This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
[PATCH 23/36] PowerPC: sqrt/sqrtf multilib for PowerPC32
- From: Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>
- To: "GNU C. Library" <libc-alpha at sourceware dot org>
- Date: Mon, 19 Aug 2013 18:29:06 -0300
- Subject: [PATCH 23/36] PowerPC: sqrt/sqrtf multilib for PowerPC32
- References: <52127ABC dot 40008 at linux dot vnet dot ibm dot com>
2013-08-19 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
* sysdeps/powerpc/fpu/w_sqrt.c: Removed unneeded file.
* sysdeps/powerpc/powerpc32/power4/fpu/w_sqrt.S: Move to ...
* sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-power4.S: ... here.
(__sqrt): Rename symbol to __sqrt_power4 and remove strong_alias,
weak_alias and compat_symbol definitions.
* sysdeps/powerpc/powerpc32/power5/fpu/w_sqrt.S : Move to ...
* powerpc32/fpu/multiarch/w_sqrt-power5.S: ... here.
(__sqrt): Rename symbol to __sqrt_power5 and remove strong_alias,
weak_alias and compat_symbol definitions.
* sysdeps/powerpc/powerpc32/power4/fpu/w_sqrtf.S: Move to ...
* powerpc32/fpu/multiarch/w_sqrtf-power4.S: ... here.
(__sqrtf): Rename symbol to __sqrtf_power4 and remove weak_alias
definition.
* sysdeps/powerpc/powerpc32/power5/fpu/w_sqrtf.S: Move to ...
* powerpc32/fpu/multiarch/w_sqrtf-power5.S: ... here.
(__sqrtf): Rename symbol to __sqrtf_power5 and remove weak_alias
definition.
* sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-ppc32.c: New file:
default sqrt implementation for PPC32.
* sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-ppc32.c: New file:
default sqrtf implementation for PPC32.
* sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt.c: New file:
multiarch sqrt for PPC32.
* sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf.c: New file:
multiarch sqrtf for PPC32.
* sysdeps/powerpc/powerpc32/fpu/multiarch/Makefile: Add sqrt and
sqrtf objects.
--
diff --git a/sysdeps/powerpc/fpu/w_sqrt.c b/sysdeps/powerpc/fpu/w_sqrt.c
deleted file mode 100644
index 70f28dd..0000000
--- a/sysdeps/powerpc/fpu/w_sqrt.c
+++ /dev/null
@@ -1,48 +0,0 @@
-/* Double-precision floating point square root wrapper.
- Copyright (C) 2004-2013 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 <fenv_libc.h>
-#include <math_ldbl_opt.h>
-
-double
-__sqrt (double x) /* wrapper sqrt */
-{
-#ifdef _IEEE_LIBM
- return __ieee754_sqrt (x);
-#else
- double z;
- z = __ieee754_sqrt (x);
- if (_LIB_VERSION == _IEEE_ || (x != x))
- return z;
-
- if (x < 0.0)
- return __kernel_standard (x, x, 26); /* sqrt(negative) */
- else
- return z;
-#endif
-}
-
-weak_alias (__sqrt, sqrt)
-#ifdef NO_LONG_DOUBLE
- strong_alias (__sqrt, __sqrtl) weak_alias (__sqrt, sqrtl)
-#endif
-#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
-compat_symbol (libm, __sqrt, sqrtl, GLIBC_2_0);
-#endif
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/Makefile b/sysdeps/powerpc/powerpc32/fpu/multiarch/Makefile
index 7c50805..22e07ce 100644
--- a/sysdeps/powerpc/powerpc32/fpu/multiarch/Makefile
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/Makefile
@@ -2,7 +2,9 @@ ifeq ($(subdir),math)
libm-sysdep_routines += s_llrintf-power6 s_llrintf-power4 s_llrintf-ppc32 \
s_llrint-power6 s_llrint-power4 s_llrint-ppc32 \
s_llround-power6 s_llround-power5+ s_llround-power4 \
- s_llround-ppc32
+ s_llround-ppc32 w_sqrt-power5 w_sqrt-power4 \
+ w_sqrt-ppc32 w_sqrtf-power5 w_sqrtf-power4 \
+ w_sqrtf-ppc32
CFLAGS-s_llround.c = -fno-builtin-llroundf
endif
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-power4.S b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-power4.S
new file mode 100644
index 0000000..35e638f
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-power4.S
@@ -0,0 +1,98 @@
+/* sqrt function. PowerPC32/POWER4 version.
+ Copyright (C) 2007-2013 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 <sysdep.h>
+#include <math_ldbl_opt.h>
+
+/* double [fp1] sqrt (double x [fp1])
+ Power4 (ISA V2.0) and above implement sqrt in hardware (not optional).
+ The fsqrt instruction generates the correct value for all inputs and
+ sets the appropriate floating point exceptions. Extended checking is
+ only needed to set errno (via __kernel_standard) if the input value
+ is negative.
+
+ The fsqrt will set FPCC and FU (Floating Point Unordered or NaN
+ to indicated that the input value was negative or NaN. Use Move to
+ Condition Register from FPSCR to copy the FPCC field to cr1. The
+ branch on summary overflow transfers control to w_sqrt to process
+ any error conditions. Otherwise we can return the result directly.
+
+ This part of the function is a leaf routine, so no need to stack a
+ frame or execute prologue/epilogue code. This means it is safe to
+ transfer directly to w_sqrt as long as the input value (f1) is
+ preserved. Putting the sqrt result into f2 (double parameter 2)
+ allows passing both the input value and sqrt result into the extended
+ wrapper so there is no need to recompute.
+
+ This tactic avoids the overhead of stacking a frame for the normal
+ (non-error) case. Until gcc supports prologue shrink-wrapping
+ this is the best we can do. */
+
+ .section ".text"
+ .machine power4
+EALIGN (__sqrt_power4, 5, 0)
+ fsqrt fp2,fp1
+ mcrfs cr1,4
+ bso- cr1,.Lw_sqrt
+ fmr fp1,fp2
+ blr
+ .align 4
+.Lw_sqrt:
+ mflr r0
+ stwu r1,-16(r1)
+ cfi_adjust_cfa_offset(16)
+ fmr fp12,fp2
+ stw r0,20(r1)
+ stw r30,8(r1)
+ cfi_offset(lr,20-16)
+ cfi_offset(r30,8-16)
+#ifdef SHARED
+ SETUP_GOT_ACCESS(r30,got_label)
+ addis r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@ha
+ addi r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@l
+ lwz r9,_LIB_VERSION@got(30)
+ lwz r0,0(r9)
+#else
+ lis r9,_LIB_VERSION@ha
+ lwz r0,_LIB_VERSION@l(r9)
+#endif
+/* if (_LIB_VERSION == _IEEE_) return z; */
+ cmpwi cr7,r0,-1
+ beq- cr7,.L4
+/* if (x != x) return z; !isnan*/
+ fcmpu cr7,fp1,fp1
+ bne- cr7,.L4
+/* if (x < 0.0)
+ return __kernel_standard (x, x, 26) */
+ fmr fp2,fp1
+ fabs fp0,fp1
+ li r3,26
+ fcmpu cr7,fp1,fp0
+ bne- cr7,.L11
+.L4:
+ lwz r0,20(r1)
+ fmr fp1,fp12
+ lwz r30,8(r1)
+ addi r1,r1,16
+ mtlr r0
+ blr
+.L11:
+ bl __kernel_standard@plt
+ fmr fp12,fp1
+ b .L4
+END(__sqrt_power4)
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-power5.S b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-power5.S
new file mode 100644
index 0000000..fcb3043
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-power5.S
@@ -0,0 +1,96 @@
+/* sqrt function. PowerPC32/POWER5 version.
+ Copyright (C) 2007-2013 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 <sysdep.h>
+#include <math_ldbl_opt.h>
+
+/* double [fp1] sqrt (double x [fp1])
+ Power4 (ISA V2.0) and above implement sqrt in hardware (not optional).
+ The fsqrt instruction generates the correct value for all inputs and
+ sets the appropriate floating point exceptions. Extended checking is
+ only needed to set errno (via __kernel_standard) if the input value
+ is negative.
+
+ So compare the input value against the absolute value of itself.
+ This will compare equal unless the value is negative (EDOM) or a NAN,
+ in which case we branch to the extend wrapper. If equal we can return
+ the result directly.
+
+ This part of the function looks like a leaf routine, so no need to
+ stack a frame or execute prologue/epilogue code. It is safe to
+ branch directly to w_sqrt as long as the input value (f1) is
+ preserved. Putting the sqrt result into f2 (float parameter 2)
+ allows passing both the input value and sqrt result into the extended
+ wrapper so there is no need to recompute.
+
+ This tactic avoids the overhead of stacking a frame for the normal
+ (non-error) case. Until gcc supports prologue shrink-wrapping
+ this is the best we can do. */
+
+ .section ".text"
+ .machine power5
+EALIGN (__sqrt_power5, 5, 0)
+ fabs fp0,fp1
+ fsqrt fp2,fp1
+ fcmpu cr1,fp0,fp1
+ bne- cr1,.Lw_sqrt
+ fmr fp1,fp2
+ blr
+ .align 4
+.Lw_sqrt:
+ mflr r0
+ stwu r1,-16(r1)
+ cfi_adjust_cfa_offset(16)
+ fmr fp12,fp2
+ stw r0,20(r1)
+ stw r30,8(r1)
+ cfi_offset(lr,20-16)
+ cfi_offset(r30,8-16)
+#ifdef SHARED
+ SETUP_GOT_ACCESS(r30,got_label)
+ addis r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@ha
+ addi r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@l
+ lwz r9,_LIB_VERSION@got(30)
+ lwz r0,0(r9)
+#else
+ lis r9,_LIB_VERSION@ha
+ lwz r0,_LIB_VERSION@l(r9)
+#endif
+/* if (_LIB_VERSION == _IEEE_) return z; */
+ cmpwi cr7,r0,-1
+ beq- cr7,.L4
+/* if (x != x) return z; !isnan*/
+ fcmpu cr7,fp1,fp1
+ bne- cr7,.L4
+/* if (x < 0.0)
+ return __kernel_standard (x, x, 26) */
+ fmr fp2,fp1
+ li r3,26
+ bne- cr1,.L11
+.L4:
+ lwz r0,20(r1)
+ fmr fp1,fp12
+ lwz r30,8(r1)
+ addi r1,r1,16
+ mtlr r0
+ blr
+.L11:
+ bl __kernel_standard@plt
+ fmr fp12,fp1
+ b .L4
+END (__sqrt_power5)
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-ppc32.c b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-ppc32.c
new file mode 100644
index 0000000..5d07a7b
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt-ppc32.c
@@ -0,0 +1,30 @@
+/* PowerPC32 default implementation for sqrt.
+ Copyright (C) 2013 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>
+
+/* The PPC32 default implementation will fallback to __ieee754_sqrt symbol
+ from sysdeps/powerpc/fpu/e_sqrt.c */
+
+#define __sqrtf __sqrtf_ppc32
+#undef weak_alias
+#define weak_alias(a, b)
+#undef strong_alias
+#define strong_alias(a, b)
+
+#include <math/w_sqrtf.c>
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt.c b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt.c
new file mode 100644
index 0000000..f77eeb2
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrt.c
@@ -0,0 +1,43 @@
+/* Multiple versions of sqrt.
+ Copyright (C) 2013 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_ldbl_opt.h>
+#include <shlib-compat.h>
+#include "init-arch.h"
+
+extern __typeof (__sqrt) __sqrt_ppc32 attribute_hidden;
+extern __typeof (__sqrt) __sqrt_power4 attribute_hidden;
+extern __typeof (__sqrt) __sqrt_power5 attribute_hidden;
+
+libc_ifunc (__sqrt,
+ (hwcap & PPC_FEATURE_POWER5)
+ ? __sqrt_power5 :
+ (hwcap & PPC_FEATURE_POWER4)
+ ? __sqrt_power4
+ : __sqrt_ppc32);
+
+weak_alias (__sqrt, sqrt)
+
+#ifdef NO_LONG_DOUBLE
+strong_alias (__sqrt, __sqrtl)
+weak_alias (__sqrt, sqrtl)
+#endif
+#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
+compat_symbol (libm, __sqrt, sqrtl, GLIBC_2_0);
+#endif
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-power4.S b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-power4.S
new file mode 100644
index 0000000..f3524a2
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-power4.S
@@ -0,0 +1,98 @@
+/* sqrtf function. PowerPC32/POWER4 version.
+ Copyright (C) 2007-2013 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 <sysdep.h>
+#include <math_ldbl_opt.h>
+
+/* float [fp1] sqrts (float x [fp1])
+ Power4 (ISA V2.0) and above implement sqrt in hardware (not optional).
+ The fsqrts instruction generates the correct value for all inputs and
+ sets the appropriate floating point exceptions. Extended checking is
+ only needed to set errno (via __kernel_standard) if the input value
+ is negative.
+
+ The fsqrts will set FPCC and FU (Floating Point Unordered or NaN
+ to indicated that the input value was negative or NaN. Use Move to
+ Condition Register from FPSCR to copy the FPCC field to cr1. The
+ branch on summary overflow transfers control to w_sqrt to process
+ any error conditions. Otherwise we can return the result directly.
+
+ This part of the function is a leaf routine, so no need to stack a
+ frame or execute prologue/epilogue code. This means it is safe to
+ transfer directly to w_sqrt as long as the input value (f1) is
+ preserved. Putting the sqrt result into f2 (float parameter 2)
+ allows passing both the input value and sqrt result into the extended
+ wrapper so there is no need to recompute.
+
+ This tactic avoids the overhead of stacking a frame for the normal
+ (non-error) case. Until gcc supports prologue shrink-wrapping
+ this is the best we can do. */
+
+ .section ".text"
+ .machine power4
+EALIGN (__sqrtf_power4, 5, 0)
+ fsqrts fp2,fp1
+ mcrfs cr1,4
+ bso- cr1,.Lw_sqrtf
+ fmr fp1,fp2
+ blr
+ .align 4
+.Lw_sqrtf:
+ mflr r0
+ stwu r1,-16(r1)
+ cfi_adjust_cfa_offset(16)
+ fmr fp12,fp2
+ stw r0,20(r1)
+ stw r30,8(r1)
+ cfi_offset(lr,20-16)
+ cfi_offset(r30,8-16)
+#ifdef SHARED
+ SETUP_GOT_ACCESS(r30,got_label)
+ addis r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@ha
+ addi r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@l
+ lwz r9,_LIB_VERSION@got(30)
+ lwz r0,0(r9)
+#else
+ lis r9,_LIB_VERSION@ha
+ lwz r0,_LIB_VERSION@l(r9)
+#endif
+/* if (_LIB_VERSION == _IEEE_) return z; */
+ cmpwi cr7,r0,-1
+ beq- cr7,.L4
+/* if (x != x, 0) return z; !isnan */
+ fcmpu cr7,fp1,fp1
+ bne- cr7,.L4
+/* if (x < 0.0)
+ return __kernel_standard (x, x, 126) */
+ fmr fp2,fp1
+ fabs fp0,fp1
+ li r3,126
+ fcmpu cr7,1,0
+ bne- cr7,.L11
+.L4:
+ lwz r0,20(r1)
+ fmr fp1,fp12
+ lwz r30,8(r1)
+ addi r1,r1,16
+ mtlr r0
+ blr
+.L11:
+ bl __kernel_standard@plt
+ fmr fp12,fp1
+ b .L4
+END (__sqrtf_power4)
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-power5.S b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-power5.S
new file mode 100644
index 0000000..aae9e91
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-power5.S
@@ -0,0 +1,96 @@
+/* sqrtf function. PowerPC32/POWER5 version.
+ Copyright (C) 2007-2013 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 <sysdep.h>
+#include <math_ldbl_opt.h>
+
+/* float [fp1] sqrts (float x [fp1])
+ Power4 (ISA V2.0) and above implement sqrt in hardware (not optional).
+ The fsqrts instruction generates the correct value for all inputs and
+ sets the appropriate floating point exceptions. Extended checking is
+ only needed to set errno (via __kernel_standard) if the input value
+ is negative.
+
+ So compare the input value against the absolute value of itself.
+ This will compare equal unless the value is negative (EDOM) or a NAN,
+ in which case we branch to the extend wrapper. If equal we can return
+ the result directly.
+
+ This part of the function looks like a leaf routine, so no need to
+ stack a frame or execute prologue/epilogue code. It is safe to
+ branch directly to w_sqrt as long as the input value (f1) is
+ preserved. Putting the sqrt result into f2 (float parameter 2)
+ allows passing both the input value and sqrt result into the extended
+ wrapper so there is no need to recompute.
+
+ This tactic avoids the overhead of stacking a frame for the normal
+ (non-error) case. Until gcc supports prologue shrink-wrapping
+ this is the best we can do. */
+
+ .section ".text"
+ .machine power5
+EALIGN (__sqrtf_power5, 5, 0)
+ fabs fp0,fp1
+ fsqrts fp2,fp1
+ fcmpu cr1,fp0,fp1
+ bne- cr1,.Lw_sqrtf
+ fmr fp1,fp2
+ blr
+ .align 4
+.Lw_sqrtf:
+ mflr r0
+ stwu r1,-16(r1)
+ cfi_adjust_cfa_offset(16)
+ fmr fp12,fp2
+ stw r0,20(r1)
+ stw r30,8(r1)
+ cfi_offset(lr,20-16)
+ cfi_offset(r30,8-16)
+#ifdef SHARED
+ SETUP_GOT_ACCESS(r30,got_label)
+ addis r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@ha
+ addi r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@l
+ lwz r9,_LIB_VERSION@got(30)
+ lwz r0,0(r9)
+#else
+ lis r9,_LIB_VERSION@ha
+ lwz r0,_LIB_VERSION@l(r9)
+#endif
+/* if (_LIB_VERSION == _IEEE_) return z; */
+ cmpwi cr7,r0,-1
+ beq- cr7,.L4
+/* if (x != x, 0) return z; !isnan */
+ fcmpu cr7,fp1,fp1
+ bne- cr7,.L4
+/* if (x < 0.0)
+ return __kernel_standard (x, x, 126) */
+ fmr fp2,fp1
+ li r3,126
+ bne- cr1,.L11
+.L4:
+ lwz r0,20(r1)
+ fmr fp1,fp12
+ lwz r30,8(r1)
+ addi r1,r1,16
+ mtlr r0
+ blr
+.L11:
+ bl __kernel_standard@plt
+ fmr fp12,fp1
+ b .L4
+END (__sqrtf_power5)
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-ppc32.c b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-ppc32.c
new file mode 100644
index 0000000..3702e44
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf-ppc32.c
@@ -0,0 +1,30 @@
+/* PowerPC32 default implementation for sqrtf.
+ Copyright (C) 2013 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>
+
+/* The PPC32 default implementation will fallback to __ieee754_sqrt symbol
+ from sysdeps/powerpc/fpu/e_sqrt.c */
+
+#define __sqrt __sqrt_ppc32
+#undef weak_alias
+#define weak_alias(a, b)
+#undef strong_alias
+#define strong_alias(a, b)
+
+#include <math/w_sqrt.c>
diff --git a/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf.c b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf.c
new file mode 100644
index 0000000..8af86c4
--- /dev/null
+++ b/sysdeps/powerpc/powerpc32/fpu/multiarch/w_sqrtf.c
@@ -0,0 +1,35 @@
+/* Multiple versions of sqrtf.
+ Copyright (C) 2013 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_ldbl_opt.h>
+#include <shlib-compat.h>
+#include "init-arch.h"
+
+extern __typeof (__sqrtf) __sqrtf_ppc32 attribute_hidden;
+extern __typeof (__sqrtf) __sqrtf_power4 attribute_hidden;
+extern __typeof (__sqrtf) __sqrtf_power5 attribute_hidden;
+
+libc_ifunc (__sqrtf,
+ (hwcap & PPC_FEATURE_POWER5)
+ ? __sqrtf_power5 :
+ (hwcap & PPC_FEATURE_POWER4)
+ ? __sqrtf_power4
+ : __sqrtf_ppc32);
+
+weak_alias (__sqrtf, sqrtf)
diff --git a/sysdeps/powerpc/powerpc32/power4/fpu/w_sqrt.S b/sysdeps/powerpc/powerpc32/power4/fpu/w_sqrt.S
deleted file mode 100644
index 3648e4a..0000000
--- a/sysdeps/powerpc/powerpc32/power4/fpu/w_sqrt.S
+++ /dev/null
@@ -1,109 +0,0 @@
-/* sqrt function. PowerPC32 version.
- Copyright (C) 2007-2013 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 <sysdep.h>
-#include <math_ldbl_opt.h>
-
-/* double [fp1] sqrt (double x [fp1])
- Power4 (ISA V2.0) and above implement sqrt in hardware (not optional).
- The fsqrt instruction generates the correct value for all inputs and
- sets the appropriate floating point exceptions. Extended checking is
- only needed to set errno (via __kernel_standard) if the input value
- is negative.
-
- The fsqrt will set FPCC and FU (Floating Point Unordered or NaN
- to indicated that the input value was negative or NaN. Use Move to
- Condition Register from FPSCR to copy the FPCC field to cr1. The
- branch on summary overflow transfers control to w_sqrt to process
- any error conditions. Otherwise we can return the result directly.
-
- This part of the function is a leaf routine, so no need to stack a
- frame or execute prologue/epilogue code. This means it is safe to
- transfer directly to w_sqrt as long as the input value (f1) is
- preserved. Putting the sqrt result into f2 (double parameter 2)
- allows passing both the input value and sqrt result into the extended
- wrapper so there is no need to recompute.
-
- This tactic avoids the overhead of stacking a frame for the normal
- (non-error) case. Until gcc supports prologue shrink-wrapping
- this is the best we can do. */
-
- .section ".text"
- .machine power4
-EALIGN (__sqrt, 5, 0)
- fsqrt fp2,fp1
- mcrfs cr1,4
- bso- cr1,.Lw_sqrt
- fmr fp1,fp2
- blr
- .align 4
-.Lw_sqrt:
- mflr r0
- stwu r1,-16(r1)
- cfi_adjust_cfa_offset(16)
- fmr fp12,fp2
- stw r0,20(r1)
- stw r30,8(r1)
- cfi_offset(lr,20-16)
- cfi_offset(r30,8-16)
-#ifdef SHARED
- SETUP_GOT_ACCESS(r30,got_label)
- addis r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@ha
- addi r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@l
- lwz r9,_LIB_VERSION@got(30)
- lwz r0,0(r9)
-#else
- lis r9,_LIB_VERSION@ha
- lwz r0,_LIB_VERSION@l(r9)
-#endif
-/* if (_LIB_VERSION == _IEEE_) return z; */
- cmpwi cr7,r0,-1
- beq- cr7,.L4
-/* if (x != x) return z; !isnan*/
- fcmpu cr7,fp1,fp1
- bne- cr7,.L4
-/* if (x < 0.0)
- return __kernel_standard (x, x, 26) */
- fmr fp2,fp1
- fabs fp0,fp1
- li r3,26
- fcmpu cr7,fp1,fp0
- bne- cr7,.L11
-.L4:
- lwz r0,20(r1)
- fmr fp1,fp12
- lwz r30,8(r1)
- addi r1,r1,16
- mtlr r0
- blr
-.L11:
- bl __kernel_standard@plt
- fmr fp12,fp1
- b .L4
- END (__sqrt)
-
-weak_alias (__sqrt, sqrt)
-
-#ifdef NO_LONG_DOUBLE
-weak_alias (__sqrt, sqrtl)
-strong_alias (__sqrt, __sqrtl)
-#endif
-#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
-compat_symbol (libm, __sqrt, sqrtl, GLIBC_2_0)
-#endif
-
diff --git a/sysdeps/powerpc/powerpc32/power4/fpu/w_sqrtf.S b/sysdeps/powerpc/powerpc32/power4/fpu/w_sqrtf.S
deleted file mode 100644
index 153843c..0000000
--- a/sysdeps/powerpc/powerpc32/power4/fpu/w_sqrtf.S
+++ /dev/null
@@ -1,101 +0,0 @@
-/* sqrtf function. PowerPC32 version.
- Copyright (C) 2007-2013 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 <sysdep.h>
-#include <math_ldbl_opt.h>
-
-/* float [fp1] sqrts (float x [fp1])
- Power4 (ISA V2.0) and above implement sqrt in hardware (not optional).
- The fsqrts instruction generates the correct value for all inputs and
- sets the appropriate floating point exceptions. Extended checking is
- only needed to set errno (via __kernel_standard) if the input value
- is negative.
-
- The fsqrts will set FPCC and FU (Floating Point Unordered or NaN
- to indicated that the input value was negative or NaN. Use Move to
- Condition Register from FPSCR to copy the FPCC field to cr1. The
- branch on summary overflow transfers control to w_sqrt to process
- any error conditions. Otherwise we can return the result directly.
-
- This part of the function is a leaf routine, so no need to stack a
- frame or execute prologue/epilogue code. This means it is safe to
- transfer directly to w_sqrt as long as the input value (f1) is
- preserved. Putting the sqrt result into f2 (float parameter 2)
- allows passing both the input value and sqrt result into the extended
- wrapper so there is no need to recompute.
-
- This tactic avoids the overhead of stacking a frame for the normal
- (non-error) case. Until gcc supports prologue shrink-wrapping
- this is the best we can do. */
-
- .section ".text"
- .machine power4
-EALIGN (__sqrtf, 5, 0)
- fsqrts fp2,fp1
- mcrfs cr1,4
- bso- cr1,.Lw_sqrtf
- fmr fp1,fp2
- blr
- .align 4
-.Lw_sqrtf:
- mflr r0
- stwu r1,-16(r1)
- cfi_adjust_cfa_offset(16)
- fmr fp12,fp2
- stw r0,20(r1)
- stw r30,8(r1)
- cfi_offset(lr,20-16)
- cfi_offset(r30,8-16)
-#ifdef SHARED
- SETUP_GOT_ACCESS(r30,got_label)
- addis r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@ha
- addi r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@l
- lwz r9,_LIB_VERSION@got(30)
- lwz r0,0(r9)
-#else
- lis r9,_LIB_VERSION@ha
- lwz r0,_LIB_VERSION@l(r9)
-#endif
-/* if (_LIB_VERSION == _IEEE_) return z; */
- cmpwi cr7,r0,-1
- beq- cr7,.L4
-/* if (x != x, 0) return z; !isnan */
- fcmpu cr7,fp1,fp1
- bne- cr7,.L4
-/* if (x < 0.0)
- return __kernel_standard (x, x, 126) */
- fmr fp2,fp1
- fabs fp0,fp1
- li r3,126
- fcmpu cr7,1,0
- bne- cr7,.L11
-.L4:
- lwz r0,20(r1)
- fmr fp1,fp12
- lwz r30,8(r1)
- addi r1,r1,16
- mtlr r0
- blr
-.L11:
- bl __kernel_standard@plt
- fmr fp12,fp1
- b .L4
- END (__sqrtf)
-
-weak_alias (__sqrtf, sqrtf)
-
diff --git a/sysdeps/powerpc/powerpc32/power5/fpu/w_sqrt.S b/sysdeps/powerpc/powerpc32/power5/fpu/w_sqrt.S
deleted file mode 100644
index ed11d5a..0000000
--- a/sysdeps/powerpc/powerpc32/power5/fpu/w_sqrt.S
+++ /dev/null
@@ -1,107 +0,0 @@
-/* sqrt function. PowerPC32 version.
- Copyright (C) 2007-2013 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 <sysdep.h>
-#include <math_ldbl_opt.h>
-
-/* double [fp1] sqrt (double x [fp1])
- Power4 (ISA V2.0) and above implement sqrt in hardware (not optional).
- The fsqrt instruction generates the correct value for all inputs and
- sets the appropriate floating point exceptions. Extended checking is
- only needed to set errno (via __kernel_standard) if the input value
- is negative.
-
- So compare the input value against the absolute value of itself.
- This will compare equal unless the value is negative (EDOM) or a NAN,
- in which case we branch to the extend wrapper. If equal we can return
- the result directly.
-
- This part of the function looks like a leaf routine, so no need to
- stack a frame or execute prologue/epilogue code. It is safe to
- branch directly to w_sqrt as long as the input value (f1) is
- preserved. Putting the sqrt result into f2 (float parameter 2)
- allows passing both the input value and sqrt result into the extended
- wrapper so there is no need to recompute.
-
- This tactic avoids the overhead of stacking a frame for the normal
- (non-error) case. Until gcc supports prologue shrink-wrapping
- this is the best we can do. */
-
- .section ".text"
- .machine power4
-EALIGN (__sqrt, 5, 0)
- fabs fp0,fp1
- fsqrt fp2,fp1
- fcmpu cr1,fp0,fp1
- bne- cr1,.Lw_sqrt
- fmr fp1,fp2
- blr
- .align 4
-.Lw_sqrt:
- mflr r0
- stwu r1,-16(r1)
- cfi_adjust_cfa_offset(16)
- fmr fp12,fp2
- stw r0,20(r1)
- stw r30,8(r1)
- cfi_offset(lr,20-16)
- cfi_offset(r30,8-16)
-#ifdef SHARED
- SETUP_GOT_ACCESS(r30,got_label)
- addis r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@ha
- addi r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@l
- lwz r9,_LIB_VERSION@got(30)
- lwz r0,0(r9)
-#else
- lis r9,_LIB_VERSION@ha
- lwz r0,_LIB_VERSION@l(r9)
-#endif
-/* if (_LIB_VERSION == _IEEE_) return z; */
- cmpwi cr7,r0,-1
- beq- cr7,.L4
-/* if (x != x) return z; !isnan*/
- fcmpu cr7,fp1,fp1
- bne- cr7,.L4
-/* if (x < 0.0)
- return __kernel_standard (x, x, 26) */
- fmr fp2,fp1
- li r3,26
- bne- cr1,.L11
-.L4:
- lwz r0,20(r1)
- fmr fp1,fp12
- lwz r30,8(r1)
- addi r1,r1,16
- mtlr r0
- blr
-.L11:
- bl __kernel_standard@plt
- fmr fp12,fp1
- b .L4
- END (__sqrt)
-
-weak_alias (__sqrt, sqrt)
-
-#ifdef NO_LONG_DOUBLE
-weak_alias (__sqrt, sqrtl)
-strong_alias (__sqrt, __sqrtl)
-#endif
-#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
-compat_symbol (libm, __sqrt, sqrtl, GLIBC_2_0)
-#endif
-
diff --git a/sysdeps/powerpc/powerpc32/power5/fpu/w_sqrtf.S b/sysdeps/powerpc/powerpc32/power5/fpu/w_sqrtf.S
deleted file mode 100644
index 2049172..0000000
--- a/sysdeps/powerpc/powerpc32/power5/fpu/w_sqrtf.S
+++ /dev/null
@@ -1,99 +0,0 @@
-/* sqrtf function. PowerPC32 version.
- Copyright (C) 2007-2013 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 <sysdep.h>
-#include <math_ldbl_opt.h>
-
-/* float [fp1] sqrts (float x [fp1])
- Power4 (ISA V2.0) and above implement sqrt in hardware (not optional).
- The fsqrts instruction generates the correct value for all inputs and
- sets the appropriate floating point exceptions. Extended checking is
- only needed to set errno (via __kernel_standard) if the input value
- is negative.
-
- So compare the input value against the absolute value of itself.
- This will compare equal unless the value is negative (EDOM) or a NAN,
- in which case we branch to the extend wrapper. If equal we can return
- the result directly.
-
- This part of the function looks like a leaf routine, so no need to
- stack a frame or execute prologue/epilogue code. It is safe to
- branch directly to w_sqrt as long as the input value (f1) is
- preserved. Putting the sqrt result into f2 (float parameter 2)
- allows passing both the input value and sqrt result into the extended
- wrapper so there is no need to recompute.
-
- This tactic avoids the overhead of stacking a frame for the normal
- (non-error) case. Until gcc supports prologue shrink-wrapping
- this is the best we can do. */
-
- .section ".text"
- .machine power4
-EALIGN (__sqrtf, 5, 0)
- fabs fp0,fp1
- fsqrts fp2,fp1
- fcmpu cr1,fp0,fp1
- bne- cr1,.Lw_sqrtf
- fmr fp1,fp2
- blr
- .align 4
-.Lw_sqrtf:
- mflr r0
- stwu r1,-16(r1)
- cfi_adjust_cfa_offset(16)
- fmr fp12,fp2
- stw r0,20(r1)
- stw r30,8(r1)
- cfi_offset(lr,20-16)
- cfi_offset(r30,8-16)
-#ifdef SHARED
- SETUP_GOT_ACCESS(r30,got_label)
- addis r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@ha
- addi r30,r30,_GLOBAL_OFFSET_TABLE_-got_label@l
- lwz r9,_LIB_VERSION@got(30)
- lwz r0,0(r9)
-#else
- lis r9,_LIB_VERSION@ha
- lwz r0,_LIB_VERSION@l(r9)
-#endif
-/* if (_LIB_VERSION == _IEEE_) return z; */
- cmpwi cr7,r0,-1
- beq- cr7,.L4
-/* if (x != x, 0) return z; !isnan */
- fcmpu cr7,fp1,fp1
- bne- cr7,.L4
-/* if (x < 0.0)
- return __kernel_standard (x, x, 126) */
- fmr fp2,fp1
- li r3,126
- bne- cr1,.L11
-.L4:
- lwz r0,20(r1)
- fmr fp1,fp12
- lwz r30,8(r1)
- addi r1,r1,16
- mtlr r0
- blr
-.L11:
- bl __kernel_standard@plt
- fmr fp12,fp1
- b .L4
- END (__sqrtf)
-
-weak_alias (__sqrtf, sqrtf)
-
--
1.7.1