PowerPC - logb[f|l] optimization for POWER7

This patch provides optimized logb (1.2x on PPC32 and 2.5x on PPC64),
logbf (1.1x on PPC32 and 2.2x on PPC64), and logbl (1.3x on PPC32 and
50% on PPC64) for the POWER7 processor.

3 files changed, 207 insertions, 0 deletions

diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/s_logb.c b/sysdeps/powerpc/powerpc32/power7/fpu/s_logb.c
new file mode 100644
index 0000000000..71ee148149
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/power7/fpu/s_logb.c
@@ -0,0 +1,75 @@
+/* logb(). PowerPC/POWER7 version.
+   Copyright (C) 2012 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include "math_private.h"
+
+/* This implementation avoids FP to INT conversions by using VSX
+   bitwise instructions over FP values.  */
+
+static const double two1div52 = 2.220446049250313e-16;	/* 1/2**52  */
+static const double two10m1   = -1023.0;		/* 2**10 -1  */
+
+/* FP mask to extract the exponent.  */
+static const union {
+  unsigned long long mask;
+  double d;
+} mask = { 0x7ff0000000000000ULL };
+
+double
+__logb (double x)
+{
+  double ret;
+
+  if (__builtin_expect (x == 0.0, 0))
+    /* Raise FE_DIVBYZERO and return -HUGE_VAL[LF].  */
+    return -1.0 / __builtin_fabs (x);
+
+  /* ret = x & 0x7ff0000000000000;  */
+  asm (
+    "xxland %x0,%x1,%x2\n"
+    "fcfid  %0,%0"
+    : "=f" (ret)
+    : "f" (x), "f" (mask.d));
+  /* ret = (ret >> 52) - 1023.0;  */
+  ret = (ret * two1div52) + two10m1;
+  if (__builtin_expect (ret > -two10m1, 0))
+    /* Multiplication is used to set logb (+-INF) = INF.  */
+    return (x * x);
+  else if (__builtin_expect (ret == two10m1, 0))
+    {
+      /* POSIX specifies that denormal numbers are treated as
+         though they were normalized.  */
+      int32_t lx, ix;
+      int m1, m2, ma;
+
+      EXTRACT_WORDS (ix , lx, x);
+      m1 = (ix == 0) ? 0 : __builtin_clz (ix);
+      m2 = (lx == 0) ? 0 : __builtin_clz (lx);
+      ma = (m1 == 0) ? m2 + 32 : m1;
+      return -1022.0 + (double)(11 - ma);
+    }
+  /* Test to avoid logb_downward (0.0) == -0.0.  */
+  return ret == -0.0 ? 0.0 : ret;
+}
+
+weak_alias (__logb, logb)
+
+#ifdef NO_LONG_DOUBLE
+strong_alias (__logb, __logbl)
+weak_alias (__logb, logbl)
+#endif
diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/s_logbf.c b/sysdeps/powerpc/powerpc32/power7/fpu/s_logbf.c
new file mode 100644
index 0000000000..aa8499a14f
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/power7/fpu/s_logbf.c
@@ -0,0 +1,60 @@
+/* logbf(). PowerPC/POWER7 version.
+   Copyright (C) 2012 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include "math_private.h"
+
+/* This implementation avoids FP to INT conversions by using VSX
+   bitwise instructions over FP values.  */
+
+static const double two1div52 = 2.220446049250313e-16;	/* 1/2**52  */
+static const double two10m1   = -1023.0;		/* -2**10 + 1  */
+static const double two7m1    = -127.0;			/* -2**7 + 1  */
+
+/* FP mask to extract the exponent.  */
+static const union {
+  unsigned long long mask;
+  double d;
+} mask = { 0x7ff0000000000000ULL };
+
+float
+__logbf (float x)
+{
+  /* VSX operation are all done internally as double.  */
+  double ret;
+
+  if (__builtin_expect (x == 0.0, 0))
+    /* Raise FE_DIVBYZERO and return -HUGE_VAL[LF].  */
+    return -1.0 / __builtin_fabsf (x);
+
+  /* ret = x & 0x7f800000;  */
+  asm (
+    "xxland %x0,%x1,%x2\n"
+    "fcfid  %0,%0"
+    : "=f"(ret)
+    : "f" (x), "f" (mask.d));
+  /* ret = (ret >> 52) - 1023.0, since ret is double.  */
+  ret = (ret * two1div52) + two10m1;
+  if (__builtin_expect (ret > -two7m1, 0))
+    /* Multiplication is used to set logb (+-INF) = INF.  */
+    return (x * x);
+  /* Since operations are done with double we don't need
+     additional tests for subnormal numbers.
+     The test is to avoid logb_downward (0.0) == -0.0.  */
+  return ret == -0.0 ? 0.0 : ret;
+}
+weak_alias (__logbf, logbf)
diff --git a/sysdeps/powerpc/powerpc32/power7/fpu/s_logbl.c b/sysdeps/powerpc/powerpc32/power7/fpu/s_logbl.c
new file mode 100644
index 0000000000..03942ca53d
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/power7/fpu/s_logbl.c
@@ -0,0 +1,72 @@
+/* logbl(). PowerPC/POWER7 version.
+   Copyright (C) 2012 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <math.h>
+#include <math_private.h>
+#include <math_ldbl_opt.h>
+
+/* This implementation avoids FP to INT conversions by using VSX
+   bitwise instructions over FP values.  */
+
+static const double two1div52 = 2.220446049250313e-16;	/* 1/2**52  */
+static const double two10m1   = -1023.0;		/* 2**10 -1  */
+
+/* FP mask to extract the exponent.  */
+static const union {
+  unsigned long long mask;
+  double d;
+} mask = { 0x7ff0000000000000ULL };
+
+long double
+__logbl (long double x)
+{
+  double xh, xl;
+  double ret;
+
+  if (__builtin_expect (x == 0.0L, 0))
+    /* Raise FE_DIVBYZERO and return -HUGE_VAL[LF].  */
+    return -1.0L / __builtin_fabsl (x);
+
+  ldbl_unpack (x, &xh, &xl);
+  /* ret = x & 0x7ff0000000000000;  */
+  asm (
+    "xxland %x0,%x1,%x2\n"
+    "fcfid  %0,%0"
+    : "=f" (ret)
+    : "f" (xh), "f" (mask.d));
+  /* ret = (ret >> 52) - 1023.0;  */
+  ret = (ret * two1div52) + two10m1;
+  if (__builtin_expect (ret > -two10m1, 0))
+    /* Multiplication is used to set logb (+-INF) = INF.  */
+    return (xh * xh);
+  else if (__builtin_expect (ret == two10m1, 0))
+    {
+      int64_t lx, hx;
+      int m1, m2, ma;
+
+      GET_LDOUBLE_WORDS64 (hx, lx, x);
+      m1 = (hx == 0) ? 0 : __builtin_clzll (hx);
+      m2 = (lx == 0) ? 0 : __builtin_clzll (lx);
+      ma = (m1 == 0) ? m2 + 64 : m1;
+      return -1022.0 + (double)(11 - ma);
+    }
+  /* Test to avoid logb_downward (0.0) == -0.0.  */
+  return ret == -0.0 ? 0.0 : ret;
+}
+
+long_double_symbol (libm, __logbl, logbl);