[8256] | 1 | ; $Id: sinf.asm 98103 2023-01-17 14:15:46Z vboxsync $
|
---|
| 2 | ;; @file
|
---|
[96242] | 3 | ; IPRT - No-CRT sinf - AMD64 & X86.
|
---|
[8256] | 4 | ;
|
---|
| 5 |
|
---|
| 6 | ;
|
---|
[98103] | 7 | ; Copyright (C) 2006-2023 Oracle and/or its affiliates.
|
---|
[5999] | 8 | ;
|
---|
[96407] | 9 | ; This file is part of VirtualBox base platform packages, as
|
---|
| 10 | ; available from https://www.virtualbox.org.
|
---|
[5999] | 11 | ;
|
---|
[96407] | 12 | ; This program is free software; you can redistribute it and/or
|
---|
| 13 | ; modify it under the terms of the GNU General Public License
|
---|
| 14 | ; as published by the Free Software Foundation, in version 3 of the
|
---|
| 15 | ; License.
|
---|
| 16 | ;
|
---|
| 17 | ; This program is distributed in the hope that it will be useful, but
|
---|
| 18 | ; WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
| 19 | ; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
---|
| 20 | ; General Public License for more details.
|
---|
| 21 | ;
|
---|
| 22 | ; You should have received a copy of the GNU General Public License
|
---|
| 23 | ; along with this program; if not, see <https://www.gnu.org/licenses>.
|
---|
| 24 | ;
|
---|
[5999] | 25 | ; The contents of this file may alternatively be used under the terms
|
---|
| 26 | ; of the Common Development and Distribution License Version 1.0
|
---|
[96407] | 27 | ; (CDDL), a copy of it is provided in the "COPYING.CDDL" file included
|
---|
| 28 | ; in the VirtualBox distribution, in which case the provisions of the
|
---|
[5999] | 29 | ; CDDL are applicable instead of those of the GPL.
|
---|
| 30 | ;
|
---|
| 31 | ; You may elect to license modified versions of this file under the
|
---|
| 32 | ; terms and conditions of either the GPL or the CDDL or both.
|
---|
| 33 | ;
|
---|
[96407] | 34 | ; SPDX-License-Identifier: GPL-3.0-only OR CDDL-1.0
|
---|
| 35 | ;
|
---|
[8256] | 36 |
|
---|
[96240] | 37 |
|
---|
| 38 | %define RT_ASM_WITH_SEH64
|
---|
[8256] | 39 | %include "iprt/asmdefs.mac"
|
---|
[96240] | 40 | %include "iprt/x86.mac"
|
---|
[8256] | 41 |
|
---|
[96240] | 42 |
|
---|
[8256] | 43 | BEGINCODE
|
---|
| 44 |
|
---|
[96240] | 45 |
|
---|
| 46 | ;;
|
---|
| 47 | ; Compute the sine of rd, measured in radians.
|
---|
| 48 | ;
|
---|
| 49 | ; @returns st(0) / xmm0
|
---|
| 50 | ; @param rd [rbp + xCB*2] / xmm0
|
---|
| 51 | ;
|
---|
[96242] | 52 | RT_NOCRT_BEGINPROC sinf
|
---|
[96240] | 53 | push xBP
|
---|
| 54 | SEH64_PUSH_xBP
|
---|
| 55 | mov xBP, xSP
|
---|
| 56 | SEH64_SET_FRAME_xBP 0
|
---|
| 57 | sub xSP, 20h
|
---|
| 58 | SEH64_ALLOCATE_STACK 20h
|
---|
| 59 | SEH64_END_PROLOGUE
|
---|
[96060] | 60 |
|
---|
[96240] | 61 | %ifdef RT_OS_WINDOWS
|
---|
| 62 | ;
|
---|
| 63 | ; Make sure we use full precision and not the windows default of 53 bits.
|
---|
| 64 | ;
|
---|
| 65 | fnstcw [xBP - 20h]
|
---|
| 66 | mov ax, [xBP - 20h]
|
---|
| 67 | or ax, X86_FCW_PC_64 ; includes both bits, so no need to clear the mask.
|
---|
| 68 | mov [xBP - 1ch], ax
|
---|
| 69 | fldcw [xBP - 1ch]
|
---|
| 70 | %endif
|
---|
| 71 |
|
---|
| 72 | ;
|
---|
| 73 | ; Load the input into st0.
|
---|
| 74 | ;
|
---|
[96060] | 75 | %ifdef RT_ARCH_AMD64
|
---|
[96242] | 76 | movss [xBP - 10h], xmm0
|
---|
| 77 | fld dword [xBP - 10h]
|
---|
[96060] | 78 | %else
|
---|
[96242] | 79 | fld dword [xBP + xCB*2]
|
---|
[96060] | 80 | %endif
|
---|
[8256] | 81 |
|
---|
[96240] | 82 | ;
|
---|
| 83 | ; We examin the input and weed out non-finit numbers first.
|
---|
| 84 | ;
|
---|
| 85 | fxam
|
---|
| 86 | fnstsw ax
|
---|
| 87 | and ax, X86_FSW_C3 | X86_FSW_C2 | X86_FSW_C0
|
---|
| 88 | cmp ax, X86_FSW_C2 ; Normal finite number (excluding zero)
|
---|
| 89 | je .finite
|
---|
| 90 | cmp ax, X86_FSW_C3 ; Zero
|
---|
| 91 | je .zero
|
---|
| 92 | cmp ax, X86_FSW_C3 | X86_FSW_C2 ; Denormals - treat them as zero.
|
---|
| 93 | je .zero
|
---|
| 94 | cmp ax, X86_FSW_C0 ; NaN - must handle it special,
|
---|
| 95 | je .nan
|
---|
[8256] | 96 |
|
---|
[96240] | 97 | ; Pass infinities and unsupported inputs to fsin, assuming it does the right thing.
|
---|
| 98 | .do_sin:
|
---|
| 99 | fsin
|
---|
| 100 | jmp .return_val
|
---|
| 101 |
|
---|
| 102 | ;
|
---|
| 103 | ; Finite number.
|
---|
| 104 | ;
|
---|
| 105 | .finite:
|
---|
| 106 | ; For very tiny numbers, 0 < abs(input) < 2**-25, we can return the
|
---|
| 107 | ; input value directly.
|
---|
| 108 | fld st0 ; duplicate st0
|
---|
| 109 | fabs ; make it an absolute (positive) value.
|
---|
| 110 | fld qword [.s_r64Tiny xWrtRIP]
|
---|
| 111 | fcomip st1 ; compare s_r64Tiny and fabs(input)
|
---|
| 112 | ja .return_tiny_number_as_is ; jump if fabs(input) is smaller
|
---|
| 113 |
|
---|
| 114 | ; FSIN is documented to be reasonable for the range ]-3pi/4,3pi/4[, so
|
---|
| 115 | ; while we have fabs(input) loaded already, check for that here and
|
---|
| 116 | ; allow rtNoCrtMathSinCore to assume it won't see values very close to
|
---|
| 117 | ; zero, except by cos -> sin conversion where they won't be relevant to
|
---|
| 118 | ; any assumpttions about precision approximation.
|
---|
| 119 | fld qword [.s_r64FSinOkay xWrtRIP]
|
---|
| 120 | fcomip st1
|
---|
| 121 | ffreep st0 ; drop the fabs(input) value
|
---|
| 122 | ja .do_sin
|
---|
| 123 |
|
---|
| 124 | ;
|
---|
| 125 | ; Call common sine/cos worker.
|
---|
| 126 | ;
|
---|
[96242] | 127 | mov ecx, 0 ; float
|
---|
[96240] | 128 | extern NAME(rtNoCrtMathSinCore)
|
---|
| 129 | call NAME(rtNoCrtMathSinCore)
|
---|
| 130 |
|
---|
| 131 | ;
|
---|
| 132 | ; Run st0.
|
---|
| 133 | ;
|
---|
| 134 | .return_val:
|
---|
[96060] | 135 | %ifdef RT_ARCH_AMD64
|
---|
[96242] | 136 | fstp dword [xBP - 10h]
|
---|
| 137 | movss xmm0, [xBP - 10h]
|
---|
[96060] | 138 | %endif
|
---|
[96240] | 139 | %ifdef RT_OS_WINDOWS
|
---|
| 140 | fldcw [xBP - 20h] ; restore original
|
---|
| 141 | %endif
|
---|
| 142 | .return:
|
---|
| 143 | leave
|
---|
| 144 | ret
|
---|
| 145 |
|
---|
| 146 | ;
|
---|
| 147 | ; As explained already, we can return tiny numbers directly too as the
|
---|
[96242] | 148 | ; output from sinf(input) = input given our precision.
|
---|
[96240] | 149 | ; We can skip the st0 -> xmm0 translation here, so follow the same path
|
---|
| 150 | ; as .zero & .nan, after we've removed the fabs(input) value.
|
---|
| 151 | ;
|
---|
| 152 | .return_tiny_number_as_is:
|
---|
| 153 | ffreep st0
|
---|
| 154 |
|
---|
| 155 | ;
|
---|
[96242] | 156 | ; sinf(+/-0.0) = +/-0.0 (preserve the sign)
|
---|
[96240] | 157 | ; We can skip the st0 -> xmm0 translation here, so follow the .nan code path.
|
---|
| 158 | ;
|
---|
| 159 | .zero:
|
---|
| 160 |
|
---|
| 161 | ;
|
---|
| 162 | ; Input is NaN, output it unmodified as far as we can (FLD changes SNaN
|
---|
| 163 | ; to QNaN when masked).
|
---|
| 164 | ;
|
---|
| 165 | .nan:
|
---|
| 166 | %ifdef RT_ARCH_AMD64
|
---|
| 167 | ffreep st0
|
---|
| 168 | %endif
|
---|
| 169 | jmp .return
|
---|
| 170 |
|
---|
| 171 | ALIGNCODE(8)
|
---|
| 172 | ; Ca. 2**-26, absolute value. Inputs closer to zero than this can be
|
---|
[96242] | 173 | ; returns directly as the sinf(input) value should be basically the same
|
---|
[96240] | 174 | ; given the precision we're working with and FSIN probably won't even
|
---|
| 175 | ; manage that.
|
---|
| 176 | ;; @todo experiment when FSIN gets better than this.
|
---|
| 177 | .s_r64Tiny:
|
---|
| 178 | dq 1.49011612e-8
|
---|
| 179 | ; The absolute limit of FSIN "good" range.
|
---|
| 180 | .s_r64FSinOkay:
|
---|
| 181 | dq 2.356194490192344928845 ; 3pi/4
|
---|
| 182 | ;dq 1.57079632679489661923 ; pi/2 - alternative.
|
---|
| 183 |
|
---|
[96242] | 184 | ENDPROC RT_NOCRT(sinf)
|
---|
[8256] | 185 |
|
---|