25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis/*
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * CDDL HEADER START
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis *
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * The contents of this file are subject to the terms of the
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * Common Development and Distribution License (the "License").
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * You may not use this file except in compliance with the License.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis *
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * or http://www.opensolaris.org/os/licensing.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * See the License for the specific language governing permissions
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * and limitations under the License.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis *
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * When distributing Covered Code, include this CDDL HEADER in each
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * If applicable, add the following below this CDDL HEADER, with the
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * fields enclosed by brackets "[]" replaced with your own identifying
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * information: Portions Copyright [yyyy] [name of copyright owner]
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis *
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * CDDL HEADER END
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis */
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis/*
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis */
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis/*
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * Copyright 2005 Sun Microsystems, Inc. All rights reserved.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis * Use is subject to license terms.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis */
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis .file "expm1f.s"
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis#include "libm.h"
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr JasiukajtisLIBM_ANSI_PRAGMA_WEAK(expm1f,function)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis .data
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis .align 4
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis.mhundred: .float -100.0
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis ENTRY(expm1f)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movl 4(%esp),%ecx / ecx <-- x
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis andl $0x7fffffff,%ecx / ecx <-- |x|
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis cmpl $0x3f317217,%ecx / Is |x| < ln(2)?
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis jbe .shortcut / If so, take a shortcut.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis cmpl $0x7f800000,%ecx / |x| >= INF?
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis jae .not_finite / if so, x is not finite
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis flds 4(%esp) / push x
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis subl $8,%esp / save RP and set round-to-64-bits
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstcw (%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw (%esp),%ax
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw %ax,4(%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis orw $0x0300,%ax
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw %ax,(%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fldcw (%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fldl2e / push log2e }not for xtndd_dbl
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fmulp %st,%st(1) / z = x*log2e }not for xtndd_dbl
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fld %st(0) / duplicate stack top
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis frndint / [z],z
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fucom / This and the next 3 instructions
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstsw %ax / add 10 clocks to runtime of the
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis sahf / main branch, but save about 265
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis je .z_integral / upon detection of integral z.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis / [z] != 0, compute exp(x) and then subtract one to get expm1(x)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fxch / z,[z]
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fsub %st(1),%st / z-[z],[z]
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis f2xm1 / 2**(z-[z])-1,[z]
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis / avoid spurious underflow when scaling to compute exp(x)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis PIC_SETUP(1)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis flds PIC_L(.mhundred)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis PIC_WRAPUP
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fucom %st(2) / if -100 !< [z], then use -100
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstsw %ax
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis sahf
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis jb .got_int_part
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fxch %st(2)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis.got_int_part:
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstp %st(0) / 2**(z-[z])-1,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fld1 / 1,2**(z-[z])-1,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis faddp %st,%st(1) / 2**(z-[z]) ,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fscale / exp(x) ,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fld1 / 1,exp(x) ,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fsubrp %st,%st(1) / exp(x)-1 ,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstp %st(1)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstcw (%esp) / restore old RP
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw (%esp),%dx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis andw $0xfcff,%dx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw 4(%esp),%cx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis andw $0x0300,%cx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis orw %dx,%cx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw %cx,(%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fldcw (%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis add $8,%esp
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis ret
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis.z_integral: / here, z is integral
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstp %st(0) / ,z
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis / avoid spurious underflow when scaling to compute exp(x)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis PIC_SETUP(2)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis flds PIC_L(.mhundred)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis PIC_WRAPUP
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fucom %st(1) / if -100 !< [z], then use -100
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstsw %ax
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis sahf
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis jb .scale_wont_ovfl
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fxch %st(1)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis.scale_wont_ovfl:
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstp %st(0) / max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fld1 / 1,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fscale / exp(x) ,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fld1 / 1,exp(x) ,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fsubrp %st,%st(1) / exp(x)-1 ,max([z],-100)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstp %st(1)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fstcw (%esp) / restore old RP
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw (%esp),%dx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis andw $0xfcff,%dx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw 4(%esp),%cx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis andw $0x0300,%cx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis orw %dx,%cx
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movw %cx,(%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fldcw (%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis add $8,%esp
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis ret
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis.shortcut:
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis / Here, |x| < ln(2), so |z| = |x*log2(e)| < 1,
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis / whence z is in f2xm1's domain.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis flds 4(%esp) / push x
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fldl2e / push log2e }not for xtndd_dbl
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fmulp %st,%st(1) / z = x*log2e }not for xtndd_dbl
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis f2xm1 / 2**(x*log2(e))-1 = e**x - 1
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis ret
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis.not_finite:
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis ja .NaN_or_pinf / branch if x is NaN
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis movl 4(%esp),%eax / eax <-- x
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis andl $0x80000000,%eax / here, x is infinite, but +/-?
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis jz .NaN_or_pinf / branch if x = +INF
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fld1 / Here, x = -inf, so return -1
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fchs
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis ret
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis.NaN_or_pinf:
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis / Here, x = NaN or +inf, so load x and return immediately.
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis flds 4(%esp)
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis fwait
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis ret
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis .align 4
25c28e83beb90e7c80452a7c818c5e6f73a07dc8Piotr Jasiukajtis SET_SIZE(expm1f)