Re: [PATCH] Improve -Ofast vectorization of std::sin etc. (PR libstdc++/81706)

[Jakub Jelinek <jakub at redhat dot com>]

On Mon, Aug 07, 2017 at 05:27:42PM +0200, Jakub Jelinek wrote:
> > This should really be a separate function.  Perhaps "merge_one_attribute"?
> 
> If it is outlined without the first 7 lines, i.e. just the body of if (b),
> then it could be duplicate_one_attribute (tree *, tree, const char *);
> called like if (b) duplicate_one_attribute (&DECL_ATTRIBUTES (b), s, "omp declare simd");
> If it is duplicated as whole, it should be called
> duplicate_one_attr_to_builtin or something similar.
> In any case, it could be in tree.c or attribs.c.
> 
> The primary question is if we want this behavior, or if we should go the
> libstdc++ patch routine (and for Jonathan the question is if he knows
> why __builtin_XXXf has been used there rather than the ::XXXf).

Here is updated patch that commonizes big part of that into
duplicate_one_attribute.  Bootstrapped/regtested on x86_64-linux and
i686-linux.  The question stands, do we want to go this way or using some
libstdc++ solution?

2017-09-01  Jakub Jelinek  <jakub@redhat.com>

	PR libstdc++/81706
	* attribs.c (attribute_value_equal): Use omp_declare_simd_clauses_equal
	for comparison of OMP_CLAUSEs regardless of flag_openmp{,_simd}.
	(duplicate_one_attribute): New function.
	* attribs.h (duplicate_one_attribute): New declaration.

	* c-decl.c (merge_decls): Copy "omp declare simd" attributes from
	newdecl to corresponding __builtin_ if any.

	* decl.c (duplicate_decls): Copy "omp declare simd" attributes from
	newdecl to corresponding __builtin_ if any.

	* gcc.target/i386/pr81706.c: New test.
	* g++.dg/ext/pr81706.C: New test.

--- gcc/attribs.c.jj	2017-09-01 09:25:37.000000000 +0200
+++ gcc/attribs.c	2017-09-01 09:54:28.581146071 +0200
@@ -1116,9 +1116,9 @@ attribute_value_equal (const_tree attr1,
 				     TREE_VALUE (attr2)) == 1);
     }
 
-  if ((flag_openmp || flag_openmp_simd)
-      && TREE_VALUE (attr1) && TREE_VALUE (attr2)
+  if (TREE_VALUE (attr1)
       && TREE_CODE (TREE_VALUE (attr1)) == OMP_CLAUSE
+      && TREE_VALUE (attr2)
       && TREE_CODE (TREE_VALUE (attr2)) == OMP_CLAUSE)
     return omp_declare_simd_clauses_equal (TREE_VALUE (attr1),
 					   TREE_VALUE (attr2));
@@ -1310,6 +1310,31 @@ merge_decl_attributes (tree olddecl, tre
 			   DECL_ATTRIBUTES (newdecl));
 }
 
+/* Duplicate all attributes with name NAME in ATTR list to *ATTRS if
+   they are missing there.  */
+
+void
+duplicate_one_attribute (tree *attrs, tree attr, const char *name)
+{
+  if (!attr)
+    return;
+  tree a = lookup_attribute (name, *attrs);
+  while (attr)
+    {
+      tree a2;
+      for (a2 = a; a2; a2 = lookup_attribute (name, TREE_CHAIN (a2)))
+	if (attribute_value_equal (attr, a2))
+	  break;
+      if (!a2)
+	{
+	  a2 = copy_node (attr);
+	  TREE_CHAIN (a2) = *attrs;
+	  *attrs = a2;
+	}
+      attr = lookup_attribute (name, TREE_CHAIN (attr));
+    }
+}
+
 #if TARGET_DLLIMPORT_DECL_ATTRIBUTES
 
 /* Specialization of merge_decl_attributes for various Windows targets.
--- gcc/attribs.h.jj	2017-09-01 09:25:37.000000000 +0200
+++ gcc/attribs.h	2017-09-01 09:54:57.366809765 +0200
@@ -77,6 +77,11 @@ extern tree remove_attribute (const char
 
 extern tree merge_attributes (tree, tree);
 
+/* Duplicate all attributes with name NAME in ATTR list to *ATTRS if
+   they are missing there.  */
+
+extern void duplicate_one_attribute (tree *, tree, const char *);
+
 /* Given two Windows decl attributes lists, possibly including
    dllimport, return a list of their union .  */
 extern tree merge_dllimport_decl_attributes (tree, tree);
--- gcc/c/c-decl.c.jj	2017-09-01 09:25:40.707410605 +0200
+++ gcc/c/c-decl.c	2017-09-01 09:49:57.419314085 +0200
@@ -2569,6 +2569,17 @@ merge_decls (tree newdecl, tree olddecl,
 			set_builtin_decl_declared_p (fncode, true);
 		      break;
 		    }
+
+		  tree s = lookup_attribute ("omp declare simd",
+					     DECL_ATTRIBUTES (newdecl));
+		  if (s)
+		    {
+		      tree b
+			= builtin_decl_explicit (DECL_FUNCTION_CODE (newdecl));
+		      if (b)
+			duplicate_one_attribute (&DECL_ATTRIBUTES (b), s,
+						 "omp declare simd");
+		    }
 		}
 	    }
 	  else
--- gcc/cp/decl.c.jj	2017-09-01 09:26:24.748892739 +0200
+++ gcc/cp/decl.c	2017-09-01 09:55:52.940160495 +0200
@@ -2470,6 +2470,16 @@ next_arg:;
 		  break;
 		}
 	    }
+
+	  tree s = lookup_attribute ("omp declare simd",
+				     DECL_ATTRIBUTES (newdecl));
+	  if (s)
+	    {
+	      tree b = builtin_decl_explicit (DECL_FUNCTION_CODE (newdecl));
+	      if (b)
+		duplicate_one_attribute (&DECL_ATTRIBUTES (b), s,
+					 "omp declare simd");
+	    }
 	}
       if (new_defines_function)
 	/* If defining a function declared with other language
--- gcc/testsuite/gcc.target/i386/pr81706.c.jj	2017-09-01 09:40:46.345761553 +0200
+++ gcc/testsuite/gcc.target/i386/pr81706.c	2017-09-01 09:40:46.345761553 +0200
@@ -0,0 +1,32 @@
+/* PR libstdc++/81706 */
+/* { dg-do compile } */
+/* { dg-options "-O3 -mavx2 -mno-avx512f" } */
+/* { dg-final { scan-assembler "call\[^\n\r]_ZGVdN4v_cos" } } */
+/* { dg-final { scan-assembler "call\[^\n\r]_ZGVdN4v_sin" } } */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+extern double cos (double) __attribute__ ((nothrow, leaf, simd ("notinbranch")));
+extern double sin (double) __attribute__ ((nothrow, leaf, simd ("notinbranch")));
+#ifdef __cplusplus
+}
+#endif
+double p[1024] = { 1.0 };
+double q[1024] = { 1.0 };
+
+void
+foo (void)
+{
+  int i;
+  for (i = 0; i < 1024; i++)
+    p[i] = cos (q[i]);
+}
+
+void
+bar (void)
+{
+  int i;
+  for (i = 0; i < 1024; i++)
+    p[i] = __builtin_sin (q[i]);
+}
--- gcc/testsuite/g++.dg/ext/pr81706.C.jj	2017-09-01 09:40:46.345761553 +0200
+++ gcc/testsuite/g++.dg/ext/pr81706.C	2017-09-01 09:40:46.345761553 +0200
@@ -0,0 +1,32 @@
+// PR libstdc++/81706
+// { dg-do compile { target i?86-*-* x86_64-*-* } }
+// { dg-options "-O3 -mavx2 -mno-avx512f" }
+// { dg-final { scan-assembler "call\[^\n\r]_ZGVdN4v_cos" } }
+// { dg-final { scan-assembler "call\[^\n\r]_ZGVdN4v_sin" } }
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+extern double cos (double) __attribute__ ((nothrow, leaf, simd ("notinbranch")));
+extern double sin (double) __attribute__ ((nothrow, leaf, simd ("notinbranch")));
+#ifdef __cplusplus
+}
+#endif
+double p[1024] = { 1.0 };
+double q[1024] = { 1.0 };
+
+void
+foo (void)
+{
+  int i;
+  for (i = 0; i < 1024; i++)
+    p[i] = cos (q[i]);
+}
+
+void
+bar (void)
+{
+  int i;
+  for (i = 0; i < 1024; i++)
+    p[i] = __builtin_sin (q[i]);
+}


	Jakub