This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
[PATCH] PowerPC - Optimization for nearbyint[f]
- From: Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>
- To: "GNU C. Library" <libc-alpha at sourceware dot org>
- Date: Wed, 30 Nov 2011 11:32:42 -0200
- Subject: [PATCH] PowerPC - Optimization for nearbyint[f]
This patch provides throughput boost for the nearbyint[f] functions
for POWER. For POWER7, it improves performance for nearbyintf from
5 (ppc32) to 6 times (ppc64) and for nearbyint from 2.5 up to 5
times. For POWER6 it improves nearbyintf up to 4x (ppc64) and
nearbyint up to 2x.
--
2011-11-18 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
* math/libm-test.inc: Added more nerabyint tests.
* sysdeps/powerpc/powerpc32/fpu/s_nearbyint.S: New file: nearbyint
optimized for PPC32.
* sysdeps/powerpc/powerpc32/fpu/s_nearbyintf.S: New file: nearbyintf
optimized for PPC32.
* sysdeps/powerpc/powerpc64/fpu/s_nearbyint.S: New file: nearbyint
optimized for PPC64.
* sysdeps/powerpc/powerpc64/fpu/s_nearbyintf.S: New file: nearbyintf
optimized for PPC64.
diff --git a/math/libm-test.inc b/math/libm-test.inc
index a451174..2843997 100644
--- a/math/libm-test.inc
+++ b/math/libm-test.inc
@@ -4618,6 +4618,10 @@ nearbyint_test (void)
TEST_f_f (nearbyint, minus_infty, minus_infty);
TEST_f_f (nearbyint, nan_value, nan_value);
+ /* Subnormal values */
+ TEST_f_f (nearbyint, -8.98847e+307, -8.98847e+307);
+ TEST_f_f (nearbyint, -4.45015e-308, minus_zero);
+
/* Default rounding mode is round to nearest. */
TEST_f_f (nearbyint, 0.5, 0.0);
TEST_f_f (nearbyint, 1.5, 2.0);
diff --git a/sysdeps/powerpc/powerpc32/fpu/s_nearbyint.S b/sysdeps/powerpc/powerpc32/fpu/s_nearbyint.S
new file mode 100644
index 0000000..0102aa3
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/s_nearbyint.S
@@ -0,0 +1,80 @@
+/* Round to int floating-point values. PowerPC32 version.
+ Copyright (C) 2011 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+ Contributed by Adhemerval Zanella <azanella@br.ibm.com>, 2011
+
+ 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, write to the Free
+ Software Foundation, Inc., 1 Franklin Street, Fifth Floor, Boston MA
+ 02110-1301 USA. */
+
+/* This has been coded in assembler because GCC makes such a mess of it
+ when it's coded in C. */
+
+#include <sysdep.h>
+#include <math_ldbl_opt.h>
+
+
+/* double [fp1] nearbyint(double [fp1] x) */
+
+ .section .rodata.cst4,"aM",@progbits,4
+ .align 2
+.LC0: /* 2**52 */
+ .long 0x59800000 /* TWO52: 2**52 */
+
+ .section ".text"
+ENTRY (__nearbyint)
+#ifdef SHARED
+ mflr r11
+ cfi_register(lr,r11)
+ bcl 20,31,1f
+1: mflr r9
+ addis r9,r9,.LC0-1b@ha
+ lfs fp13,.LC0-1b@l(r9)
+ mtlr r11
+ cfi_same_value (lr)
+#else
+ lis r9,.LC0@ha
+ lfs fp13,.LC0@l(r9)
+#endif
+ fabs fp0,fp1
+ fsub fp12,fp13,fp13 /* generate 0.0 */
+ fcmpu cr7,fp0,fp13 /* if (fabs(x) > TWO52 */
+ bgelr cr7
+ fcmpu cr7,fp1,fp12 /* if (x > 0.0 */
+ ble cr7,L(lessthanzero)
+ mtfsb0 4*cr7+lt /* Disable FE_INEXACT exception */
+ fadd fp0,fp1,fp13 /* x += TWO52 */
+ fsub fp1,fp0,fp13 /* x -= TWO52 */
+ fabs fp1,fp1 /* if (x == 0.0 */
+ mtfsb0 4*cr1+eq /* Clear any FE_INEXACT exception */
+ blr
+L(lessthanzero):
+ bgelr cr7
+ mtfsb0 4*cr7+lt /* Disable FE_INEXACT exception */
+ fsub fp0,fp13,fp1 /* x -= TWO52 */
+ fsub fp0,fp0,fp13 /* x += TWO52 */
+ fneg fp1,fp0 /* if (x == 0.0) */
+ mtfsb0 4*cr1+eq /* Clear any FE_INEXACT exception */
+ blr
+END (__nearbyint)
+
+weak_alias (__nearbyint, nearbyint)
+
+#ifdef NO_LONG_DOUBLE
+weak_alias (__nearbyint, nearbyintl)
+strong_alias (__nearbyint, __nearbyintl)
+#endif
+#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
+compat_symbol (libm, __nearbyint, nearbyintl, GLIBC_2_0)
+#endif
diff --git a/sysdeps/powerpc/powerpc32/fpu/s_nearbyintf.S b/sysdeps/powerpc/powerpc32/fpu/s_nearbyintf.S
new file mode 100644
index 0000000..5190228
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/s_nearbyintf.S
@@ -0,0 +1,70 @@
+/* Round to int floating-point values. PowerPC32 version.
+ Copyright (C) 2011 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+ Contributed by Adhemerval Zanella <azanella@br.ibm.com>, 2011
+
+ 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, write to the Free
+ Software Foundation, Inc., 1 Franklin Street, Fifth Floor, Boston MA
+ 02110-1301 USA. */
+
+/* This has been coded in assembler because GCC makes such a mess of it
+ when it's coded in C. */
+
+#include <sysdep.h>
+
+
+/* float [fp1] nearbyintf(float [fp1] x) */
+
+ .section .rodata.cst4,"aM",@progbits,4
+ .align 2
+.LC0:
+ .long 0x4B000000 /* TWO23: 2**23 */
+
+ .section ".text"
+ENTRY (__nearbyintf)
+#ifdef SHARED
+ mflr r11
+ cfi_register(lr,r11)
+ bcl 20,31,1f
+1: mflr r9
+ addis r9,r9,.LC0-1b@ha
+ lfs fp13,.LC0-1b@l(r9)
+ mtlr r11
+ cfi_same_value (lr)
+#else
+ lis r9,.LC0@ha
+ lfs fp13,.LC0@l(r9)
+#endif
+ fabs fp0,fp1
+ fsub fp12,fp13,fp13 /* generate 0.0 */
+ fcmpu cr7,fp0,fp13 /* if (fabs(x) > TWO23 */
+ bgelr cr7
+ fcmpu cr7,fp1,fp12 /* if (x > 0.0 */
+ ble cr7,L(lessthanzero)
+ mtfsb0 4*cr7+lt /* Disable FE_INEXACT exception */
+ fadds fp0,fp1,fp13 /* x += TWO23 */
+ fsubs fp1,fp0,fp13 /* x -= TWO23 */
+ mtfsb0 4*cr1+eq /* Clear any FE_INEXACT exception */
+ blr
+L(lessthanzero):
+ bgelr cr7
+ mtfsb0 4*cr7+lt /* Disable FE_INEXACT exception */
+ fsubs fp0,fp13,fp1 /* x -= TWO23 */
+ fsubs fp0,fp0,fp13 /* x += TWO23 */
+ fneg fp1,fp0 /* if (x == 0.0) */
+ mtfsb0 4*cr1+eq /* Clear any FE_INEXACT exception */
+ blr
+END (__nearbyintf)
+
+weak_alias (__nearbyintf, nearbyintf)
diff --git a/sysdeps/powerpc/powerpc64/fpu/s_nearbyint.S b/sysdeps/powerpc/powerpc64/fpu/s_nearbyint.S
new file mode 100644
index 0000000..436160f
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/fpu/s_nearbyint.S
@@ -0,0 +1,68 @@
+/* Round to int floating-point values. PowerPC64 version.
+ Copyright (C) 2011 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+ Contributed by Adhemerval Zanella <azanella@br.ibm.com>, 2011
+
+ 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, write to the Free
+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ 02111-1307 USA. */
+
+/* This has been coded in assembler because GCC makes such a mess of it
+ when it's coded in C. */
+
+#include <sysdep.h>
+#include <math_ldbl_opt.h>
+
+
+/* double [fp1] nearbyint(double [fp1] x) */
+
+ .section ".toc","aw"
+.LC0: /* 2**52 */
+ .tc FD_43300000_0[TC],0x4330000000000000
+ .section ".text"
+
+EALIGN (__nearbyint, 4, 0)
+ CALL_MCOUNT 0
+ fabs fp0,fp1
+ lfd fp13,.LC0@toc(2)
+ fcmpu cr7,fp0,fp13 /* if (fabs(x) > TWO52) */
+ bgelr cr7
+ fsub fp12,fp13,fp13 /* generate 0.0 */
+ fcmpu cr7,fp1,fp12 /* if (x > 0.0) */
+ ble cr7, L(lessthanzero)
+ mtfsb0 4*cr7+lt /* Disable FE_INEXACT exception */
+ fadd fp1,fp1,fp13 /* x+= TWO52 */
+ fsub fp1,fp1,fp13 /* x-= TWO52 */
+ fabs fp1,fp1 /* if (x == 0.0) */
+ mtfsb0 4*cr1+eq /* Clear any FE_INEXACT exception */
+ blr /* x = 0.0; */
+L(lessthanzero):
+ bgelr cr7 /* if (x < 0.0) */
+ mtfsb0 4*cr7+lt
+ fsub fp1,fp1,fp13 /* x -= TWO52 */
+ fadd fp1,fp1,fp13 /* x += TWO52 */
+ fnabs fp1,fp1 /* if (x == 0.0) */
+ mtfsb0 4*cr1+eq
+ blr /* x = -0.0; */
+END (__nearbyint)
+
+weak_alias (__nearbyint, nearbyint)
+
+#ifdef NO_LONG_DOUBLE
+weak_alias (__nearbyint, nearbyint)
+strong_alias (__nearbyint, __nearbyintl)
+#endif
+#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
+compat_symbol (libm, __nearbyint, nearbyintl, GLIBC_2_0)
+#endif
diff --git a/sysdeps/powerpc/powerpc64/fpu/s_nearbyintf.S b/sysdeps/powerpc/powerpc64/fpu/s_nearbyintf.S
new file mode 100644
index 0000000..c337e79
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/fpu/s_nearbyintf.S
@@ -0,0 +1,59 @@
+/* Round to int floating-point values. PowerPC64 version.
+ Copyright (C) 2011 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+ Contributed by Adhemerval Zanella <azanella@br.ibm.com>, 2011
+
+ 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, write to the Free
+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
+ 02111-1307 USA. */
+
+/* This has been coded in assembler because GCC makes such a mess of it
+ when it's coded in C. */
+
+#include <sysdep.h>
+
+
+/* float [fp1] nearbyintf(float [fp1]) */
+
+ .section ".toc","aw"
+.LC0: /* 2**23 */
+ .tc FD_4b000000_0[TC],0x4b00000000000000
+ .section ".text"
+
+EALIGN (__nearbyintf, 4, 0)
+ CALL_MCOUNT 0
+ fabs fp0,fp1
+ lfs fp13,.LC0@toc(2)
+ fcmpu cr7,fp0,fp13 /* if (fabs(x) > TWO52) */
+ bgelr cr7
+ fsubs fp12,fp13,fp13 /* generate 0.0 */
+ fcmpu cr7,fp1,fp12 /* if (x > 0.0) */
+ ble cr7, L(lessthanzero)
+ mtfsb0 4*cr7+lt /* Disable FE_INEXACT exception */
+ fadds fp1,fp1,fp13 /* x += TWO23 */
+ fsubs fp1,fp1,fp13 /* x -= TWO23 */
+ fabs fp1,fp1 /* if (x == 0.0) */
+ mtfsb0 4*cr1+eq /* Clear pending FE_INEXACT exception */
+ blr /* x = 0.0; */
+L(lessthanzero):
+ bgelr cr7 /* if (x < 0.0) */
+ mtfsb0 4*cr7+lt /* Disable FE_INEXACT exception */
+ fsubs fp1,fp1,fp13 /* x -= TWO23 */
+ fadds fp1,fp1,fp13 /* x += TWO23 */
+ fnabs fp1,fp1 /* if (x == 0.0) */
+ mtfsb0 4*cr1+eq /* Clear pending FE_INEXACT exception */
+ blr /* x = -0.0; */
+END (__nearbyintf)
+
+weak_alias (__nearbyintf, nearbyintf)
--
1.6.0.2