0dc9e0996a
As the current libc includes a libm implementation, with the new libc this is needed. Unneeded (for the moment) archs have been removed.
77 lines
1.8 KiB
ArmAsm
77 lines
1.8 KiB
ArmAsm
/*
|
|
* Written by J.T. Conklin <jtc@NetBSD.org>.
|
|
* Public domain.
|
|
*/
|
|
|
|
/*
|
|
* Modified by Lex Wennmacher <wennmach@NetBSD.org>
|
|
* Still public domain.
|
|
*/
|
|
|
|
#include <machine/asm.h>
|
|
|
|
#include "abi.h"
|
|
|
|
RCSID("$NetBSD: s_log1pf.S,v 1.10 2003/09/16 18:17:11 wennmach Exp $")
|
|
|
|
/*
|
|
* The log1pf() function is provided to compute an accurate value of
|
|
* log(1 + x), even for tiny values of x. The i387 FPU provides the
|
|
* fyl2xp1 instruction for this purpose. However, the range of this
|
|
* instruction is limited to:
|
|
* -(1 - (sqrt(2) / 2)) <= x <= sqrt(2) - 1
|
|
* -0.292893 <= x <= 0.414214
|
|
* at least on older processor versions.
|
|
*
|
|
* log1pf() is implemented by testing the range of the argument.
|
|
* If it is appropriate for fyl2xp1, this instruction is used.
|
|
* Else, we compute log1pf(x) = ln(2)*ld(1 + x) the traditional way
|
|
* (using fyl2x).
|
|
*
|
|
* The range testing costs speed, but as the rationale for the very
|
|
* existence of this function is accuracy, we accept that.
|
|
*
|
|
* In order to reduce the cost for testing the range, we check if
|
|
* the argument is in the range
|
|
* -0.25 <= x <= 0.25
|
|
* which can be done with just one conditional branch. If x is
|
|
* inside this range, we use fyl2xp1. Outside of this range,
|
|
* the use of fyl2x is accurate enough.
|
|
*
|
|
*/
|
|
|
|
.text
|
|
.align 4
|
|
ENTRY(log1pf)
|
|
XMM_ONE_ARG_FLOAT_PROLOGUE
|
|
flds ARG_FLOAT_ONE
|
|
fabs
|
|
fld1 /* ... x 1 */
|
|
fadd %st(0) /* ... x 2 */
|
|
fadd %st(0) /* ... x 4 */
|
|
fld1 /* ... 4 1 */
|
|
fdivp /* ... x 0.25 */
|
|
fcompp
|
|
fnstsw %ax
|
|
andb $69,%ah
|
|
jne use_fyl2x
|
|
jmp use_fyl2xp1
|
|
|
|
.align 4
|
|
use_fyl2x:
|
|
fldln2
|
|
flds ARG_FLOAT_ONE
|
|
fld1
|
|
faddp
|
|
fyl2x
|
|
XMM_FLOAT_EPILOGUE
|
|
ret
|
|
|
|
.align 4
|
|
use_fyl2xp1:
|
|
fldln2
|
|
flds ARG_FLOAT_ONE
|
|
fyl2xp1
|
|
XMM_FLOAT_EPILOGUE
|
|
ret
|