powerpc: Add optimized ilogb* for POWER9

The instructions xsxexpdp and xsxexpqp introduced on POWER9 extract
the exponent from a double-precision and quad-precision floating-point
respectively, thus they can be used to improve ilogb, ilogbf and ilogbf128.

diff --git a/sysdeps/powerpc/fpu/math_private.h b/sysdeps/powerpc/fpu/math_private.h
index 91b13617499152548c66d68db1c6899d00a91010..21628f3bdad1dfa0dc25f3cc48ae486f1d8f879e 100644 (file)
--- a/sysdeps/powerpc/fpu/math_private.h
+++ b/sysdeps/powerpc/fpu/math_private.h
@@ -25,7 +25,28 @@
 
 #include_next <math_private.h>
 
-#if defined _ARCH_PWR9 && __HAVE_DISTINCT_FLOAT128
+#ifdef _ARCH_PWR9
+
+#if __GNUC_PREREQ (8, 0)
+# define _GL_HAS_BUILTIN_ILOGB 1
+#elif defined __has_builtin
+# define _GL_HAS_BUILTIN_ILOGB __has_builtin (__builtin_vsx_scalar_extract_exp)
+#else
+# define _GL_HAS_BUILTIN_ILOGB 0
+#endif
+
+#define __builtin_test_dc_ilogbf __builtin_test_dc_ilogb
+#define __builtin_ilogbf __builtin_ilogb
+
+#define __builtin_test_dc_ilogb(x, y) \
+        __builtin_vsx_scalar_test_data_class_dp(x, y)
+#define __builtin_ilogb(x) __builtin_vsx_scalar_extract_exp(x) - 0x3ff
+
+#define __builtin_test_dc_ilogbf128(x, y) \
+        __builtin_vsx_scalar_test_data_class_qp(x, y)
+#define __builtin_ilogbf128(x) __builtin_vsx_scalar_extract_expq(x) - 0x3fff
+
+#if __HAVE_DISTINCT_FLOAT128
 extern __always_inline _Float128
 __ieee754_sqrtf128 (_Float128 __x)
 {
@@ -34,5 +55,8 @@ __ieee754_sqrtf128 (_Float128 __x)
   return __z;
 }
 #endif
+#else /* !_ARCH_PWR9 */
+#define _GL_HAS_BUILTIN_ILOGB 0
+#endif
 
 #endif /* _PPC_MATH_PRIVATE_H_ */

diff --git a/sysdeps/powerpc/powerpc64/le/fpu/w_ilogb_template.c b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogb_template.c
new file mode 100644 (file)
index 0000000..b5c1c0a
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogb_template.c
@@ -0,0 +1,30 @@
+#include <math.h>
+#include <errno.h>
+#include <limits.h>
+#include <math_private.h>
+#include <fenv.h>
+
+#if _GL_HAS_BUILTIN_ILOGB
+int
+M_DECL_FUNC (__ilogb) (FLOAT x)
+{
+  int r;
+  /* Check for exceptional cases.  */
+  if (! M_SUF(__builtin_test_dc_ilogb) (x, 0x7f))
+    r = M_SUF (__builtin_ilogb) (x);
+  else
+    /* Fallback to the generic ilogb if x is NaN, Inf or subnormal.  */
+    r = M_SUF (__ieee754_ilogb) (x);
+  if (__builtin_expect (r == FP_ILOGB0, 0)
+      || __builtin_expect (r == FP_ILOGBNAN, 0)
+      || __builtin_expect (r == INT_MAX, 0))
+    {
+      __set_errno (EDOM);
+      __feraiseexcept (FE_INVALID);
+    }
+  return r;
+}
+declare_mgen_alias (__ilogb, ilogb)
+#else
+#include <math/w_ilogb_template.c>
+#endif

diff --git a/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
new file mode 100644 (file)
index 0000000..205f154
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/le/fpu/w_ilogbl.c
@@ -0,0 +1,4 @@
+/* Skip the optimization for long double as ibm128 does not provide an
+   optimized builtin. */
+#include <math-type-macros-ldouble.h>
+#include <math/w_ilogb_template.c>