Hi!
glibc for -ffast-math annotates a couple of math functions with simd
attribute, so that one can use vectorized versions with 4/8/16 vectorization
factor.
If one uses ::cos or ::cosf or std::cos(double), this works just fine, but not
when using std::cos(float). This is because the libstdc++ headers call
__builtin_cosf, but the builtin function doesn't have the simd attribute,
only ::cosf does.
Attached are 2 patches to improve this.
The first one is a C/C++ FE change, which arranges that if we add simd
attribute to say ::cosf, then calls to __builtin_cosf will act as if
__builtin_cosf also has the attribute. While other attributes aren't
handled this way, perhaps a small precedent to such change is that if
somebody uses typeof (cosf) cosf __asm ("foobar"); then calls to
__builtin_cosf if they expand into a library call will call foobar, not
cosf.
The other patch is instead a libstdc++ change, not using __builtin_cosf
etc., but ::cosf.
Both patches have been (separately) bootstrapped/regtested on x86_64-linux
and i686-linux.
Jakub
2017-08-07 Jakub Jelinek <ja...@redhat.com>
PR libstdc++/81706
* tree.c (attribute_value_equal): Use omp_declare_simd_clauses_equal
for comparison of OMP_CLAUSEs regardless of flag_openmp{,_simd}.
* 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/tree.c.jj 2017-07-29 09:48:40.000000000 +0200
+++ gcc/tree.c 2017-08-04 12:06:35.636072718 +0200
@@ -5022,8 +5022,8 @@ 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_VALUE (attr2)
&& TREE_CODE (TREE_VALUE (attr1)) == OMP_CLAUSE
&& TREE_CODE (TREE_VALUE (attr2)) == OMP_CLAUSE)
return omp_declare_simd_clauses_equal (TREE_VALUE (attr1),
--- gcc/c/c-decl.c.jj 2017-07-31 11:31:15.000000000 +0200
+++ gcc/c/c-decl.c 2017-08-04 12:39:48.113226134 +0200
@@ -2566,6 +2566,36 @@ 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)
+ {
+ tree s2 = lookup_attribute ("omp declare simd",
+ DECL_ATTRIBUTES (b));
+ while (s)
+ {
+ tree s3;
+ for (s3 = s2; s3;
+ s3 = lookup_attribute ("omp declare simd",
+ TREE_CHAIN (s3)))
+ if (attribute_value_equal (s, s3))
+ break;
+ if (!s3)
+ {
+ s3 = copy_node (s);
+ TREE_CHAIN (s3) = DECL_ATTRIBUTES (b);
+ DECL_ATTRIBUTES (b) = s3;
+ }
+ s = lookup_attribute ("omp declare simd",
+ TREE_CHAIN (s));
+ }
+ }
+ }
}
}
else
--- gcc/cp/decl.c.jj 2017-08-01 19:23:10.000000000 +0200
+++ gcc/cp/decl.c 2017-08-04 12:44:44.773780568 +0200
@@ -2456,6 +2456,35 @@ 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)
+ {
+ tree s2 = lookup_attribute ("omp declare simd",
+ DECL_ATTRIBUTES (b));
+ while (s)
+ {
+ tree s3;
+ for (s3 = s2; s3;
+ s3 = lookup_attribute ("omp declare simd",
+ TREE_CHAIN (s3)))
+ if (attribute_value_equal (s, s3))
+ break;
+ if (!s3)
+ {
+ s3 = copy_node (s);
+ TREE_CHAIN (s3) = DECL_ATTRIBUTES (b);
+ DECL_ATTRIBUTES (b) = s3;
+ }
+ s = lookup_attribute ("omp declare simd",
+ TREE_CHAIN (s));
+ }
+ }
+ }
}
if (new_defines_function)
/* If defining a function declared with other language
--- gcc/testsuite/gcc.target/i386/pr81706.c.jj 2017-08-06 23:50:46.511337565
+0200
+++ gcc/testsuite/gcc.target/i386/pr81706.c 2017-08-06 23:50:35.000000000
+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 2017-08-06 23:51:09.318065575
+0200
+++ gcc/testsuite/g++.dg/ext/pr81706.C 2017-08-06 23:51:38.577716630 +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]);
+}
2017-08-07 Jakub Jelinek <ja...@redhat.com>
PR libstdc++/81706
* include/c_global/cmath (std::cos, std::exp, std::log,
std::pow, std::sin): Call ::FNf instead of __builtin_FNf in
float overloads.
--- libstdc++-v3/include/c_global/cmath.jj 2017-07-24 10:57:58.000000000
+0200
+++ libstdc++-v3/include/c_global/cmath 2017-08-04 10:16:59.484637951 +0200
@@ -182,7 +182,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
#ifndef __CORRECT_ISO_CPP_MATH_H_PROTO
inline _GLIBCXX_CONSTEXPR float
cos(float __x)
- { return __builtin_cosf(__x); }
+ { return ::cosf(__x); }
inline _GLIBCXX_CONSTEXPR long double
cos(long double __x)
@@ -220,7 +220,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
#ifndef __CORRECT_ISO_CPP_MATH_H_PROTO
inline _GLIBCXX_CONSTEXPR float
exp(float __x)
- { return __builtin_expf(__x); }
+ { return ::expf(__x); }
inline _GLIBCXX_CONSTEXPR long double
exp(long double __x)
@@ -336,7 +336,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
#ifndef __CORRECT_ISO_CPP_MATH_H_PROTO
inline _GLIBCXX_CONSTEXPR float
log(float __x)
- { return __builtin_logf(__x); }
+ { return ::logf(__x); }
inline _GLIBCXX_CONSTEXPR long double
log(long double __x)
@@ -386,7 +386,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
#ifndef __CORRECT_ISO_CPP_MATH_H_PROTO
inline _GLIBCXX_CONSTEXPR float
pow(float __x, float __y)
- { return __builtin_powf(__x, __y); }
+ { return ::powf(__x, __y); }
inline _GLIBCXX_CONSTEXPR long double
pow(long double __x, long double __y)
@@ -423,7 +423,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
#ifndef __CORRECT_ISO_CPP_MATH_H_PROTO
inline _GLIBCXX_CONSTEXPR float
sin(float __x)
- { return __builtin_sinf(__x); }
+ { return ::sinf(__x); }
inline _GLIBCXX_CONSTEXPR long double
sin(long double __x)
