--- /dev/null
+Upstream-Status: Pending\r
+\r
+2011-03-22 Joseph Myers <joseph@codesourcery.com>\r
+\r
+ Merge from SG++ 2.11:\r
+\r
+ 2010-10-05 Nathan Froyd <froydnj@codesourcery.com>\r
+\r
+ Issue #9382\r
+\r
+ * sysdeps/powerpc/powerpc32/603e/: New directory.\r
+ * sysdeps/unix/sysv/linux/powerpc/powerpc32/e500mc/: New directory.\r
+ * sysdeps/unix/sysv/linux/powerpc/powerpc32/603e/: New directory.\r
+ * sysdeps/unix/sysv/linux/powerpc/powerpc32/7400/: New directory.\r
+ * sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrtf.c: Update.\r
+ * sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrt.c: Update.\r
+ * sysdeps/powerpc/powerpc64/e5500/fpu/Implies: New file.\r
+\r
+Index: libc/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrt.c\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrt.c\r
+@@ -0,0 +1,134 @@\r
++/* Double-precision floating point square root.\r
++ Copyright (C) 2010 Free Software Foundation, Inc.\r
++ This file is part of the GNU C Library.\r
++\r
++ The GNU C Library is free software; you can redistribute it and/or\r
++ modify it under the terms of the GNU Lesser General Public\r
++ License as published by the Free Software Foundation; either\r
++ version 2.1 of the License, or (at your option) any later version.\r
++\r
++ The GNU C Library is distributed in the hope that it will be useful,\r
++ but WITHOUT ANY WARRANTY; without even the implied warranty of\r
++ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU\r
++ Lesser General Public License for more details.\r
++\r
++ You should have received a copy of the GNU Lesser General Public\r
++ License along with the GNU C Library; if not, write to the Free\r
++ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA\r
++ 02111-1307 USA. */\r
++\r
++#include <math.h>\r
++#include <math_private.h>\r
++#include <fenv_libc.h>\r
++#include <inttypes.h>\r
++\r
++#include <sysdep.h>\r
++#include <ldsodefs.h>\r
++\r
++static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 };\r
++static const ieee_float_shape_type a_inf = {.word = 0x7f800000 };\r
++static const float two108 = 3.245185536584267269e+32;\r
++static const float twom54 = 5.551115123125782702e-17;\r
++static const float half = 0.5;\r
++\r
++/* The method is based on the descriptions in:\r
++\r
++ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5;\r
++ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9\r
++\r
++ We find the actual square root and half of its reciprocal\r
++ simultaneously. */\r
++\r
++#ifdef __STDC__\r
++double\r
++__ieee754_sqrt (double b)\r
++#else\r
++double\r
++__ieee754_sqrt (b)\r
++ double b;\r
++#endif\r
++{\r
++ if (__builtin_expect (b > 0, 1))\r
++ {\r
++ double y, g, h, d, r;\r
++ ieee_double_shape_type u;\r
++\r
++ if (__builtin_expect (b != a_inf.value, 1))\r
++ {\r
++ fenv_t fe;\r
++\r
++ fe = fegetenv_register ();\r
++\r
++ u.value = b;\r
++\r
++ relax_fenv_state ();\r
++\r
++ __asm__ ("frsqrte %[estimate], %[x]\n"\r
++ : [estimate] "=f" (y) : [x] "f" (b));\r
++\r
++ /* Following Muller et al, page 168, equation 5.20.\r
++\r
++ h goes to 1/(2*sqrt(b))\r
++ g goes to sqrt(b).\r
++\r
++ We need three iterations to get within 1ulp. */\r
++\r
++ /* Indicate that these can be performed prior to the branch. GCC\r
++ insists on sinking them below the branch, however; it seems like\r
++ they'd be better before the branch so that we can cover any latency\r
++ from storing the argument and loading its high word. Oh well. */\r
++\r
++ g = b * y;\r
++ h = 0.5 * y;\r
++ \r
++ /* Handle small numbers by scaling. */\r
++ if (__builtin_expect ((u.parts.msw & 0x7ff00000) <= 0x02000000, 0))\r
++ return __ieee754_sqrt (b * two108) * twom54;\r
++\r
++#define FMADD(a_, c_, b_) \\r
++ ({ double __r; \\r
++ __asm__ ("fmadd %[r], %[a], %[c], %[b]\n" \\r
++ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \\r
++ __r;})\r
++#define FNMSUB(a_, c_, b_) \\r
++ ({ double __r; \\r
++ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \\r
++ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \\r
++ __r;})\r
++\r
++ r = FNMSUB (g, h, half);\r
++ g = FMADD (g, r, g);\r
++ h = FMADD (h, r, h);\r
++\r
++ r = FNMSUB (g, h, half);\r
++ g = FMADD (g, r, g);\r
++ h = FMADD (h, r, h);\r
++\r
++ r = FNMSUB (g, h, half);\r
++ g = FMADD (g, r, g);\r
++ h = FMADD (h, r, h);\r
++\r
++ /* g is now +/- 1ulp, or exactly equal to, the square root of b. */\r
++\r
++ /* Final refinement. */\r
++ d = FNMSUB (g, g, b);\r
++\r
++ fesetenv_register (fe);\r
++ return FMADD (d, h, g);\r
++ }\r
++ }\r
++ else if (b < 0)\r
++ {\r
++ /* For some reason, some PowerPC32 processors don't implement\r
++ FE_INVALID_SQRT. */\r
++#ifdef FE_INVALID_SQRT\r
++ feraiseexcept (FE_INVALID_SQRT);\r
++\r
++ fenv_union_t u = { .fenv = fegetenv_register () };\r
++ if ((u.l[1] & FE_INVALID) == 0)\r
++#endif\r
++ feraiseexcept (FE_INVALID);\r
++ b = a_nan.value;\r
++ }\r
++ return f_wash (b);\r
++}\r
+Index: libc/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrtf.c\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/powerpc/powerpc32/603e/fpu/e_sqrtf.c\r
+@@ -0,0 +1,101 @@\r
++/* Single-precision floating point square root.\r
++ Copyright (C) 2010 Free Software Foundation, Inc.\r
++ This file is part of the GNU C Library.\r
++\r
++ The GNU C Library is free software; you can redistribute it and/or\r
++ modify it under the terms of the GNU Lesser General Public\r
++ License as published by the Free Software Foundation; either\r
++ version 2.1 of the License, or (at your option) any later version.\r
++\r
++ The GNU C Library is distributed in the hope that it will be useful,\r
++ but WITHOUT ANY WARRANTY; without even the implied warranty of\r
++ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU\r
++ Lesser General Public License for more details.\r
++\r
++ You should have received a copy of the GNU Lesser General Public\r
++ License along with the GNU C Library; if not, write to the Free\r
++ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA\r
++ 02111-1307 USA. */\r
++\r
++#include <math.h>\r
++#include <math_private.h>\r
++#include <fenv_libc.h>\r
++#include <inttypes.h>\r
++\r
++#include <sysdep.h>\r
++#include <ldsodefs.h>\r
++\r
++static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 };\r
++static const ieee_float_shape_type a_inf = {.word = 0x7f800000 };\r
++static const float threehalf = 1.5;\r
++\r
++/* The method is based on the descriptions in:\r
++\r
++ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5;\r
++ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9\r
++\r
++ We find the reciprocal square root and use that to compute the actual\r
++ square root. */\r
++\r
++#ifdef __STDC__\r
++float\r
++__ieee754_sqrtf (float b)\r
++#else\r
++float\r
++__ieee754_sqrtf (b)\r
++ float b;\r
++#endif\r
++{\r
++ if (__builtin_expect (b > 0, 1))\r
++ {\r
++#define FMSUB(a_, c_, b_) \\r
++ ({ double __r; \\r
++ __asm__ ("fmsub %[r], %[a], %[c], %[b]\n" \\r
++ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \\r
++ __r;})\r
++#define FNMSUB(a_, c_, b_) \\r
++ ({ double __r; \\r
++ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \\r
++ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \\r
++ __r;})\r
++\r
++ if (__builtin_expect (b != a_inf.value, 1))\r
++ {\r
++ double y, x;\r
++ fenv_t fe;\r
++\r
++ fe = fegetenv_register ();\r
++\r
++ relax_fenv_state ();\r
++\r
++ /* Compute y = 1.5 * b - b. Uses fewer constants than y = 0.5 * b. */\r
++ y = FMSUB (threehalf, b, b);\r
++\r
++ /* Initial estimate. */\r
++ __asm__ ("frsqrte %[x], %[b]\n" : [x] "=f" (x) : [b] "f" (b));\r
++\r
++ /* Iterate. x_{n+1} = x_n * (1.5 - y * (x_n * x_n)). */\r
++ x = x * FNMSUB (y, x * x, threehalf);\r
++ x = x * FNMSUB (y, x * x, threehalf);\r
++ x = x * FNMSUB (y, x * x, threehalf);\r
++\r
++ /* All done. */\r
++ fesetenv_register (fe);\r
++ return x * b;\r
++ }\r
++ }\r
++ else if (b < 0)\r
++ {\r
++ /* For some reason, some PowerPC32 processors don't implement\r
++ FE_INVALID_SQRT. */\r
++#ifdef FE_INVALID_SQRT\r
++ feraiseexcept (FE_INVALID_SQRT);\r
++\r
++ fenv_union_t u = { .fenv = fegetenv_register () };\r
++ if ((u.l[1] & FE_INVALID) == 0)\r
++#endif\r
++ feraiseexcept (FE_INVALID);\r
++ b = a_nan.value;\r
++ }\r
++ return f_washf (b);\r
++}\r
+Index: libc/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrt.c\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrt.c\r
+@@ -0,0 +1,134 @@\r
++/* Double-precision floating point square root.\r
++ Copyright (C) 2010 Free Software Foundation, Inc.\r
++ This file is part of the GNU C Library.\r
++\r
++ The GNU C Library is free software; you can redistribute it and/or\r
++ modify it under the terms of the GNU Lesser General Public\r
++ License as published by the Free Software Foundation; either\r
++ version 2.1 of the License, or (at your option) any later version.\r
++\r
++ The GNU C Library is distributed in the hope that it will be useful,\r
++ but WITHOUT ANY WARRANTY; without even the implied warranty of\r
++ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU\r
++ Lesser General Public License for more details.\r
++\r
++ You should have received a copy of the GNU Lesser General Public\r
++ License along with the GNU C Library; if not, write to the Free\r
++ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA\r
++ 02111-1307 USA. */\r
++\r
++#include <math.h>\r
++#include <math_private.h>\r
++#include <fenv_libc.h>\r
++#include <inttypes.h>\r
++\r
++#include <sysdep.h>\r
++#include <ldsodefs.h>\r
++\r
++static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 };\r
++static const ieee_float_shape_type a_inf = {.word = 0x7f800000 };\r
++static const float two108 = 3.245185536584267269e+32;\r
++static const float twom54 = 5.551115123125782702e-17;\r
++static const float half = 0.5;\r
++\r
++/* The method is based on the descriptions in:\r
++\r
++ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5;\r
++ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9\r
++\r
++ We find the actual square root and half of its reciprocal\r
++ simultaneously. */\r
++\r
++#ifdef __STDC__\r
++double\r
++__ieee754_sqrt (double b)\r
++#else\r
++double\r
++__ieee754_sqrt (b)\r
++ double b;\r
++#endif\r
++{\r
++ if (__builtin_expect (b > 0, 1))\r
++ {\r
++ double y, g, h, d, r;\r
++ ieee_double_shape_type u;\r
++\r
++ if (__builtin_expect (b != a_inf.value, 1))\r
++ {\r
++ fenv_t fe;\r
++\r
++ fe = fegetenv_register ();\r
++\r
++ u.value = b;\r
++\r
++ relax_fenv_state ();\r
++\r
++ __asm__ ("frsqrte %[estimate], %[x]\n"\r
++ : [estimate] "=f" (y) : [x] "f" (b));\r
++\r
++ /* Following Muller et al, page 168, equation 5.20.\r
++\r
++ h goes to 1/(2*sqrt(b))\r
++ g goes to sqrt(b).\r
++\r
++ We need three iterations to get within 1ulp. */\r
++\r
++ /* Indicate that these can be performed prior to the branch. GCC\r
++ insists on sinking them below the branch, however; it seems like\r
++ they'd be better before the branch so that we can cover any latency\r
++ from storing the argument and loading its high word. Oh well. */\r
++\r
++ g = b * y;\r
++ h = 0.5 * y;\r
++ \r
++ /* Handle small numbers by scaling. */\r
++ if (__builtin_expect ((u.parts.msw & 0x7ff00000) <= 0x02000000, 0))\r
++ return __ieee754_sqrt (b * two108) * twom54;\r
++\r
++#define FMADD(a_, c_, b_) \\r
++ ({ double __r; \\r
++ __asm__ ("fmadd %[r], %[a], %[c], %[b]\n" \\r
++ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \\r
++ __r;})\r
++#define FNMSUB(a_, c_, b_) \\r
++ ({ double __r; \\r
++ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \\r
++ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \\r
++ __r;})\r
++\r
++ r = FNMSUB (g, h, half);\r
++ g = FMADD (g, r, g);\r
++ h = FMADD (h, r, h);\r
++\r
++ r = FNMSUB (g, h, half);\r
++ g = FMADD (g, r, g);\r
++ h = FMADD (h, r, h);\r
++\r
++ r = FNMSUB (g, h, half);\r
++ g = FMADD (g, r, g);\r
++ h = FMADD (h, r, h);\r
++\r
++ /* g is now +/- 1ulp, or exactly equal to, the square root of b. */\r
++\r
++ /* Final refinement. */\r
++ d = FNMSUB (g, g, b);\r
++\r
++ fesetenv_register (fe);\r
++ return FMADD (d, h, g);\r
++ }\r
++ }\r
++ else if (b < 0)\r
++ {\r
++ /* For some reason, some PowerPC32 processors don't implement\r
++ FE_INVALID_SQRT. */\r
++#ifdef FE_INVALID_SQRT\r
++ feraiseexcept (FE_INVALID_SQRT);\r
++\r
++ fenv_union_t u = { .fenv = fegetenv_register () };\r
++ if ((u.l[1] & FE_INVALID) == 0)\r
++#endif\r
++ feraiseexcept (FE_INVALID);\r
++ b = a_nan.value;\r
++ }\r
++ return f_wash (b);\r
++}\r
+Index: libc/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrtf.c\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/powerpc/powerpc64/e5500/fpu/e_sqrtf.c\r
+@@ -0,0 +1,101 @@\r
++/* Single-precision floating point square root.\r
++ Copyright (C) 2010 Free Software Foundation, Inc.\r
++ This file is part of the GNU C Library.\r
++\r
++ The GNU C Library is free software; you can redistribute it and/or\r
++ modify it under the terms of the GNU Lesser General Public\r
++ License as published by the Free Software Foundation; either\r
++ version 2.1 of the License, or (at your option) any later version.\r
++\r
++ The GNU C Library is distributed in the hope that it will be useful,\r
++ but WITHOUT ANY WARRANTY; without even the implied warranty of\r
++ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU\r
++ Lesser General Public License for more details.\r
++\r
++ You should have received a copy of the GNU Lesser General Public\r
++ License along with the GNU C Library; if not, write to the Free\r
++ Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA\r
++ 02111-1307 USA. */\r
++\r
++#include <math.h>\r
++#include <math_private.h>\r
++#include <fenv_libc.h>\r
++#include <inttypes.h>\r
++\r
++#include <sysdep.h>\r
++#include <ldsodefs.h>\r
++\r
++static const ieee_float_shape_type a_nan = {.word = 0x7fc00000 };\r
++static const ieee_float_shape_type a_inf = {.word = 0x7f800000 };\r
++static const float threehalf = 1.5;\r
++\r
++/* The method is based on the descriptions in:\r
++\r
++ _The Handbook of Floating-Pointer Arithmetic_ by Muller et al., chapter 5;\r
++ _IA-64 and Elementary Functions: Speed and Precision_ by Markstein, chapter 9\r
++\r
++ We find the reciprocal square root and use that to compute the actual\r
++ square root. */\r
++\r
++#ifdef __STDC__\r
++float\r
++__ieee754_sqrtf (float b)\r
++#else\r
++float\r
++__ieee754_sqrtf (b)\r
++ float b;\r
++#endif\r
++{\r
++ if (__builtin_expect (b > 0, 1))\r
++ {\r
++#define FMSUB(a_, c_, b_) \\r
++ ({ double __r; \\r
++ __asm__ ("fmsub %[r], %[a], %[c], %[b]\n" \\r
++ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \\r
++ __r;})\r
++#define FNMSUB(a_, c_, b_) \\r
++ ({ double __r; \\r
++ __asm__ ("fnmsub %[r], %[a], %[c], %[b]\n" \\r
++ : [r] "=f" (__r) : [a] "f" (a_), [c] "f" (c_), [b] "f" (b_)); \\r
++ __r;})\r
++\r
++ if (__builtin_expect (b != a_inf.value, 1))\r
++ {\r
++ double y, x;\r
++ fenv_t fe;\r
++\r
++ fe = fegetenv_register ();\r
++\r
++ relax_fenv_state ();\r
++\r
++ /* Compute y = 1.5 * b - b. Uses fewer constants than y = 0.5 * b. */\r
++ y = FMSUB (threehalf, b, b);\r
++\r
++ /* Initial estimate. */\r
++ __asm__ ("frsqrte %[x], %[b]\n" : [x] "=f" (x) : [b] "f" (b));\r
++\r
++ /* Iterate. x_{n+1} = x_n * (1.5 - y * (x_n * x_n)). */\r
++ x = x * FNMSUB (y, x * x, threehalf);\r
++ x = x * FNMSUB (y, x * x, threehalf);\r
++ x = x * FNMSUB (y, x * x, threehalf);\r
++\r
++ /* All done. */\r
++ fesetenv_register (fe);\r
++ return x * b;\r
++ }\r
++ }\r
++ else if (b < 0)\r
++ {\r
++ /* For some reason, some PowerPC32 processors don't implement\r
++ FE_INVALID_SQRT. */\r
++#ifdef FE_INVALID_SQRT\r
++ feraiseexcept (FE_INVALID_SQRT);\r
++\r
++ fenv_union_t u = { .fenv = fegetenv_register () };\r
++ if ((u.l[1] & FE_INVALID) == 0)\r
++#endif\r
++ feraiseexcept (FE_INVALID);\r
++ b = a_nan.value;\r
++ }\r
++ return f_washf (b);\r
++}\r
+Index: libc/sysdeps/unix/sysv/linux/powerpc/powerpc32/603e/fpu/Implies\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/unix/sysv/linux/powerpc/powerpc32/603e/fpu/Implies\r
+@@ -0,0 +1 @@\r
++powerpc/powerpc32/603e/fpu\r
+Index: libc/sysdeps/unix/sysv/linux/powerpc/powerpc32/7400/fpu/Implies\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/unix/sysv/linux/powerpc/powerpc32/7400/fpu/Implies\r
+@@ -0,0 +1 @@\r
++powerpc/powerpc32/603e/fpu\r
+Index: libc/sysdeps/unix/sysv/linux/powerpc/powerpc32/e500mc/fpu/Implies\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/unix/sysv/linux/powerpc/powerpc32/e500mc/fpu/Implies\r
+@@ -0,0 +1 @@\r
++powerpc/powerpc32/603e/fpu\r
+Index: libc/sysdeps/unix/sysv/linux/powerpc/powerpc64/e5500/fpu/Implies\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/unix/sysv/linux/powerpc/powerpc64/e5500/fpu/Implies\r
+@@ -0,0 +1 @@\r
++powerpc/powerpc64/e5500/fpu\r
+Index: libc/sysdeps/unix/sysv/linux/powerpc/powerpc32/e5500/fpu/Implies\r
+===================================================================\r
+--- /dev/null\r
++++ libc/sysdeps/unix/sysv/linux/powerpc/powerpc32/e5500/fpu/Implies\r
+@@ -0,0 +1 @@\r
++powerpc/powerpc32/603e/fpu\r
