2008-06-12 Jakub Jelinek <jakub@redhat.com> * sysdeps/powerpc/powerpc64/fpu/s_llround.S (__llround): Avoid using cr[34] registers. * sysdeps/powerpc/powerpc64/fpu/s_llroundf.S (__llroundf): Likewise. * sysdeps/powerpc/powerpc32/fpu/s_lround.S (__lround): Avoid using cr3 register. 2008-02-18 Steven Munroe <sjmunroe@us.ibm.com> [BZ #5768] * sysdeps/powerpc/powerpc32/fpu/s_isnan.S: New file. * sysdeps/powerpc/powerpc32/fpu/s_isnan.c: Removed. * sysdeps/powerpc/powerpc64/fpu/s_isnan.S: New file. * sysdeps/powerpc/powerpc64/fpu/s_isnan.c: Removed. 2007-11-20 Ryan S. Arnold <rsa@us.ibm.com> [BZ #4997] * sysdeps/powerpc/powerpc32/fpu/s_lround.S (__lround): Fixed erroneous result when x is +/-nextafter(+/-0.5,-/+1) i.e. all 1's in the mantissa. * sysdeps/powerpc/powerpc64/fpu/s_llround.S (__llround): Likewise. Also account for when x is an odd number between 2^52 and 2^53-1. * sysdeps/powerpc/powerpc64/fpu/s_llroundf.S (__llroundf): Likewise. * math/libm-test.inc (lround_test, llround_test): Added test cases to detect aforementioned erroneous conditions. powerpc-cpu/ 2008-06-12 Jakub Jelinek <jakub@redhat.com> * sysdeps/powerpc/powerpc32/powerpc64/fpu/s_llround.S (__llround): Avoid using cr[34] registers. 2008-03-03 Steven Munroe <sjmunroe@us.ibm.com> * sysdeps/powerpc/powerpc32/power6/fpu/s_isnan.S: New file. * sysdeps/powerpc/powerpc32/power6/fpu/s_isnanf.S: New file. * sysdeps/powerpc/powerpc64/power6/fpu/s_isnan.S: New file. 2007-11-20 Ryan S. Arnold <rsa@us.ibm.com> [BZ #4997] * sysdeps/powerpc/powerpc32/powerpc64/fpu/s_llround.S (__llround): Fixed erroneous result when x is +/-nextafter(+/-0.5,-/+1) i.e. all 1's in the mantissa. Also account for when x is an odd number between 2^52 and 2^53-1. 2007-10-26 Steven Munroe <sjmunroe@us.ibm.com> * sysdeps/powerpc/powerpc32/powerpc64/fpu/w_sqrt.S: New File. * sysdeps/powerpc/powerpc32/powerpc64/fpu/w_sqrtf.S: New File. 2007-05-21 Steven Munroe <sjmunroe@us.ibm.com> * sysdeps/powerpc/powerpc32/powerpc64/fpu/slowexp.c: New file. * sysdeps/powerpc/powerpc64/power4/fpu/slowexp.c: New file. * sysdeps/powerpc/powerpc64/power4/fpu/w_sqrt.c: New file. --- libc/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/w_sqrtf.c.jj 2008-08-27 15:23:37.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/w_sqrtf.c 2007-06-03 23:15:58.000000000 +0200 @@ -0,0 +1,60 @@ +/* Single-precision floating point square root wrapper. + Copyright (C) 2004, 2007 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 <sysdep.h> +#include <ldsodefs.h> + +#ifdef __STDC__ +float +__sqrtf (float x) /* wrapper sqrtf */ +#else +float +__sqrtf (x) /* wrapper sqrtf */ + float x; +#endif +{ +#ifdef _IEEE_LIBM + return __ieee754_sqrtf (x); +#else + float z; +/* Power4 (ISA V2.0) and above implement sqrtf in hardware. */ + __asm __volatile ( + " fsqrts %0,%1\n" + : "=f" (z) + : "f" (x)); + + if (__builtin_expect (_LIB_VERSION == _IEEE_, 0)) + return z; + + if (__builtin_expect (x != x, 0)) + return z; + + if (__builtin_expect (x < 0.0, 0)) + /* sqrtf(negative) */ + return (float) __kernel_standard ((double) x, (double) x, 126); + else + return z; +#endif +} + +weak_alias (__sqrtf, sqrtf) --- libc/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/w_sqrt.c.jj 2008-08-27 15:23:34.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/w_sqrt.c 2007-06-03 23:15:53.000000000 +0200 @@ -0,0 +1,62 @@ +/* Double-precision floating point square root wrapper. + Copyright (C) 2004, 2007 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_ldbl_opt.h> +#include "math.h" +#include "math_private.h" +#include <fenv_libc.h> + +#ifdef __STDC__ +double +__sqrt (double x) /* wrapper sqrt */ +#else +double +__sqrt (x) /* wrapper sqrt */ + double x; +#endif +{ + double z; +/* Power4 (ISA V2.0) and above implement sqrt in hardware. */ + __asm __volatile ( + " fsqrt %0,%1\n" + : "=f" (z) + : "f" (x)); +#ifdef _IEEE_LIBM + return z; +#else + if (__builtin_expect (_LIB_VERSION == _IEEE_, 0)) + return z; + + if (__builtin_expect (x != x, 0)) + return z; + + if (__builtin_expect (x < 0.0, 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 --- libc/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/fpu/s_isnan.S.jj 2008-08-27 15:24:46.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc64/power6/fpu/s_isnan.S 2008-04-10 07:58:59.000000000 +0200 @@ -0,0 +1,60 @@ +/* isnan(). PowerPC64 version. + Copyright (C) 2008 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 <sysdep.h> +#include <math_ldbl_opt.h> + +/* int __isnan(x) */ + .machine power6 +EALIGN (__isnan, 4, 0) + CALL_MCOUNT 0 + stfd fp1,-8(r1) /* copy FPR to GPR */ + ori r1,r1,0 + ld r4,-8(r1) + lis r0,0x7ff0 + sldi r0,r0,32 /* const long r0 0x7ff00000 00000000 */ + clrldi r4,r4,1 /* x = fabs(x) */ + cmpd cr7,r4,r0 /* if (fabs(x) <= inf) */ + li r3,0 /* then return 0 */ + blelr+ cr7 + li r3,1 /* else return 1 */ + blr + END (__isnan) + +hidden_def (__isnan) +weak_alias (__isnan, isnan) + +/* It turns out that the 'double' version will also always work for + single-precision. */ +strong_alias (__isnan, __isnanf) +hidden_def (__isnanf) +weak_alias (__isnanf, isnanf) + +#ifdef NO_LONG_DOUBLE +strong_alias (__isnan, __isnanl) +weak_alias (__isnan, isnanl) +#endif + +#ifndef IS_IN_libm +# if LONG_DOUBLE_COMPAT(libc, GLIBC_2_0) +compat_symbol (libc, __isnan, __isnanl, GLIBC_2_0); +compat_symbol (libc, isnan, isnanl, GLIBC_2_0); +# endif +#endif + --- libc/powerpc-cpu/sysdeps/powerpc/powerpc32/powerpc64/fpu/w_sqrt.S.jj 2008-08-27 15:10:03.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc32/powerpc64/fpu/w_sqrt.S 2008-04-10 06:52:55.000000000 +0200 @@ -0,0 +1,119 @@ +/* sqrt function. PowerPC32 version. + Copyright (C) 2007 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., 1 Franklin Street, Fifth Floor, Boston MA + 02110-1301 USA. */ + +#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. Extented 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 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) + cfi_offset(r30,8) +#ifdef SHARED +# ifdef HAVE_ASM_PPC_REL16 + bcl 20,31,.LCF1 +.LCF1: + mflr r30 + addis r30,r30,_GLOBAL_OFFSET_TABLE_-.LCF1@ha + addi r30,r30,_GLOBAL_OFFSET_TABLE_-.LCF1@l + lwz r9,_LIB_VERSION@got(30) + lwz r0,0(r9) +# else + bl _GLOBAL_OFFSET_TABLE_@local-4 + mflr r30 + lwz r9,_LIB_VERSION@got(30) + lwz r0,0(r9) +# endif +#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 + --- libc/powerpc-cpu/sysdeps/powerpc/powerpc32/powerpc64/fpu/w_sqrtf.S.jj 2008-08-27 15:10:06.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc32/powerpc64/fpu/w_sqrtf.S 2008-04-10 06:53:05.000000000 +0200 @@ -0,0 +1,111 @@ +/* sqrtf function. PowerPC32 version. + Copyright (C) 2007 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., 1 Franklin Street, Fifth Floor, Boston MA + 02110-1301 USA. */ + +#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. Extented 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 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) + cfi_offset(r30,8) +#ifdef SHARED +# ifdef HAVE_ASM_PPC_REL16 + bcl 20,31,.LCF1 +.LCF1: + mflr r30 + addis r30,r30,_GLOBAL_OFFSET_TABLE_-.LCF1@ha + addi r30,r30,_GLOBAL_OFFSET_TABLE_-.LCF1@l + lwz r9,_LIB_VERSION@got(30) + lwz r0,0(r9) +# else + bl _GLOBAL_OFFSET_TABLE_@local-4 + mflr r30 + lwz r9,_LIB_VERSION@got(30) + lwz r0,0(r9) +# endif +#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) + --- libc/powerpc-cpu/sysdeps/powerpc/powerpc32/powerpc64/fpu/s_llround.S.jj 2008-08-27 11:00:42.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc32/powerpc64/fpu/s_llround.S 2008-08-13 14:05:38.000000000 +0200 @@ -1,5 +1,5 @@ /* llround function. PowerPC32 on PowerPC64 version. - Copyright (C) 2004, 2006 Free Software Foundation, Inc. + Copyright (C) 2004, 2006, 2007, 2008 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 @@ -20,15 +20,15 @@ #include <sysdep.h> #include <math_ldbl_opt.h> - .section .rodata.cst8,"aM",@progbits,8 - .align 2 -.LC0: /* 0.0 */ + .section .rodata.cst12,"aM",@progbits,12 + .align 3 + .LC0: /* 0x1.0000000000000p+52 == 2^52 */ + .long 0x43300000 .long 0x00000000 -.LC1: /* 0.5 */ - .long 0x3f000000 + .long 0x3f000000 /* Use this for 0.5 */ .section ".text" - + /* long [r3] lround (float x [fp1]) IEEE 1003.1 lround function. IEEE specifies "round to the nearest integer value, rounding halfway cases away from zero, regardless of @@ -37,7 +37,15 @@ tie, choose the one that is even (least significant bit o).". So we can't use the PowerPC "round to Nearest" mode. Instead we set "round toward Zero" mode and round by adding +-0.5 before rounding - to the integer value. */ + to the integer value. + + It is necessary to detect when x is (+-)0x1.fffffffffffffp-2 + because adding +-0.5 in this case will cause an erroneous shift, + carry and round. We simply return 0 if 0.5 > x > -0.5. Likewise + if x is and odd number between +-(2^52 and 2^53-1) a shift and + carry will erroneously round if biased with +-0.5. Therefore if x + is greater/less than +-2^52 we don't need to bias the number with + +-0.5. */ ENTRY (__llround) stwu r1,-16(r1) @@ -57,30 +65,41 @@ ENTRY (__llround) # endif mtlr r11 cfi_same_value (lr) - lfs fp12,0(r9) - lfs fp10,.LC1-.LC0(r9) + lfd fp9,0(r9) + lfs fp10,8(r9) #else - lis r9,.LC0@ha - lis r10,.LC1@ha - lfs fp12,.LC0@l(r9) - lfs fp10,.LC1@l(r10) + lis r9,.LC0@ha + lfd fp9,.LC0@l(r9) /* Load 2^52 into fpr9. */ + lfs fp10,.LC0@l+8(r9) /* Load 0.5 into fpr10. */ #endif - fcmpu cr6,fp1,fp12 /* if (x > 0.0) */ - ble- cr6,.L4 - fadd fp1,fp1,fp10 /* x+= 0.5; */ -.L9: - fctidz fp2,fp1 /* Convert To Integer DW round toward 0. */ - stfd fp2,8(r1) - nop /* Ensure the following load is in a different dispatch */ - nop /* group to avoid pipe stall on POWER4&5. */ + fabs fp2,fp1 /* Get the absolute value of x. */ + fsub fp12,fp10,fp10 /* Compute 0.0 into fpr12. */ + fcmpu cr6,fp2,fp10 /* if |x| < 0.5 */ + fcmpu cr7,fp2,fp9 /* if |x| >= 2^52 */ + fcmpu cr1,fp1,fp12 /* x is negative? x < 0.0 */ + blt- cr6,.Lretzero /* 0.5 > x < -0.5 so just return 0. */ + bge- cr7,.Lnobias /* 2^52 > x < -2^52 just convert with no bias. */ + fadd fp3,fp2,fp10 /* |x|+=0.5 bias to prepare to round. */ + bge cr1,.Lconvert /* x is positive so don't negate x. */ + fnabs fp3,fp3 /* -(|x|+=0.5) */ +.Lconvert: + fctidz fp4,fp3 /* Convert to Integer double word round toward 0. */ + stfd fp4,8(r1) + nop + nop nop - lwz r4,12(r1) + lwz r4,12(r1) /* Load return as integer. */ lwz r3,8(r1) +.Lout: addi r1,r1,16 blr -.L4: - fsub fp1,fp1,fp10 /* x-= 0.5; */ - b .L9 +.Lretzero: /* 0.5 > x > -0.5 */ + li r3,0 /* return 0. */ + li r4,0 + b .Lout +.Lnobias: + fmr fp3,fp1 + b .Lconvert END (__llround) weak_alias (__llround, llround) --- libc/powerpc-cpu/sysdeps/powerpc/powerpc32/power6/fpu/s_isnan.S.jj 2008-08-27 15:18:02.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc32/power6/fpu/s_isnan.S 2008-04-10 07:58:20.000000000 +0200 @@ -0,0 +1,62 @@ +/* isnan(). PowerPC32 version. + Copyright (C) 2008 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 <sysdep.h> +#include <math_ldbl_opt.h> + +/* int __isnan(x) */ + .machine power6 +EALIGN (__isnan, 4, 0) + stwu r1,-32(r1) + cfi_adjust_cfa_offset (32) + ori r1,r1,0 + stfd fp1,24(r1) /* copy FPR to GPR */ + ori r1,r1,0 + lwz r4,24(r1) + lwz r5,28(r1) + lis r0,0x7ff0 /* const long r0 0x7ff00000 00000000 */ + clrlwi r4,r4,1 /* x = fabs(x) */ + cmpw cr7,r4,r0 /* if (fabs(x) =< inf) */ + cmpwi cr6,r5,0 + li r3,0 /* then return 0 */ + addi r1,r1,32 + cfi_adjust_cfa_offset (-32) + bltlr+ cr7 + bgt- cr7,L(NaN) + beqlr+ cr6 +L(NaN): + li r3,1 /* else return 1 */ + blr + END (__isnan) + +hidden_def (__isnan) +weak_alias (__isnan, isnan) + +#ifdef NO_LONG_DOUBLE +strong_alias (__isnan, __isnanl) +weak_alias (__isnan, isnanl) +#endif + +#ifndef IS_IN_libm +# if LONG_DOUBLE_COMPAT(libc, GLIBC_2_0) +compat_symbol (libc, __isnan, __isnanl, GLIBC_2_0); +compat_symbol (libc, isnan, isnanl, GLIBC_2_0); +# endif +#endif + --- libc/powerpc-cpu/sysdeps/powerpc/powerpc32/power6/fpu/s_isnanf.S.jj 2008-08-27 15:18:05.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc32/power6/fpu/s_isnanf.S 2008-04-10 07:58:33.000000000 +0200 @@ -0,0 +1,46 @@ +/* isnanf(). PowerPC32 version. + Copyright (C) 2008 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 <sysdep.h> +#include <math_ldbl_opt.h> + +/* int __isnanf(x) */ + .machine power6 +EALIGN (__isnanf, 4, 0) + stwu r1,-32(r1) + cfi_adjust_cfa_offset (32) + ori r1,r1,0 + stfs fp1,24(r1) /* copy FPR to GPR */ + ori r1,r1,0 + lwz r4,24(r1) + lis r0,0x7f80 /* const long r0 0x7f800000 */ + clrlwi r4,r4,1 /* x = fabs(x) */ + cmpw cr7,r4,r0 /* if (fabs(x) =< inf) */ + li r3,0 /* then return 0 */ + addi r1,r1,32 + cfi_adjust_cfa_offset (-32) + blelr+ cr7 +L(NaN): + li r3,1 /* else return 1 */ + blr + END (__isnan) + +hidden_def (__isnanf) +weak_alias (__isnanf, isnanf) + --- libc/powerpc-cpu/sysdeps/powerpc/powerpc32/power6/fpu/s_llroundf.S.jj 2008-08-27 11:00:43.000000000 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc32/power6/fpu/s_llroundf.S 2008-08-27 15:20:38.000000000 +0200 @@ -1,2 +1 @@ /* __llroundf is in s_llround.S */ -/* __llroundf is in s_llround.S */ --- libc/sysdeps/powerpc/powerpc64/fpu/s_isnan.c.jj 2006-01-28 01:07:31.000000000 +0100 +++ libc/sysdeps/powerpc/powerpc64/fpu/s_isnan.c 2008-08-27 15:15:44.000000000 +0200 @@ -1,7 +0,0 @@ -#include <sysdeps/powerpc/fpu/s_isnan.c> -#ifndef IS_IN_libm -# if LONG_DOUBLE_COMPAT(libc, GLIBC_2_0) -compat_symbol (libc, __isnan, __isnanl, GLIBC_2_0); -compat_symbol (libc, isnan, isnanl, GLIBC_2_0); -# endif -#endif --- libc/sysdeps/powerpc/powerpc64/fpu/s_isnan.S.jj 2008-08-27 15:15:50.000000000 +0200 +++ libc/sysdeps/powerpc/powerpc64/fpu/s_isnan.S 2008-04-10 16:15:43.000000000 +0200 @@ -0,0 +1,57 @@ +/* isnan(). PowerPC64 version. + Copyright (C) 2008 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 <sysdep.h> +#include <math_ldbl_opt.h> + +/* int __isnan(x) */ + .machine power4 +EALIGN (__isnan, 4, 0) + CALL_MCOUNT 0 + mffs fp0 + mtfsb0 4*cr6+lt /* reset_fpscr_bit (FPSCR_VE) */ + fcmpu cr7,fp1,fp1 + mtfsf 255,fp0 + li r3,0 + beqlr+ cr7 /* (x == x) then not a NAN */ + li r3,1 /* else must be a NAN */ + blr + END (__isnan) + +hidden_def (__isnan) +weak_alias (__isnan, isnan) + +/* It turns out that the 'double' version will also always work for + single-precision. */ +strong_alias (__isnan, __isnanf) +hidden_def (__isnanf) +weak_alias (__isnanf, isnanf) + +#ifdef NO_LONG_DOUBLE +strong_alias (__isnan, __isnanl) +weak_alias (__isnan, isnanl) +#endif + +#ifndef IS_IN_libm +# if LONG_DOUBLE_COMPAT(libc, GLIBC_2_0) +compat_symbol (libc, __isnan, __isnanl, GLIBC_2_0); +compat_symbol (libc, isnan, isnanl, GLIBC_2_0); +# endif +#endif + --- libc/sysdeps/powerpc/powerpc64/fpu/s_llround.S.jj 2006-01-28 01:07:40.000000000 +0100 +++ libc/sysdeps/powerpc/powerpc64/fpu/s_llround.S 2008-08-13 14:05:38.000000000 +0200 @@ -1,5 +1,5 @@ /* llround function. PowerPC64 version. - Copyright (C) 2004, 2006 Free Software Foundation, Inc. + Copyright (C) 2004, 2006, 2007, 2008 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 @@ -21,13 +21,13 @@ #include <math_ldbl_opt.h> .section ".toc","aw" -.LC0: /* -0.0 */ - .tc FD_00000000_0[TC],0x0000000000000000 +.LC0: /* 2^52 */ + .tc FD_43300000_0[TC],0x4330000000000000 .LC1: /* 0.5 */ .tc FD_3fe00000_0[TC],0x3fe0000000000000 .section ".text" -/* long long [r3] llround (float x [fp1]) +/* long long [r3] llround (double x [fp1]) IEEE 1003.1 llround function. IEEE specifies "round to the nearest integer value, rounding halfway cases away from zero, regardless of the current rounding mode." However PowerPC Architecture defines @@ -35,26 +35,45 @@ tie, choose the one that is even (least significant bit o).". So we can't use the PowerPC "round to Nearest" mode. Instead we set "round toward Zero" mode and round by adding +-0.5 before rounding - to the integer value. */ + to the integer value. + + It is necessary to detect when x is (+-)0x1.fffffffffffffp-2 + because adding +-0.5 in this case will cause an erroneous shift, + carry and round. We simply return 0 if 0.5 > x > -0.5. Likewise + if x is and odd number between +-(2^52 and 2^53-1) a shift and + carry will erroneously round if biased with +-0.5. Therefore if x + is greater/less than +-2^52 we don't need to bias the number with + +-0.5. */ ENTRY (__llround) CALL_MCOUNT 0 - lfd fp12,.LC0@toc(2) - lfd fp10,.LC1@toc(2) - fcmpu cr6,fp1,fp12 /* if (x > 0.0) */ - ble- cr6,.L4 - fadd fp1,fp1,fp10 /* x+= 0.5; */ -.L9: - fctidz fp2,fp1 /* Convert To Integer DW llround toward 0. */ - stfd fp2,-16(r1) - nop /* Insure the following load is in a different dispatch group */ - nop /* to avoid pipe stall on POWER4&5. */ + lfd fp9,.LC0@toc(2) /* Load 2^52 into fpr9. */ + lfd fp10,.LC1@toc(2)/* Load 0.5 into fpr10. */ + fabs fp2,fp1 /* Get the absolute value of x. */ + fsub fp12,fp10,fp10 /* Compute 0.0 into fp12. */ + fcmpu cr6,fp2,fp10 /* if |x| < 0.5 */ + fcmpu cr7,fp2,fp9 /* if |x| >= 2^52 */ + fcmpu cr1,fp1,fp12 /* x is negative? x < 0.0 */ + blt- cr6,.Lretzero /* 0.5 > x < -0.5 so just return 0. */ + bge- cr7,.Lnobias /* 2^52 > x < -2^52 just convert with no bias. */ + fadd fp3,fp2,fp10 /* |x|+=0.5 bias to prepare to round. */ + bge cr1,.Lconvert /* x is positive so don't negate x. */ + fnabs fp3,fp3 /* -(|x|+=0.5) */ +.Lconvert: + fctidz fp4,fp3 /* Convert to Integer double word round toward 0. */ + stfd fp4,-16(r1) + nop + nop nop - ld r3,-16(r1) + ld r3,-16(r1) /* Load return as integer. */ +.Lout: blr -.L4: - fsub fp1,fp1,fp10 /* x-= 0.5; */ - b .L9 +.Lretzero: /* 0.5 > x > -0.5 */ + li r3,0 /* return 0. */ + b .Lout +.Lnobias: + fmr fp3,fp1 + b .Lconvert END (__llround) strong_alias (__llround, __lround) --- libc/sysdeps/powerpc/powerpc64/fpu/s_llroundf.S.jj 2006-01-28 01:14:39.000000000 +0100 +++ libc/sysdeps/powerpc/powerpc64/fpu/s_llroundf.S 2008-08-13 14:05:38.000000000 +0200 @@ -1,5 +1,5 @@ /* llroundf function. PowerPC64 version. - Copyright (C) 2004, 2006 Free Software Foundation, Inc. + Copyright (C) 2004, 2006, 2007, 2008 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 @@ -20,8 +20,8 @@ #include <sysdep.h> .section ".toc","aw" -.LC0: /* -0.0 */ - .tc FD_00000000_0[TC],0x0000000000000000 +.LC0: /* 2^23 */ + .tc FD_41600000_0[TC],0x4160000000000000 .LC1: /* 0.5 */ .tc FD_3fe00000_0[TC],0x3fe0000000000000 .section ".text" @@ -34,24 +34,45 @@ tie, choose the one that is even (least significant bit o).". So we can't use the PowerPC "round to Nearest" mode. Instead we set "round toward Zero" mode and round by adding +-0.5 before rounding - to the integer value. */ + to the integer value. + + It is necessary to detect when x is (+-)0x1.fffffffffffffp-2 + because adding +-0.5 in this case will cause an erroneous shift, + carry and round. We simply return 0 if 0.5 > x > -0.5. Likewise + if x is and odd number between +-(2^23 and 2^24-1) a shift and + carry will erroneously round if biased with +-0.5. Therefore if x + is greater/less than +-2^23 we don't need to bias the number with + +-0.5. */ ENTRY (__llroundf) CALL_MCOUNT 0 - lfd fp12,.LC0@toc(2) - lfd fp10,.LC1@toc(2) - fcmpu cr6,fp1,fp12 /* if (x < 0.0) */ - fsubs fp3,fp1,fp10 /* x-= 0.5; */ - ble- cr6,.L9 - fadds fp3,fp1,fp10 /* x+= 0.5; */ -.L9: - fctidz fp2,fp3 /* Convert To Integer DW round toward 0. */ - stfd fp2,-16(r1) - nop /* Insure the following load is in a different dispatch group */ - nop /* to avoid pipe stall on POWER4&5. */ + lfd fp9,.LC0@toc(2) /* Load 2^23 into fpr9. */ + lfd fp10,.LC1@toc(2)/* Load 0.5 into fpr10. */ + fabs fp2,fp1 /* Get the absolute value of x. */ + fsub fp12,fp10,fp10 /* Compute 0.0 into fp12. */ + fcmpu cr6,fp2,fp10 /* if |x| < 0.5 */ + fcmpu cr7,fp2,fp9 /* if |x| >= 2^23 */ + fcmpu cr1,fp1,fp12 /* x is negative? x < 0.0 */ + blt- cr6,.Lretzero /* 0.5 > x < -0.5 so just return 0. */ + bge- cr7,.Lnobias /* 2^23 > x < -2^23 just convert with no bias. */ + fadd fp3,fp2,fp10 /* |x|+=0.5 bias to prepare to round. */ + bge cr1,.Lconvert /* x is positive so don't negate x. */ + fnabs fp3,fp3 /* -(|x|+=0.5) */ +.Lconvert: + fctidz fp4,fp3 /* Convert to Integer double word round toward 0. */ + stfd fp4,-16(r1) + nop + nop nop - ld r3,-16(r1) + ld r3,-16(r1) /* Load return as integer. */ +.Lout: blr +.Lretzero: /* 0.5 > x > -0.5 */ + li r3,0 /* return 0. */ + b .Lout +.Lnobias: + fmr fp3,fp1 + b .Lconvert END (__llroundf) strong_alias (__llroundf, __lroundf) --- libc/sysdeps/powerpc/powerpc32/fpu/s_isnan.c.jj 2006-01-28 01:07:27.000000000 +0100 +++ libc/sysdeps/powerpc/powerpc32/fpu/s_isnan.c 2008-08-27 15:15:14.000000000 +0200 @@ -1,7 +0,0 @@ -#include <sysdeps/powerpc/fpu/s_isnan.c> -#ifndef IS_IN_libm -# if LONG_DOUBLE_COMPAT(libc, GLIBC_2_0) -compat_symbol (libc, __isnan, __isnanl, GLIBC_2_0); -compat_symbol (libc, isnan, isnanl, GLIBC_2_0); -# endif -#endif --- libc/sysdeps/powerpc/powerpc32/fpu/s_lround.S.jj 2006-01-28 01:07:39.000000000 +0100 +++ libc/sysdeps/powerpc/powerpc32/fpu/s_lround.S 2008-08-13 14:05:38.000000000 +0200 @@ -1,5 +1,5 @@ /* lround function. PowerPC32 version. - Copyright (C) 2004, 2006 Free Software Foundation, Inc. + Copyright (C) 2004, 2006, 2007, 2008 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 @@ -20,13 +20,10 @@ #include <sysdep.h> #include <math_ldbl_opt.h> - .section .rodata.cst8,"aM",@progbits,8 + .section .rodata.cst4,"aM",@progbits,4 .align 2 -.LC0: /* 0.0 */ - .long 0x00000000 -.LC1: /* 0.5 */ +.LC0: /* 0.5 */ .long 0x3f000000 - .section ".text" /* long [r3] lround (float x [fp1]) @@ -37,7 +34,10 @@ tie, choose the one that is even (least significant bit o).". So we can't use the PowerPC "round to Nearest" mode. Instead we set "round toward Zero" mode and round by adding +-0.5 before rounding - to the integer value. */ + to the integer value. It is necessary to detect when x is + (+-)0x1.fffffffffffffp-2 because adding +-0.5 in this case will + cause an erroneous shift, carry and round. We simply return 0 if + 0.5 > x > -0.5. */ ENTRY (__lround) stwu r1,-16(r1) @@ -49,40 +49,40 @@ ENTRY (__lround) bcl 20,31,1f 1: mflr r9 addis r9,r9,.LC0-1b@ha - addi r9,r9,.LC0-1b@l + lfs fp10,.LC0-1b@l(r9) # else bl _GLOBAL_OFFSET_TABLE_@local-4 mflr r10 lwz r9,.LC0@got(10) + lfs fp10,0(r9) # endif mtlr r11 cfi_same_value (lr) - lfs fp12,0(r9) #else lis r9,.LC0@ha - lfs fp12,.LC0@l(r9) -#endif -#ifdef SHARED - lfs fp10,.LC1-.LC0(r9) -#else - lis r9,.LC1@ha - lfs fp10,.LC1@l(r9) + lfs fp10,.LC0@l(r9) #endif - fcmpu cr6,fp1,fp12 /* if (x > 0.0) */ - ble- cr6,.L4 - fadd fp1,fp1,fp10 /* x+= 0.5; */ -.L9: - fctiwz fp2,fp1 /* Convert To Integer DW lround toward 0. */ - stfd fp2,8(r1) + fabs fp2, fp1 /* Get the absolute value of x. */ + fsub fp12,fp10,fp10 /* Compute 0.0. */ + fcmpu cr6, fp2, fp10 /* if |x| < 0.5 */ + fcmpu cr7, fp1, fp12 /* x is negative? x < 0.0 */ + blt- cr6,.Lretzero + fadd fp3,fp2,fp10 /* |x|+=0.5 bias to prepare to round. */ + bge cr7,.Lconvert /* x is positive so don't negate x. */ + fnabs fp3,fp3 /* -(|x|+=0.5) */ +.Lconvert: + fctiwz fp4,fp3 /* Convert to Integer word lround toward 0. */ + stfd fp4,8(r1) nop /* Ensure the following load is in a different dispatch */ nop /* group to avoid pipe stall on POWER4&5. */ nop - lwz r3,12(r1) + lwz r3,12(r1) /* Load return as integer. */ +.Lout: addi r1,r1,16 blr -.L4: - fsub fp1,fp1,fp10 /* x-= 0.5; */ - b .L9 +.Lretzero: /* when 0.5 > x > -0.5 */ + li r3,0 /* return 0. */ + b .Lout END (__lround) weak_alias (__lround, lround) --- libc/sysdeps/powerpc/powerpc32/fpu/s_isnan.S.jj 2008-08-27 15:15:21.000000000 +0200 +++ libc/sysdeps/powerpc/powerpc32/fpu/s_isnan.S 2008-04-10 16:15:27.000000000 +0200 @@ -0,0 +1,56 @@ +/* isnan(). PowerPC32 version. + Copyright (C) 2008 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 <sysdep.h> +#include <math_ldbl_opt.h> + +/* int __isnan(x) */ + .machine power4 +EALIGN (__isnan, 4, 0) + mffs fp0 + mtfsb0 4*cr6+lt /* reset_fpscr_bit (FPSCR_VE) */ + fcmpu cr7,fp1,fp1 + mtfsf 255,fp0 + li r3,0 + beqlr+ cr7 /* (x == x) then not a NAN */ + li r3,1 /* else must be a NAN */ + blr + END (__isnan) + +hidden_def (__isnan) +weak_alias (__isnan, isnan) + +/* It turns out that the 'double' version will also always work for + single-precision. */ +strong_alias (__isnan, __isnanf) +hidden_def (__isnanf) +weak_alias (__isnanf, isnanf) + +#ifdef NO_LONG_DOUBLE +strong_alias (__isnan, __isnanl) +weak_alias (__isnan, isnanl) +#endif + +#ifndef IS_IN_libm +# if LONG_DOUBLE_COMPAT(libc, GLIBC_2_0) +compat_symbol (libc, __isnan, __isnanl, GLIBC_2_0); +compat_symbol (libc, isnan, isnanl, GLIBC_2_0); +# endif +#endif + --- libc/powerpc-cpu/sysdeps/powerpc/powerpc32/powerpc64/fpu/slowexp.c 2008-05-09 14:58:04.646001045 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc32/powerpc64/fpu/slowexp.c 2007-06-03 22:51:17.000000000 +0200 @@ -0,0 +1,66 @@ +/* + * IBM Accurate Mathematical Library + * written by International Business Machines Corp. + * Copyright (C) 2001, 2007 Free Software Foundation + * + * This program 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. + * + * This program 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 this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. + */ +/**************************************************************************/ +/* MODULE_NAME:slowexp.c */ +/* */ +/* FUNCTION:slowexp */ +/* */ +/* FILES NEEDED:mpa.h */ +/* mpa.c mpexp.c */ +/* */ +/*Converting from double precision to Multi-precision and calculating */ +/* e^x */ +/**************************************************************************/ +#include "math_private.h" + +#ifdef NO_LONG_DOUBLE +#include "mpa.h" +void __mpexp(mp_no *x, mp_no *y, int p); +#endif + +/*Converting from double precision to Multi-precision and calculating e^x */ +double __slowexp(double x) { +#ifdef NO_LONG_DOUBLE + double w,z,res,eps=3.0e-26; + int p; + mp_no mpx, mpy, mpz,mpw,mpeps,mpcor; + + p=6; + __dbl_mp(x,&mpx,p); /* Convert a double precision number x */ + /* into a multiple precision number mpx with prec. p. */ + __mpexp(&mpx, &mpy, p); /* Multi-Precision exponential function */ + __dbl_mp(eps,&mpeps,p); + __mul(&mpeps,&mpy,&mpcor,p); + __add(&mpy,&mpcor,&mpw,p); + __sub(&mpy,&mpcor,&mpz,p); + __mp_dbl(&mpw, &w, p); + __mp_dbl(&mpz, &z, p); + if (w == z) return w; + else { /* if calculating is not exactly */ + p = 32; + __dbl_mp(x,&mpx,p); + __mpexp(&mpx, &mpy, p); + __mp_dbl(&mpy, &res, p); + return res; + } +#else + return (double) __ieee754_expl((long double)x); +#endif +} --- libc/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/slowexp.c 2008-05-09 14:58:04.646001045 +0200 +++ libc/powerpc-cpu/sysdeps/powerpc/powerpc64/power4/fpu/slowexp.c 2007-06-03 22:51:17.000000000 +0200 @@ -0,0 +1,66 @@ +/* + * IBM Accurate Mathematical Library + * written by International Business Machines Corp. + * Copyright (C) 2001, 2007 Free Software Foundation + * + * This program 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. + * + * This program 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 this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. + */ +/**************************************************************************/ +/* MODULE_NAME:slowexp.c */ +/* */ +/* FUNCTION:slowexp */ +/* */ +/* FILES NEEDED:mpa.h */ +/* mpa.c mpexp.c */ +/* */ +/*Converting from double precision to Multi-precision and calculating */ +/* e^x */ +/**************************************************************************/ +#include "math_private.h" + +#ifdef NO_LONG_DOUBLE +#include "mpa.h" +void __mpexp(mp_no *x, mp_no *y, int p); +#endif + +/*Converting from double precision to Multi-precision and calculating e^x */ +double __slowexp(double x) { +#ifdef NO_LONG_DOUBLE + double w,z,res,eps=3.0e-26; + int p; + mp_no mpx, mpy, mpz,mpw,mpeps,mpcor; + + p=6; + __dbl_mp(x,&mpx,p); /* Convert a double precision number x */ + /* into a multiple precision number mpx with prec. p. */ + __mpexp(&mpx, &mpy, p); /* Multi-Precision exponential function */ + __dbl_mp(eps,&mpeps,p); + __mul(&mpeps,&mpy,&mpcor,p); + __add(&mpy,&mpcor,&mpw,p); + __sub(&mpy,&mpcor,&mpz,p); + __mp_dbl(&mpw, &w, p); + __mp_dbl(&mpz, &z, p); + if (w == z) return w; + else { /* if calculating is not exactly */ + p = 32; + __dbl_mp(x,&mpx,p); + __mpexp(&mpx, &mpy, p); + __mp_dbl(&mpy, &res, p); + return res; + } +#else + return (double) __ieee754_expl((long double)x); +#endif +} --- libc/math/libm-test.inc 24 Mar 2008 19:57:48 -0000 1.78 +++ libc/math/libm-test.inc 11 Apr 2008 19:32:30 -0000 1.79 @@ -4300,6 +4300,17 @@ lround_test (void) # endif TEST_f_l (lround, 2097152.5, 2097153); TEST_f_l (lround, -2097152.5, -2097153); + /* nextafter(0.5,-1) */ + TEST_f_l (lround, 0x1.fffffffffffffp-2, 0); + /* nextafter(-0.5,1) */ + TEST_f_l (lround, -0x1.fffffffffffffp-2, 0); +#else + /* nextafter(0.5,-1) */ + TEST_f_l (lround, 0x1.fffffp-2, 0); + /* nextafter(-0.5,1) */ + TEST_f_l (lround, -0x1.fffffp-2, 0); + TEST_f_l (lround, 0x1.fffffep+23, 16777215); + TEST_f_l (lround, -0x1.fffffep+23, -16777215); #endif END (lround); } @@ -4359,8 +4370,40 @@ llround_test (void) TEST_f_L (llround, 4294967295.5, 4294967296LL); /* 0x200000000 */ TEST_f_L (llround, 8589934591.5, 8589934592LL); + + /* nextafter(0.5,-1) */ + TEST_f_L (llround, 0x1.fffffffffffffp-2, 0); + /* nextafter(-0.5,1) */ + TEST_f_L (llround, -0x1.fffffffffffffp-2, 0); + /* On PowerPC an exponent of '52' is the largest incrementally + * representable sequence of whole-numbers in the 'double' range. We test + * lround to make sure that a guard bit set during the lround operation + * hasn't forced an erroneous shift giving us an incorrect result. The odd + * numbers between +-(2^52+1 and 2^53-1) are affected since they have the + * rightmost bit set. */ + /* +-(2^52+1) */ + TEST_f_L (llround, 0x1.0000000000001p+52,4503599627370497LL); + TEST_f_L (llround, -0x1.0000000000001p+52,-4503599627370497LL); + /* +-(2^53-1): Input is the last (positive and negative) incrementally + * representable whole-number in the 'double' range that might round + * erroneously. */ + TEST_f_L (llround, 0x1.fffffffffffffp+52, 9007199254740991LL); + TEST_f_L (llround, -0x1.fffffffffffffp+52, -9007199254740991LL); +#else + /* nextafter(0.5,-1) */ + TEST_f_L (llround, 0x1.fffffep-2, 0); + /* nextafter(-0.5,1) */ + TEST_f_L (llround, -0x1.fffffep-2, 0); + /* As above, on PowerPC an exponent of '23' is the largest incrementally + * representable sequence of whole-numbers in the 'float' range. + * Likewise, numbers between +-(2^23+1 and 2^24-1) are affected. */ + TEST_f_L (llround, 0x1.000002p+23,8388609); + TEST_f_L (llround, -0x1.000002p+23,-8388609); + TEST_f_L (llround, 0x1.fffffep+23, 16777215); + TEST_f_L (llround, -0x1.fffffep+23, -16777215); #endif + #ifdef TEST_LDOUBLE /* The input can only be represented in long double. */ TEST_f_L (llround, 4503599627370495.5L, 4503599627370496LL);