glibc: Drop fsqrt implementation patches for fsl ppc
PATCH 1/2 glibc: Drop fsqrt implementation patches for fsl ppc
--- ...5500-e6500-603e-fsqrt-implementation.patch | 1581 ----------------- ...undefined-reference-to-__sqrt_finite.patch | 205 --- ...-are-now-inline-functions-and-call-o.patch | 384 ---- ...-are-now-inline-functions-and-call-o.patch | 58 - meta/recipes-core/glibc/glibc_2.34.bb | 4 - 5 files changed, 2232 deletions(-) delete mode 100644 meta/recipes-core/glibc/glibc/0009-fsl-e500-e5500-e6500-603e-fsqrt-implementation.patch delete mode 100644 meta/recipes-core/glibc/glibc/0010-ppc-sqrt-Fix-undefined-reference-to-__sqrt_finite.patch delete mode 100644 meta/recipes-core/glibc/glibc/0011-__ieee754_sqrt-f-are-now-inline-functions-and-call-o.patch delete mode 100644 meta/recipes-core/glibc/glibc/0014-__ieee754_sqrt-f-are-now-inline-functions-and-call-o.patch
meta/recipes-core/glibc/glibc/0009-fsl-e500-e5500-e6500-603e-fsqrt-implementation.patch deleted +0 -1581
@@ -1,1581 +0,0 @@ -From 74923ca4b1ae0ed5a2478e7d265b37534f6815d7 Mon Sep 17 00:00:00 2001 -From: Khem Raj <[email protected]> -Date: Wed, 18 Mar 2015 00:01:50 +0000 -Subject: [PATCH] fsl e500/e5500/e6500/603e fsqrt implementation - -Upstream-Status: Pending -Signed-off-by: Edmar Wienskoski <[email protected]> -Signed-off-by: Khem Raj <[email protected]> ---- - sysdeps/powerpc/powerpc32/603e/fpu/e_sqrt.c | 134 ++++++++++++++++++ - sysdeps/powerpc/powerpc32/603e/fpu/e_sqrtf.c | 101 +++++++++++++ - sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrt.c | 134 ++++++++++++++++++ - .../powerpc/powerpc32/e500mc/fpu/e_sqrtf.c | 101 +++++++++++++ - sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrt.c | 134 ++++++++++++++++++ - sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrtf.c | 101 +++++++++++++ - sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrt.c | 134 ++++++++++++++++++ - sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrtf.c | 101 +++++++++++++ - sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrt.c | 134 ++++++++++++++++++ - sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrtf.c | 101 +++++++++++++ - sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrt.c | 134 ++++++++++++++++++ - sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrtf.c | 101 +++++++++++++ - .../linux/powerpc/powerpc32/603e/fpu/Implies | 1 + - .../powerpc/powerpc32/e300c3/fpu/Implies | 2 + - .../powerpc/powerpc32/e500mc/fpu/Implies | 1 + - .../linux/powerpc/powerpc32/e5500/fpu/Implies | 1 + - .../linux/powerpc/powerpc32/e6500/fpu/Implies | 1 + - .../linux/powerpc/powerpc64/e5500/fpu/Implies | 1 + - .../linux/powerpc/powerpc64/e6500/fpu/Implies | 1 + - 19 files changed, 1418 insertions(+) - create mode 100644 sysdeps/powerpc/powerpc32/603e/fpu/e_sqrt.c - create mode 100644 sysdeps/powerpc/powerpc32/603e/fpu/e_sqrtf.c - create mode 100644 sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrt.c - create mode 100644 sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrtf.c - create mode 100644 sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrt.c - create mode 100644 sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrtf.c - create mode 100644 sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrt.c - create mode 100644 sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrtf.c - create mode 100644 sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrt.c - create mode 100644 sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrtf.c - create mode 100644 sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrt.c - create mode 100644 sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrtf.c - create mode 100644 sysdeps/unix/sysv/linux/powerpc/powerpc32/603e/fpu/Implies - create mode 100644 sysdeps/unix/sysv/linux/powerpc/powerpc32/e300c3/fpu/Implies - create mode 100644 sysdeps/unix/sysv/linux/powerpc/powerpc32/e500mc/fpu/Implies - create mode 100644 sysdeps/unix/sysv/linux/powerpc/powerpc32/e5500/fpu/Implies - create mode 100644 sysdeps/unix/sysv/linux/powerpc/powerpc32/e6500/fpu/Implies - create mode 100644 sysdeps/unix/sysv/linux/powerpc/powerpc64/e5500/fpu/Implies - create mode 100644 sysdeps/unix/sysv/linux/powerpc/powerpc64/e6500/fpu/Implies - -diff --git a/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrt.c b/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrt.c -new file mode 100644 -index 0000000000..71e516d1c8 ---- /dev/null -+++ b/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrt.c -@@ -0,0 +1,134 @@ -+/* Double-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float two108 = 3.245185536584267269e+32; -+static const float twom54 = 5.551115123125782702e-17; -+static const float half = 0.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the actual square root and half of its reciprocal -+ simultaneously. */ -+ -+#ifdef __STDC__ -+double -+__ieee754_sqrt (double b) -+#else -+double -+__ieee754_sqrt (b) -+ double b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+ double y, g, h, d, r; -+ ieee_double_shape_type u; -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ u.value = b; -+ -+ relax_fenv_state (); -+ -+ __asm__ ("frsqrte %[estimate], %[x]\n" -+ : [estimate] "=f" (y) : [x] "f" (b)); -+ -+ /* Following Muller et al, page 168, equation 5.20. -+ -+ h goes to 1/(2*sqrt(b)) -+ g goes to sqrt(b). -+ -+ We need three iterations to get within 1ulp. */ -+ -+ /* Indicate that these can be performed prior to the branch. GCC -+ insists on sinking them below the branch, however; it seems like -+ they'd be better before the branch so that we can cover any latency -+ from storing the argument and loading its high word. Oh well. */ -+ -+ g = b * y; -+ h = 0.5 * y; -+ -+ /* Handle small numbers by scaling. */ -+ if (__builtin_expect ((u.parts.msw & 0x7ff00000) <= 0x02000000, 0)) -+ return __ieee754_sqrt (b * two108) * twom54; -+ -+#define FMADD(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmadd %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ /* g is now +/- 1ulp, or exactly equal to, the square root of b. */ -+ -+ /* Final refinement. */ -+ d = FNMSUB (g, g, b); -+ -+ fesetenv_register (fe); -+ return FMADD (d, h, g); -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_wash (b); -+} -diff --git a/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrtf.c b/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrtf.c -new file mode 100644 -index 0000000000..26fa067abf ---- /dev/null -+++ b/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrtf.c -@@ -0,0 +1,101 @@ -+/* Single-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float threehalf = 1.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the reciprocal square root and use that to compute the actual -+ square root. */ -+ -+#ifdef __STDC__ -+float -+__ieee754_sqrtf (float b) -+#else -+float -+__ieee754_sqrtf (b) -+ float b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+#define FMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ double y, x; -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ relax_fenv_state (); -+ -+ /* Compute y = 1.5 * b - b. Uses fewer constants than y = 0.5 * b. */ -+ y = FMSUB (threehalf, b, b); -+ -+ /* Initial estimate. */ -+ __asm__ ("frsqrte %[x], %[b]\n" : [x] "=f" (x) : [b] "f" (b)); -+ -+ /* Iterate. x_{n+1} = x_n * (1.5 - y * (x_n * x_n)). */ -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ -+ /* All done. */ -+ fesetenv_register (fe); -+ return x * b; -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_washf (b); -+} -diff --git a/sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrt.c b/sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrt.c -new file mode 100644 -index 0000000000..71e516d1c8 ---- /dev/null -+++ b/sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrt.c -@@ -0,0 +1,134 @@ -+/* Double-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float two108 = 3.245185536584267269e+32; -+static const float twom54 = 5.551115123125782702e-17; -+static const float half = 0.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the actual square root and half of its reciprocal -+ simultaneously. */ -+ -+#ifdef __STDC__ -+double -+__ieee754_sqrt (double b) -+#else -+double -+__ieee754_sqrt (b) -+ double b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+ double y, g, h, d, r; -+ ieee_double_shape_type u; -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ u.value = b; -+ -+ relax_fenv_state (); -+ -+ __asm__ ("frsqrte %[estimate], %[x]\n" -+ : [estimate] "=f" (y) : [x] "f" (b)); -+ -+ /* Following Muller et al, page 168, equation 5.20. -+ -+ h goes to 1/(2*sqrt(b)) -+ g goes to sqrt(b). -+ -+ We need three iterations to get within 1ulp. */ -+ -+ /* Indicate that these can be performed prior to the branch. GCC -+ insists on sinking them below the branch, however; it seems like -+ they'd be better before the branch so that we can cover any latency -+ from storing the argument and loading its high word. Oh well. */ -+ -+ g = b * y; -+ h = 0.5 * y; -+ -+ /* Handle small numbers by scaling. */ -+ if (__builtin_expect ((u.parts.msw & 0x7ff00000) <= 0x02000000, 0)) -+ return __ieee754_sqrt (b * two108) * twom54; -+ -+#define FMADD(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmadd %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ /* g is now +/- 1ulp, or exactly equal to, the square root of b. */ -+ -+ /* Final refinement. */ -+ d = FNMSUB (g, g, b); -+ -+ fesetenv_register (fe); -+ return FMADD (d, h, g); -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_wash (b); -+} -diff --git a/sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrtf.c b/sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrtf.c -new file mode 100644 -index 0000000000..26fa067abf ---- /dev/null -+++ b/sysdeps/powerpc/powerpc32/e500mc/fpu/e_sqrtf.c -@@ -0,0 +1,101 @@ -+/* Single-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float threehalf = 1.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the reciprocal square root and use that to compute the actual -+ square root. */ -+ -+#ifdef __STDC__ -+float -+__ieee754_sqrtf (float b) -+#else -+float -+__ieee754_sqrtf (b) -+ float b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+#define FMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ double y, x; -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ relax_fenv_state (); -+ -+ /* Compute y = 1.5 * b - b. Uses fewer constants than y = 0.5 * b. */ -+ y = FMSUB (threehalf, b, b); -+ -+ /* Initial estimate. */ -+ __asm__ ("frsqrte %[x], %[b]\n" : [x] "=f" (x) : [b] "f" (b)); -+ -+ /* Iterate. x_{n+1} = x_n * (1.5 - y * (x_n * x_n)). */ -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ -+ /* All done. */ -+ fesetenv_register (fe); -+ return x * b; -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_washf (b); -+} -diff --git a/sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrt.c b/sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrt.c -new file mode 100644 -index 0000000000..71e516d1c8 ---- /dev/null -+++ b/sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrt.c -@@ -0,0 +1,134 @@ -+/* Double-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float two108 = 3.245185536584267269e+32; -+static const float twom54 = 5.551115123125782702e-17; -+static const float half = 0.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the actual square root and half of its reciprocal -+ simultaneously. */ -+ -+#ifdef __STDC__ -+double -+__ieee754_sqrt (double b) -+#else -+double -+__ieee754_sqrt (b) -+ double b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+ double y, g, h, d, r; -+ ieee_double_shape_type u; -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ u.value = b; -+ -+ relax_fenv_state (); -+ -+ __asm__ ("frsqrte %[estimate], %[x]\n" -+ : [estimate] "=f" (y) : [x] "f" (b)); -+ -+ /* Following Muller et al, page 168, equation 5.20. -+ -+ h goes to 1/(2*sqrt(b)) -+ g goes to sqrt(b). -+ -+ We need three iterations to get within 1ulp. */ -+ -+ /* Indicate that these can be performed prior to the branch. GCC -+ insists on sinking them below the branch, however; it seems like -+ they'd be better before the branch so that we can cover any latency -+ from storing the argument and loading its high word. Oh well. */ -+ -+ g = b * y; -+ h = 0.5 * y; -+ -+ /* Handle small numbers by scaling. */ -+ if (__builtin_expect ((u.parts.msw & 0x7ff00000) <= 0x02000000, 0)) -+ return __ieee754_sqrt (b * two108) * twom54; -+ -+#define FMADD(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmadd %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ /* g is now +/- 1ulp, or exactly equal to, the square root of b. */ -+ -+ /* Final refinement. */ -+ d = FNMSUB (g, g, b); -+ -+ fesetenv_register (fe); -+ return FMADD (d, h, g); -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_wash (b); -+} -diff --git a/sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrtf.c b/sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrtf.c -new file mode 100644 -index 0000000000..26fa067abf ---- /dev/null -+++ b/sysdeps/powerpc/powerpc32/e5500/fpu/e_sqrtf.c -@@ -0,0 +1,101 @@ -+/* Single-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float threehalf = 1.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the reciprocal square root and use that to compute the actual -+ square root. */ -+ -+#ifdef __STDC__ -+float -+__ieee754_sqrtf (float b) -+#else -+float -+__ieee754_sqrtf (b) -+ float b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+#define FMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ double y, x; -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ relax_fenv_state (); -+ -+ /* Compute y = 1.5 * b - b. Uses fewer constants than y = 0.5 * b. */ -+ y = FMSUB (threehalf, b, b); -+ -+ /* Initial estimate. */ -+ __asm__ ("frsqrte %[x], %[b]\n" : [x] "=f" (x) : [b] "f" (b)); -+ -+ /* Iterate. x_{n+1} = x_n * (1.5 - y * (x_n * x_n)). */ -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ -+ /* All done. */ -+ fesetenv_register (fe); -+ return x * b; -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_washf (b); -+} -diff --git a/sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrt.c b/sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrt.c -new file mode 100644 -index 0000000000..71e516d1c8 ---- /dev/null -+++ b/sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrt.c -@@ -0,0 +1,134 @@ -+/* Double-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float two108 = 3.245185536584267269e+32; -+static const float twom54 = 5.551115123125782702e-17; -+static const float half = 0.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the actual square root and half of its reciprocal -+ simultaneously. */ -+ -+#ifdef __STDC__ -+double -+__ieee754_sqrt (double b) -+#else -+double -+__ieee754_sqrt (b) -+ double b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+ double y, g, h, d, r; -+ ieee_double_shape_type u; -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ u.value = b; -+ -+ relax_fenv_state (); -+ -+ __asm__ ("frsqrte %[estimate], %[x]\n" -+ : [estimate] "=f" (y) : [x] "f" (b)); -+ -+ /* Following Muller et al, page 168, equation 5.20. -+ -+ h goes to 1/(2*sqrt(b)) -+ g goes to sqrt(b). -+ -+ We need three iterations to get within 1ulp. */ -+ -+ /* Indicate that these can be performed prior to the branch. GCC -+ insists on sinking them below the branch, however; it seems like -+ they'd be better before the branch so that we can cover any latency -+ from storing the argument and loading its high word. Oh well. */ -+ -+ g = b * y; -+ h = 0.5 * y; -+ -+ /* Handle small numbers by scaling. */ -+ if (__builtin_expect ((u.parts.msw & 0x7ff00000) <= 0x02000000, 0)) -+ return __ieee754_sqrt (b * two108) * twom54; -+ -+#define FMADD(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmadd %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ /* g is now +/- 1ulp, or exactly equal to, the square root of b. */ -+ -+ /* Final refinement. */ -+ d = FNMSUB (g, g, b); -+ -+ fesetenv_register (fe); -+ return FMADD (d, h, g); -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_wash (b); -+} -diff --git a/sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrtf.c b/sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrtf.c -new file mode 100644 -index 0000000000..26fa067abf ---- /dev/null -+++ b/sysdeps/powerpc/powerpc32/e6500/fpu/e_sqrtf.c -@@ -0,0 +1,101 @@ -+/* Single-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float threehalf = 1.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the reciprocal square root and use that to compute the actual -+ square root. */ -+ -+#ifdef __STDC__ -+float -+__ieee754_sqrtf (float b) -+#else -+float -+__ieee754_sqrtf (b) -+ float b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+#define FMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ double y, x; -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ relax_fenv_state (); -+ -+ /* Compute y = 1.5 * b - b. Uses fewer constants than y = 0.5 * b. */ -+ y = FMSUB (threehalf, b, b); -+ -+ /* Initial estimate. */ -+ __asm__ ("frsqrte %[x], %[b]\n" : [x] "=f" (x) : [b] "f" (b)); -+ -+ /* Iterate. x_{n+1} = x_n * (1.5 - y * (x_n * x_n)). */ -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ -+ /* All done. */ -+ fesetenv_register (fe); -+ return x * b; -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_washf (b); -+} -diff --git a/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrt.c b/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrt.c -new file mode 100644 -index 0000000000..71e516d1c8 ---- /dev/null -+++ b/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrt.c -@@ -0,0 +1,134 @@ -+/* Double-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float two108 = 3.245185536584267269e+32; -+static const float twom54 = 5.551115123125782702e-17; -+static const float half = 0.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the actual square root and half of its reciprocal -+ simultaneously. */ -+ -+#ifdef __STDC__ -+double -+__ieee754_sqrt (double b) -+#else -+double -+__ieee754_sqrt (b) -+ double b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+ double y, g, h, d, r; -+ ieee_double_shape_type u; -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ u.value = b; -+ -+ relax_fenv_state (); -+ -+ __asm__ ("frsqrte %[estimate], %[x]\n" -+ : [estimate] "=f" (y) : [x] "f" (b)); -+ -+ /* Following Muller et al, page 168, equation 5.20. -+ -+ h goes to 1/(2*sqrt(b)) -+ g goes to sqrt(b). -+ -+ We need three iterations to get within 1ulp. */ -+ -+ /* Indicate that these can be performed prior to the branch. GCC -+ insists on sinking them below the branch, however; it seems like -+ they'd be better before the branch so that we can cover any latency -+ from storing the argument and loading its high word. Oh well. */ -+ -+ g = b * y; -+ h = 0.5 * y; -+ -+ /* Handle small numbers by scaling. */ -+ if (__builtin_expect ((u.parts.msw & 0x7ff00000) <= 0x02000000, 0)) -+ return __ieee754_sqrt (b * two108) * twom54; -+ -+#define FMADD(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmadd %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ /* g is now +/- 1ulp, or exactly equal to, the square root of b. */ -+ -+ /* Final refinement. */ -+ d = FNMSUB (g, g, b); -+ -+ fesetenv_register (fe); -+ return FMADD (d, h, g); -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_wash (b); -+} -diff --git a/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrtf.c b/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrtf.c -new file mode 100644 -index 0000000000..26fa067abf ---- /dev/null -+++ b/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrtf.c -@@ -0,0 +1,101 @@ -+/* Single-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float threehalf = 1.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the reciprocal square root and use that to compute the actual -+ square root. */ -+ -+#ifdef __STDC__ -+float -+__ieee754_sqrtf (float b) -+#else -+float -+__ieee754_sqrtf (b) -+ float b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+#define FMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ double y, x; -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ relax_fenv_state (); -+ -+ /* Compute y = 1.5 * b - b. Uses fewer constants than y = 0.5 * b. */ -+ y = FMSUB (threehalf, b, b); -+ -+ /* Initial estimate. */ -+ __asm__ ("frsqrte %[x], %[b]\n" : [x] "=f" (x) : [b] "f" (b)); -+ -+ /* Iterate. x_{n+1} = x_n * (1.5 - y * (x_n * x_n)). */ -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ -+ /* All done. */ -+ fesetenv_register (fe); -+ return x * b; -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_washf (b); -+} -diff --git a/sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrt.c b/sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrt.c -new file mode 100644 -index 0000000000..71e516d1c8 ---- /dev/null -+++ b/sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrt.c -@@ -0,0 +1,134 @@ -+/* Double-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float two108 = 3.245185536584267269e+32; -+static const float twom54 = 5.551115123125782702e-17; -+static const float half = 0.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the actual square root and half of its reciprocal -+ simultaneously. */ -+ -+#ifdef __STDC__ -+double -+__ieee754_sqrt (double b) -+#else -+double -+__ieee754_sqrt (b) -+ double b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+ double y, g, h, d, r; -+ ieee_double_shape_type u; -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ u.value = b; -+ -+ relax_fenv_state (); -+ -+ __asm__ ("frsqrte %[estimate], %[x]\n" -+ : [estimate] "=f" (y) : [x] "f" (b)); -+ -+ /* Following Muller et al, page 168, equation 5.20. -+ -+ h goes to 1/(2*sqrt(b)) -+ g goes to sqrt(b). -+ -+ We need three iterations to get within 1ulp. */ -+ -+ /* Indicate that these can be performed prior to the branch. GCC -+ insists on sinking them below the branch, however; it seems like -+ they'd be better before the branch so that we can cover any latency -+ from storing the argument and loading its high word. Oh well. */ -+ -+ g = b * y; -+ h = 0.5 * y; -+ -+ /* Handle small numbers by scaling. */ -+ if (__builtin_expect ((u.parts.msw & 0x7ff00000) <= 0x02000000, 0)) -+ return __ieee754_sqrt (b * two108) * twom54; -+ -+#define FMADD(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmadd %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ r = FNMSUB (g, h, half); -+ g = FMADD (g, r, g); -+ h = FMADD (h, r, h); -+ -+ /* g is now +/- 1ulp, or exactly equal to, the square root of b. */ -+ -+ /* Final refinement. */ -+ d = FNMSUB (g, g, b); -+ -+ fesetenv_register (fe); -+ return FMADD (d, h, g); -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_wash (b); -+} -diff --git a/sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrtf.c b/sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrtf.c -new file mode 100644 -index 0000000000..26fa067abf ---- /dev/null -+++ b/sysdeps/powerpc/powerpc64/e6500/fpu/e_sqrtf.c -@@ -0,0 +1,101 @@ -+/* Single-precision floating point square root. -+ Copyright (C) 2010 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, write to the Free -+ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA -+ 02111-1307 USA. */ -+ -+#include <math.h> -+#include <math_private.h> -+#include <fenv_libc.h> -+#include <inttypes.h> -+ -+#include <sysdep.h> -+#include <ldsodefs.h> -+ -+static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 }; -+static const ieee_float_shape_type a_inf = {.word = 0x7f800000 }; -+static const float threehalf = 1.5; -+ -+/* The method is based on the descriptions in: -+ -+ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5; -+ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9 -+ -+ We find the reciprocal square root and use that to compute the actual -+ square root. */ -+ -+#ifdef __STDC__ -+float -+__ieee754_sqrtf (float b) -+#else -+float -+__ieee754_sqrtf (b) -+ float b; -+#endif -+{ -+ if (__builtin_expect (b > 0, 1)) -+ { -+#define FMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+#define FNMSUB(a_, c_, b_) \ -+ ({ double __r; \ -+ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \ -+ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \ -+ __r;}) -+ -+ if (__builtin_expect (b != a_inf.value, 1)) -+ { -+ double y, x; -+ fenv_t fe; -+ -+ fe = fegetenv_register (); -+ -+ relax_fenv_state (); -+ -+ /* Compute y = 1.5 * b - b. Uses fewer constants than y = 0.5 * b. */ -+ y = FMSUB (threehalf, b, b); -+ -+ /* Initial estimate. */ -+ __asm__ ("frsqrte %[x], %[b]\n" : [x] "=f" (x) : [b] "f" (b)); -+ -+ /* Iterate. x_{n+1} = x_n * (1.5 - y * (x_n * x_n)). */ -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ x = x * FNMSUB (y, x * x, threehalf); -+ -+ /* All done. */ -+ fesetenv_register (fe); -+ return x * b; -+ } -+ } -+ else if (b < 0) -+ { -+ /* For some reason, some PowerPC32 processors don't implement -+ FE_INVALID_SQRT. */ -+#ifdef FE_INVALID_SQRT -+ feraiseexcept (FE_INVALID_SQRT); -+ -+ fenv_union_t u = { .fenv = fegetenv_register () }; -+ if ((u.l & FE_INVALID) == 0) -+#endif -+ feraiseexcept (FE_INVALID); -+ b = a_nan.value; -+ } -+ return f_washf (b); -+} -diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc32/603e/fpu/Implies b/sysdeps/unix/sysv/linux/powerpc/powerpc32/603e/fpu/Implies -new file mode 100644 -index 0000000000..b103b4dea5 ---- /dev/null -+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/603e/fpu/Implies -@@ -0,0 +1 @@ -+powerpc/powerpc32/603e/fpu -diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc32/e300c3/fpu/Implies b/sysdeps/unix/sysv/linux/powerpc/powerpc32/e300c3/fpu/Implies -new file mode 100644 -index 0000000000..64db17fada ---- /dev/null -+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/e300c3/fpu/Implies -@@ -0,0 +1,2 @@ -+# e300c3 is a variant of 603e so use the same optimizations for sqrt -+powerpc/powerpc32/603e/fpu -diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc32/e500mc/fpu/Implies b/sysdeps/unix/sysv/linux/powerpc/powerpc32/e500mc/fpu/Implies -new file mode 100644 -index 0000000000..7eac5fcf02 ---- /dev/null -+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/e500mc/fpu/Implies -@@ -0,0 +1 @@ -+powerpc/powerpc32/e500mc/fpu -diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc32/e5500/fpu/Implies b/sysdeps/unix/sysv/linux/powerpc/powerpc32/e5500/fpu/Implies -new file mode 100644 -index 0000000000..264b2a7700 ---- /dev/null -+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/e5500/fpu/Implies -@@ -0,0 +1 @@ -+powerpc/powerpc32/e5500/fpu -diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc32/e6500/fpu/Implies b/sysdeps/unix/sysv/linux/powerpc/powerpc32/e6500/fpu/Implies -new file mode 100644 -index 0000000000..a25934467b ---- /dev/null -+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc32/e6500/fpu/Implies -@@ -0,0 +1 @@ -+powerpc/powerpc32/e6500/fpu -diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc64/e5500/fpu/Implies b/sysdeps/unix/sysv/linux/powerpc/powerpc64/e5500/fpu/Implies -new file mode 100644 -index 0000000000..a7bc854be8 ---- /dev/null -+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc64/e5500/fpu/Implies -@@ -0,0 +1 @@ -+powerpc/powerpc64/e5500/fpu -diff --git a/sysdeps/unix/sysv/linux/powerpc/powerpc64/e6500/fpu/Implies b/sysdeps/unix/sysv/linux/powerpc/powerpc64/e6500/fpu/Implies -new file mode 100644 -index 0000000000..04ff8cc181 ---- /dev/null -+++ b/sysdeps/unix/sysv/linux/powerpc/powerpc64/e6500/fpu/Implies -@@ -0,0 +1 @@ -+powerpc/powerpc64/e6500/fpu
meta/recipes-core/glibc/glibc/0010-ppc-sqrt-Fix-undefined-reference-to-__sqrt_finite.patch deleted +0 -205
205 lines not on this page · raw message
meta/recipes-core/glibc/glibc/0011-__ieee754_sqrt-f-are-now-inline-functions-and-call-o.patch deleted +0 -384
384 lines not on this page · raw message
meta/recipes-core/glibc/glibc/0014-__ieee754_sqrt-f-are-now-inline-functions-and-call-o.patch deleted +0 -58
58 lines not on this page · raw message
meta/recipes-core/glibc/glibc_2.34.bb +0 -4
@@ -30,11 +30,7 @@ NATIVESDKFIXES:class-nativesdk = "\ SRC_URI = "${GLIBC_GIT_URI};branch=${SRCBRANCH};name=glibc \ ${NATIVESDKFIXES} \ - file://0009-fsl-e500-e5500-e6500-603e-fsqrt-implementation.patch \ - file://0010-ppc-sqrt-Fix-undefined-reference-to-__sqrt_finite.patch \ - file://0011-__ieee754_sqrt-f-are-now-inline-functions-and-call-o.patch \ file://0013-eglibc-run-libm-err-tab.pl-with-specific-dirs-in-S.patch \ - file://0014-__ieee754_sqrt-f-are-now-inline-functions-and-call-o.patch \ file://0015-sysdeps-gnu-configure.ac-handle-correctly-libc_cv_ro.patch \ file://0016-yes-within-the-path-sets-wrong-config-variables.patch \ file://0017-timezone-re-written-tzselect-as-posix-sh.patch \
Signature · 1 line
2.34.1
PATCH 2/2 ffmpeg: Fix build on ppc32
Fixes
| powerpc-yoe-linux/11.2.0/ld: libavformat/libavformat.so: undefined reference to `__atomic_fetch_sub_8'
--- meta/recipes-multimedia/ffmpeg/ffmpeg_4.4.1.bb | 1 + 1 file changed, 1 insertion(+)
meta/recipes-multimedia/ffmpeg/ffmpeg_4.4.1.bb +1 -0
@@ -131,6 +131,7 @@ EXTRA_OECONF:append:mips = " --extra-libs=-latomic --disable-mips32r5 --disable- --disable-loongson2 --disable-loongson3 --disable-mmi --disable-msa --disable-msa2" EXTRA_OECONF:append:riscv32 = " --extra-libs=-latomic" EXTRA_OECONF:append:armv5 = " --extra-libs=-latomic" +EXTRA_OECONF:append:powerpc = " --extra-libs=-latomic" # gold crashes on x86, another solution is to --disable-asm but thats more hacky # ld.gold: internal error in relocate_section, at ../../gold/i386.cc:3684
Signature · 1 line
2.34.1
I did test such a patch today as well since we discussed it. I'd propose this commit message since we need to say why:
"""
glibc: Drop ppc sqrt optimisations
OpenEmbedded isn't an upstream or a patch repository. These are optimisations which for reasons unknown were never merged into upstream glibc. We have no active ppc maintainers and these patches are at an evolutionary dead end. Since they are optimisations, there may be small performance regressions by removing them but OE can't carry patches like this indefinitely and these appear to have no future.
"""
Cheers,
Richard
Quoted 8 lines · Richard Purdie
I did test such a patch today as well since we discussed it. I'd propose this
commit message since we need to say why:"""
glibc: Drop ppc sqrt optimisations
Quoted 8 lines
OpenEmbedded isn't an upstream or a patch repository. These are optimisations
which for reasons unknown were never merged into upstream glibc. We have no active ppc maintainers and these patches are at an evolutionary dead end. Since they are optimisations, there may be small performance regressions by removing them but OE can't carry patches like this indefinitely and these appear to have no future.
"""
This sounds good to me. I guess will add it to commit or do you want me to send v2. This passed my local testing on qemuppc
Quoted 7 lines
Cheers,
Richard
Quoted 23 lines · Khem Raj
Quoted 8 lines · Richard Purdie
I did test such a patch today as well since we discussed it. I'd propose this
commit message since we need to say why:"""
glibc: Drop ppc sqrt optimisations
Quoted 9 lines
OpenEmbedded isn't an upstream or a patch repository. These are optimisations
which for reasons unknown were never merged into upstream glibc. We have no active ppc maintainers and these patches are at an evolutionary dead end. Since they are optimisations, there may be small performance regressions by removing them but OE can't carry patches like this indefinitely and these appear to have no future.
"""
Quoted 3 lines
This sounds good to me. I guess will add it to commit or do you want me to send v2. This passed my local testing on qemuppc
I can sort if you're ok with it, thanks.
Cheers,
Richard
Quoted 29 lines · Richard Purdie
Quoted 5 lines · Khem Raj
Quoted 1 line · Richard Purdie
I did test such a patch today as well since we discussed it. I'd
propose
Quoted 10 lines
Quoted 10 lines
this
commit message since we need to say why:"""
glibc: Drop ppc sqrt optimisations
OpenEmbedded isn't an upstream or a patch repository. These are optimisations
which for reasons unknown were never merged into upstream glibc. Wehave no
Quoted 1 line
Quoted 1 line
active ppc maintainers and these patches are at an evolutionary dead
end.
Quoted 1 line
Quoted 1 line
Since they are optimisations, there may be small performance
regressions by
Quoted 1 line
Quoted 1 line
removing them but OE can't carry patches like this indefinitely and
these
Quoted 5 lines
Quoted 3 lines
appear to have no future.
"""
Quoted 6 lines
Quoted 1 line
This sounds good to me. I guess will add it to commit or do you want me
to
Quoted 2 lines
send v2. This passed my local testing on qemuppc
I can sort if you're ok with it, thanks.
Sounds good
Quoted 6 lines
Cheers,
Richard