNCTION_DECL) is a constructor for a complete
    object.  */
@@ -2724,7 +2724,7 @@ struct GTY(()) lang_decl {
 /* Nonzero if NODE (a FUNCTION_DECL or TEMPLATE_DECL)
    is a destructor.  */
 #define DECL_DESTRUCTOR_P(NODE)				\
-  IDENTIFIER_DTOR_P (DECL_NAME (NODE))
+  DECL_CXX_DESTRUCTOR_P (STRIP_TEMPLATE (NODE))
 
 /* Nonzero if NODE (a FUNCTION_DECL) is a destructor, but not the
    specialized in-charge constructor, in-charge deleting constructor,
diff --git a/gcc/cp/mangle.c b/gcc/cp/mangle.c
index 94c4bed2848..ecd4eb066d4 100644
--- a/gcc/cp/mangle.c
+++ b/gcc/cp/mangle.c
@@ -1351,9 +1351,9 @@ write_unqualified_name (tree decl)
   else if (DECL_DECLARES_FUNCTION_P (decl))
     {
       found = true;
-      if (DECL_CONSTRUCTOR_P (decl))
+      if (IDENTIFIER_CTOR_P (DECL_NAME (decl)))
 	write_special_name_constructor (decl);
-      else if (DECL_DESTRUCTOR_P (decl))
+      else if (IDENTIFIER_DTOR_P (DECL_NAME (decl)))
 	write_special_name_destructor (decl);
       else if (DECL_CONV_FN_P (decl))
 	{
diff --git a/gcc/testsuite/g++.dg/ext/mv27.C b/gcc/testsuite/g++.dg/ext/mv27.C
new file mode 100644
index 00000000000..443a54be765
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ext/mv27.C
@@ -0,0 +1,18 @@
+// PR c++/83911
+// { dg-do compile { target i?86-*-* x86_64-*-* } }
+// { dg-require-ifunc "" }
+
+class SimdFloat
+{
+public:
+    __attribute__ ((target ("default")))
+    SimdFloat(float x) {}
+
+    __attribute__ ((target ("avx2")))
+    SimdFloat(float x) {}
+};
+
+SimdFloat foo()
+{
+    return 1;
+}

Re: RFA (make_dispatcher_decl): PATCH for c++/83911, ICE with multiversioned constructor

[Jason Merrill]

Ping

On Fri, Mar 2, 2018 at 1:23 PM, Jason Merrill <jason@redhat.com> wrote:
> As I mentioned in the PR, the problem here is that we're replacing a
> constructor with a dispatcher function which doesn't look much like a
> constructor.  This patch adjusts make_dispatcher_decl to make it look
> more like the functions it dispatches to, but other things are certain
> to break for similar reasons down the road.  A proper solution should
> be more transparent, like thunks.
>
> Tested x86_64-pc-linux-gnu.  Does this seem worth applying to fix the
> regression?

Re: RFA (make_dispatcher_decl): PATCH for c++/83911, ICE with multiversioned constructor

[Richard Biener]

On Wed, Mar 14, 2018 at 8:57 PM, Jason Merrill <jason@redhat.com> wrote:
> Ping
>
> On Fri, Mar 2, 2018 at 1:23 PM, Jason Merrill <jason@redhat.com> wrote:
>> As I mentioned in the PR, the problem here is that we're replacing a
>> constructor with a dispatcher function which doesn't look much like a
>> constructor.  This patch adjusts make_dispatcher_decl to make it look
>> more like the functions it dispatches to, but other things are certain
>> to break for similar reasons down the road.  A proper solution should
>> be more transparent, like thunks.
>>
>> Tested x86_64-pc-linux-gnu.  Does this seem worth applying to fix the
>> regression?

The patch looks reasonable to me, you probably know best whether
the cp/ parts are risky or not ;)

So - OK from my POV.

And yes, thunks may be a better representation for the dispatcher.

Richard.

Re: RFA (make_dispatcher_decl): PATCH for c++/83911, ICE with multiversioned constructor

[Jason Merrill]

[-- Attachment #1: Type: text/plain, Size: 1104 bytes --]

On Thu, Mar 15, 2018 at 4:50 AM, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Wed, Mar 14, 2018 at 8:57 PM, Jason Merrill <jason@redhat.com> wrote:
>> Ping
>>
>> On Fri, Mar 2, 2018 at 1:23 PM, Jason Merrill <jason@redhat.com> wrote:
>>> As I mentioned in the PR, the problem here is that we're replacing a
>>> constructor with a dispatcher function which doesn't look much like a
>>> constructor.  This patch adjusts make_dispatcher_decl to make it look
>>> more like the functions it dispatches to, but other things are certain
>>> to break for similar reasons down the road.  A proper solution should
>>> be more transparent, like thunks.
>>>
>>> Tested x86_64-pc-linux-gnu.  Does this seem worth applying to fix the
>>> regression?
>
> The patch looks reasonable to me, you probably know best whether
> the cp/ parts are risky or not ;)
>
> So - OK from my POV.
>
> And yes, thunks may be a better representation for the dispatcher.

It occurred to me that I could handle this more locally by deferring
the function substitution until genericization time, so this is what
I'm checking in:

[-- Attachment #2: 83911.diff --]
[-- Type: text/x-patch, Size: 3094 bytes --]

commit 05dda93e56e02859c65f93bd541dc7793bc7305c
Author: Jason Merrill <jason@redhat.com>
Date:   Fri Feb 16 16:53:47 2018 -0500

            PR c++/83911 - ICE with multiversioned constructor.
    
            * cp-gimplify.c (cp_genericize_r): Replace versioned function with
            dispatchere here.
            * call.c (build_over_call): Not here.
            PR c++/83911 - ICE with multiversioned constructor.

diff --git a/gcc/cp/call.c b/gcc/cp/call.c
index 45c22aaa312..67438ff2e94 100644
--- a/gcc/cp/call.c
+++ b/gcc/cp/call.c
@@ -8204,23 +8204,6 @@ build_over_call (struct z_candidate *cand, int flags, tsubst_flags_t complain)
 	}
     }
 
-  /* For calls to a multi-versioned function, overload resolution
-     returns the function with the highest target priority, that is,
-     the version that will checked for dispatching first.  If this
-     version is inlinable, a direct call to this version can be made
-     otherwise the call should go through the dispatcher.  */
-
-  if (DECL_FUNCTION_VERSIONED (fn)
-      && (current_function_decl == NULL
-	  || !targetm.target_option.can_inline_p (current_function_decl, fn)))
-    {
-      fn = get_function_version_dispatcher (fn);
-      if (fn == NULL)
-	return NULL;
-      if (!already_used)
-	mark_versions_used (fn);
-    }
-
   if (!already_used
       && !mark_used (fn, complain))
     return error_mark_node;
diff --git a/gcc/cp/cp-gimplify.c b/gcc/cp/cp-gimplify.c
index 0ddd435454c..653d1dcee26 100644
--- a/gcc/cp/cp-gimplify.c
+++ b/gcc/cp/cp-gimplify.c
@@ -1513,6 +1513,29 @@ cp_genericize_r (tree *stmt_p, int *walk_subtrees, void *data)
 		       == REFERENCE_TYPE))
 	    *walk_subtrees = 0;
 	}
+      /* Fall through.  */
+    case AGGR_INIT_EXPR:
+      /* For calls to a multi-versioned function, overload resolution
+	 returns the function with the highest target priority, that is,
+	 the version that will checked for dispatching first.  If this
+	 version is inlinable, a direct call to this version can be made
+	 otherwise the call should go through the dispatcher.  */
+      {
+	tree fn = cp_get_callee_fndecl (stmt);
+	if (fn && DECL_FUNCTION_VERSIONED (fn)
+	    && (current_function_decl == NULL
+		|| !targetm.target_option.can_inline_p (current_function_decl,
+							fn)))
+	  if (tree dis = get_function_version_dispatcher (fn))
+	    {
+	      mark_versions_used (dis);
+	      dis = build_address (dis);
+	      if (TREE_CODE (stmt) == CALL_EXPR)
+		CALL_EXPR_FN (stmt) = dis;
+	      else
+		AGGR_INIT_EXPR_FN (stmt) = dis;
+	    }
+      }
       break;
 
     default:
diff --git a/gcc/testsuite/g++.dg/ext/mv27.C b/gcc/testsuite/g++.dg/ext/mv27.C
new file mode 100644
index 00000000000..443a54be765
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ext/mv27.C
@@ -0,0 +1,18 @@
+// PR c++/83911
+// { dg-do compile { target i?86-*-* x86_64-*-* } }
+// { dg-require-ifunc "" }
+
+class SimdFloat
+{
+public:
+    __attribute__ ((target ("default")))
+    SimdFloat(float x) {}
+
+    __attribute__ ((target ("avx2")))
+    SimdFloat(float x) {}
+};
+
+SimdFloat foo()
+{
+    return 1;
+}

Re: RFA (make_dispatcher_decl): PATCH for c++/83911, ICE with multiversioned constructor

[Jason Merrill]

[-- Attachment #1: Type: text/plain, Size: 1384 bytes --]

On Fri, Mar 16, 2018 at 8:38 AM, Jason Merrill <jason@redhat.com> wrote:
> On Thu, Mar 15, 2018 at 4:50 AM, Richard Biener
> <richard.guenther@gmail.com> wrote:
>> On Wed, Mar 14, 2018 at 8:57 PM, Jason Merrill <jason@redhat.com> wrote:
>>> Ping
>>>
>>> On Fri, Mar 2, 2018 at 1:23 PM, Jason Merrill <jason@redhat.com> wrote:
>>>> As I mentioned in the PR, the problem here is that we're replacing a
>>>> constructor with a dispatcher function which doesn't look much like a
>>>> constructor.  This patch adjusts make_dispatcher_decl to make it look
>>>> more like the functions it dispatches to, but other things are certain
>>>> to break for similar reasons down the road.  A proper solution should
>>>> be more transparent, like thunks.
>>>>
>>>> Tested x86_64-pc-linux-gnu.  Does this seem worth applying to fix the
>>>> regression?
>>
>> The patch looks reasonable to me, you probably know best whether
>> the cp/ parts are risky or not ;)
>>
>> So - OK from my POV.
>>
>> And yes, thunks may be a better representation for the dispatcher.
>
> It occurred to me that I could handle this more locally by deferring
> the function substitution until genericization time, so this is what
> I'm checking in:

...but now that we're in stage 1, it still seems sensible to have a
single way of checking whether something is a constructor.

Tested x86_64-pc-linux-gnu, applying to trunk.

[-- Attachment #2: ctor_p.diff --]
[-- Type: text/x-patch, Size: 1163 bytes --]

commit 656d038fe0cf78a2432a8c9a047edc93af6d5b23
Author: Jason Merrill <jason@redhat.com>
Date:   Fri Mar 16 08:32:26 2018 -0400

            * cp-tree.h (DECL_CONSTRUCTOR_P): Use DECL_CXX_CONSTRUCTOR_P.
    
            (DECL_DESTRUCTOR_P): Use DECL_CXX_DESTRUCTOR_P.

diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h
index 2df158c9ea6..a4e0099a249 100644
--- a/gcc/cp/cp-tree.h
+++ b/gcc/cp/cp-tree.h
@@ -2731,7 +2731,7 @@ struct GTY(()) lang_decl {
 /* For FUNCTION_DECLs and TEMPLATE_DECLs: nonzero means that this function
    is a constructor.  */
 #define DECL_CONSTRUCTOR_P(NODE) \
-  IDENTIFIER_CTOR_P (DECL_NAME (NODE))
+  DECL_CXX_CONSTRUCTOR_P (STRIP_TEMPLATE (NODE))
 
 /* Nonzero if NODE (a FUNCTION_DECL) is a constructor for a complete
    object.  */
@@ -2760,7 +2760,7 @@ struct GTY(()) lang_decl {
 /* Nonzero if NODE (a FUNCTION_DECL or TEMPLATE_DECL)
    is a destructor.  */
 #define DECL_DESTRUCTOR_P(NODE)				\
-  IDENTIFIER_DTOR_P (DECL_NAME (NODE))
+  DECL_CXX_DESTRUCTOR_P (STRIP_TEMPLATE (NODE))
 
 /* Nonzero if NODE (a FUNCTION_DECL) is a destructor, but not the
    specialized in-charge constructor, in-charge deleting constructor,