|
Packit |
6c4009 |
/* Function sincosf vectorized with SSE4.
|
|
Packit |
6c4009 |
Copyright (C) 2014-2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
This file is part of the GNU C Library.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is free software; you can redistribute it and/or
|
|
Packit |
6c4009 |
modify it under the terms of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License as published by the Free Software Foundation; either
|
|
Packit |
6c4009 |
version 2.1 of the License, or (at your option) any later version.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is distributed in the hope that it will be useful,
|
|
Packit |
6c4009 |
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
Packit |
6c4009 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Packit |
6c4009 |
Lesser General Public License for more details.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
You should have received a copy of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License along with the GNU C Library; if not, see
|
|
Packit |
6c4009 |
<http://www.gnu.org/licenses/>. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#include <sysdep.h>
|
|
Packit |
6c4009 |
#include "svml_s_trig_data.h"
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (_ZGVbN4vl4l4_sincosf_sse4)
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
ALGORITHM DESCRIPTION:
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
1) Range reduction to [-Pi/4; +Pi/4] interval
|
|
Packit |
6c4009 |
a) Grab sign from source argument and save it.
|
|
Packit |
6c4009 |
b) Remove sign using AND operation
|
|
Packit |
6c4009 |
c) Getting octant Y by 2/Pi multiplication
|
|
Packit |
6c4009 |
d) Add "Right Shifter" value
|
|
Packit |
6c4009 |
e) Treat obtained value as integer S for destination sign setting.
|
|
Packit |
6c4009 |
SS = ((S-S&1)&2)<<30; For sin part
|
|
Packit |
6c4009 |
SC = ((S+S&1)&2)<<30; For cos part
|
|
Packit |
6c4009 |
f) Change destination sign if source sign is negative
|
|
Packit |
6c4009 |
using XOR operation.
|
|
Packit |
6c4009 |
g) Subtract "Right Shifter" (0x4B000000) value
|
|
Packit |
6c4009 |
h) Subtract Y*(PI/2) from X argument, where PI/2 divided to 4 parts:
|
|
Packit |
6c4009 |
X = X - Y*PI1 - Y*PI2 - Y*PI3 - Y*PI4;
|
|
Packit |
6c4009 |
2) Polynomial (minimax for sin within [-Pi/4; +Pi/4] interval)
|
|
Packit |
6c4009 |
a) Calculate X^2 = X * X
|
|
Packit |
6c4009 |
b) Calculate 2 polynomials for sin and cos:
|
|
Packit |
6c4009 |
RS = X * ( A0 + X^2 * (A1 + x^2 * (A2 + x^2 * (A3))));
|
|
Packit |
6c4009 |
RC = B0 + X^2 * (B1 + x^2 * (B2 + x^2 * (B3 + x^2 * (B4))));
|
|
Packit |
6c4009 |
c) Swap RS & RC if first bit of obtained value after
|
|
Packit |
6c4009 |
Right Shifting is set to 1. Using And, Andnot & Or operations.
|
|
Packit |
6c4009 |
3) Destination sign setting
|
|
Packit |
6c4009 |
a) Set shifted destination sign using XOR operation:
|
|
Packit |
6c4009 |
R1 = XOR( RS, SS );
|
|
Packit |
6c4009 |
R2 = XOR( RC, SC ). */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pushq %rbp
|
|
Packit |
6c4009 |
cfi_adjust_cfa_offset (8)
|
|
Packit |
6c4009 |
cfi_rel_offset (%rbp, 0)
|
|
Packit |
6c4009 |
movq %rsp, %rbp
|
|
Packit |
6c4009 |
cfi_def_cfa_register (%rbp)
|
|
Packit |
6c4009 |
andq $-64, %rsp
|
|
Packit |
6c4009 |
subq $320, %rsp
|
|
Packit |
6c4009 |
movq __svml_s_trig_data@GOTPCREL(%rip), %rax
|
|
Packit |
6c4009 |
movups %xmm12, 176(%rsp)
|
|
Packit |
6c4009 |
movups %xmm9, 160(%rsp)
|
|
Packit |
6c4009 |
movups __sAbsMask(%rax), %xmm12
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Absolute argument computation */
|
|
Packit |
6c4009 |
movaps %xmm12, %xmm5
|
|
Packit |
6c4009 |
andnps %xmm0, %xmm12
|
|
Packit |
6c4009 |
movups __sInvPI(%rax), %xmm7
|
|
Packit |
6c4009 |
andps %xmm0, %xmm5
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* c) Getting octant Y by 2/Pi multiplication
|
|
Packit |
6c4009 |
d) Add "Right Shifter" value. */
|
|
Packit |
6c4009 |
mulps %xmm5, %xmm7
|
|
Packit |
6c4009 |
movups %xmm10, 144(%rsp)
|
|
Packit |
6c4009 |
movups __sPI1(%rax), %xmm10
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* h) Subtract Y*(PI/2) from X argument, where PI/2 divided to 3 parts:
|
|
Packit |
6c4009 |
X = X - Y*PI1 - Y*PI2 - Y*PI3. */
|
|
Packit |
6c4009 |
movaps %xmm10, %xmm1
|
|
Packit |
6c4009 |
addps __sRShifter(%rax), %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* e) Treat obtained value as integer S for destination sign setting */
|
|
Packit |
6c4009 |
movaps %xmm7, %xmm9
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* g) Subtract "Right Shifter" (0x4B000000) value */
|
|
Packit |
6c4009 |
subps __sRShifter(%rax), %xmm7
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm1
|
|
Packit |
6c4009 |
pslld $31, %xmm9
|
|
Packit |
6c4009 |
movups __sPI2(%rax), %xmm6
|
|
Packit |
6c4009 |
movups %xmm13, 112(%rsp)
|
|
Packit |
6c4009 |
movaps %xmm5, %xmm13
|
|
Packit |
6c4009 |
movaps %xmm6, %xmm2
|
|
Packit |
6c4009 |
subps %xmm1, %xmm13
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm2
|
|
Packit |
6c4009 |
movups __sSignMask(%rax), %xmm3
|
|
Packit |
6c4009 |
movaps %xmm5, %xmm1
|
|
Packit |
6c4009 |
movups __sOneHalf(%rax), %xmm4
|
|
Packit |
6c4009 |
subps %xmm2, %xmm13
|
|
Packit |
6c4009 |
cmpnleps __sRangeReductionVal(%rax), %xmm5
|
|
Packit |
6c4009 |
movaps %xmm3, %xmm2
|
|
Packit |
6c4009 |
andps %xmm13, %xmm2
|
|
Packit |
6c4009 |
xorps %xmm2, %xmm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Result sign calculations */
|
|
Packit |
6c4009 |
xorps %xmm2, %xmm3
|
|
Packit |
6c4009 |
xorps %xmm9, %xmm3
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Add correction term 0.5 for cos() part */
|
|
Packit |
6c4009 |
addps %xmm7, %xmm4
|
|
Packit |
6c4009 |
movmskps %xmm5, %ecx
|
|
Packit |
6c4009 |
mulps %xmm4, %xmm10
|
|
Packit |
6c4009 |
mulps %xmm4, %xmm6
|
|
Packit |
6c4009 |
subps %xmm10, %xmm1
|
|
Packit |
6c4009 |
movups __sPI3(%rax), %xmm10
|
|
Packit |
6c4009 |
subps %xmm6, %xmm1
|
|
Packit |
6c4009 |
movaps %xmm10, %xmm6
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm6
|
|
Packit |
6c4009 |
mulps %xmm4, %xmm10
|
|
Packit |
6c4009 |
subps %xmm6, %xmm13
|
|
Packit |
6c4009 |
subps %xmm10, %xmm1
|
|
Packit |
6c4009 |
movups __sPI4(%rax), %xmm6
|
|
Packit |
6c4009 |
mulps %xmm6, %xmm7
|
|
Packit |
6c4009 |
mulps %xmm6, %xmm4
|
|
Packit |
6c4009 |
subps %xmm7, %xmm13
|
|
Packit |
6c4009 |
subps %xmm4, %xmm1
|
|
Packit |
6c4009 |
xorps %xmm9, %xmm13
|
|
Packit |
6c4009 |
xorps %xmm3, %xmm1
|
|
Packit |
6c4009 |
movaps %xmm13, %xmm4
|
|
Packit |
6c4009 |
movaps %xmm1, %xmm2
|
|
Packit |
6c4009 |
mulps %xmm13, %xmm4
|
|
Packit |
6c4009 |
mulps %xmm1, %xmm2
|
|
Packit |
6c4009 |
movups __sA9(%rax), %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* 2) Polynomial (minimax for sin within [-Pi/4; +Pi/4] interval)
|
|
Packit |
6c4009 |
a) Calculate X^2 = X * X
|
|
Packit |
6c4009 |
b) Calculate 2 polynomials for sin and cos:
|
|
Packit |
6c4009 |
RS = X * ( A0 + X^2 * (A1 + x^2 * (A2 + x^2 * (A3))));
|
|
Packit |
6c4009 |
RC = B0 + X^2 * (B1 + x^2 * (B2 + x^2 * (B3 + x^2 * (B4)))) */
|
|
Packit |
6c4009 |
movaps %xmm7, %xmm3
|
|
Packit |
6c4009 |
mulps %xmm4, %xmm3
|
|
Packit |
6c4009 |
mulps %xmm2, %xmm7
|
|
Packit |
6c4009 |
addps __sA7(%rax), %xmm3
|
|
Packit |
6c4009 |
addps __sA7(%rax), %xmm7
|
|
Packit |
6c4009 |
mulps %xmm4, %xmm3
|
|
Packit |
6c4009 |
mulps %xmm2, %xmm7
|
|
Packit |
6c4009 |
addps __sA5(%rax), %xmm3
|
|
Packit |
6c4009 |
addps __sA5(%rax), %xmm7
|
|
Packit |
6c4009 |
mulps %xmm4, %xmm3
|
|
Packit |
6c4009 |
mulps %xmm2, %xmm7
|
|
Packit |
6c4009 |
addps __sA3(%rax), %xmm3
|
|
Packit |
6c4009 |
addps __sA3(%rax), %xmm7
|
|
Packit |
6c4009 |
mulps %xmm3, %xmm4
|
|
Packit |
6c4009 |
mulps %xmm7, %xmm2
|
|
Packit |
6c4009 |
mulps %xmm13, %xmm4
|
|
Packit |
6c4009 |
mulps %xmm1, %xmm2
|
|
Packit |
6c4009 |
addps %xmm4, %xmm13
|
|
Packit |
6c4009 |
addps %xmm2, %xmm1
|
|
Packit |
6c4009 |
xorps %xmm12, %xmm13
|
|
Packit |
6c4009 |
testl %ecx, %ecx
|
|
Packit |
6c4009 |
jne .LBL_1_3
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_2:
|
|
Packit |
6c4009 |
cfi_remember_state
|
|
Packit |
6c4009 |
movups 160(%rsp), %xmm9
|
|
Packit |
6c4009 |
movaps %xmm13, (%rdi)
|
|
Packit |
6c4009 |
movups 144(%rsp), %xmm10
|
|
Packit |
6c4009 |
movups 176(%rsp), %xmm12
|
|
Packit |
6c4009 |
movups 112(%rsp), %xmm13
|
|
Packit |
6c4009 |
movups %xmm1, (%rsi)
|
|
Packit |
6c4009 |
movq %rbp, %rsp
|
|
Packit |
6c4009 |
cfi_def_cfa_register (%rsp)
|
|
Packit |
6c4009 |
popq %rbp
|
|
Packit |
6c4009 |
cfi_adjust_cfa_offset (-8)
|
|
Packit |
6c4009 |
cfi_restore (%rbp)
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_3:
|
|
Packit |
6c4009 |
cfi_restore_state
|
|
Packit |
6c4009 |
movups %xmm0, 128(%rsp)
|
|
Packit |
6c4009 |
movups %xmm13, 192(%rsp)
|
|
Packit |
6c4009 |
movups %xmm1, 256(%rsp)
|
|
Packit |
6c4009 |
je .LBL_1_2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
xorb %dl, %dl
|
|
Packit |
6c4009 |
xorl %eax, %eax
|
|
Packit |
6c4009 |
movups %xmm8, 48(%rsp)
|
|
Packit |
6c4009 |
movups %xmm11, 32(%rsp)
|
|
Packit |
6c4009 |
movups %xmm14, 16(%rsp)
|
|
Packit |
6c4009 |
movups %xmm15, (%rsp)
|
|
Packit |
6c4009 |
movq %rsi, 64(%rsp)
|
|
Packit |
6c4009 |
movq %r12, 104(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (12, 104)
|
|
Packit |
6c4009 |
movb %dl, %r12b
|
|
Packit |
6c4009 |
movq %r13, 96(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (13, 96)
|
|
Packit |
6c4009 |
movl %eax, %r13d
|
|
Packit |
6c4009 |
movq %r14, 88(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (14, 88)
|
|
Packit |
6c4009 |
movl %ecx, %r14d
|
|
Packit |
6c4009 |
movq %r15, 80(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (15, 80)
|
|
Packit |
6c4009 |
movq %rbx, 72(%rsp)
|
|
Packit |
6c4009 |
movq %rdi, %rbx
|
|
Packit |
6c4009 |
cfi_remember_state
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_6:
|
|
Packit |
6c4009 |
btl %r13d, %r14d
|
|
Packit |
6c4009 |
jc .LBL_1_13
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_7:
|
|
Packit |
6c4009 |
lea 1(%r13), %esi
|
|
Packit |
6c4009 |
btl %esi, %r14d
|
|
Packit |
6c4009 |
jc .LBL_1_10
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_8:
|
|
Packit |
6c4009 |
incb %r12b
|
|
Packit |
6c4009 |
addl $2, %r13d
|
|
Packit |
6c4009 |
cmpb $16, %r12b
|
|
Packit |
6c4009 |
jb .LBL_1_6
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movups 48(%rsp), %xmm8
|
|
Packit |
6c4009 |
movq %rbx, %rdi
|
|
Packit |
6c4009 |
movups 32(%rsp), %xmm11
|
|
Packit |
6c4009 |
movups 16(%rsp), %xmm14
|
|
Packit |
6c4009 |
movups (%rsp), %xmm15
|
|
Packit |
6c4009 |
movq 64(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 104(%rsp), %r12
|
|
Packit |
6c4009 |
cfi_restore (%r12)
|
|
Packit |
6c4009 |
movq 96(%rsp), %r13
|
|
Packit |
6c4009 |
cfi_restore (%r13)
|
|
Packit |
6c4009 |
movq 88(%rsp), %r14
|
|
Packit |
6c4009 |
cfi_restore (%r14)
|
|
Packit |
6c4009 |
movq 80(%rsp), %r15
|
|
Packit |
6c4009 |
cfi_restore (%r15)
|
|
Packit |
6c4009 |
movq 72(%rsp), %rbx
|
|
Packit |
6c4009 |
movups 192(%rsp), %xmm13
|
|
Packit |
6c4009 |
movups 256(%rsp), %xmm1
|
|
Packit |
6c4009 |
jmp .LBL_1_2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_10:
|
|
Packit |
6c4009 |
cfi_restore_state
|
|
Packit |
6c4009 |
movzbl %r12b, %r15d
|
|
Packit |
6c4009 |
movss 132(%rsp,%r15,8), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(sinf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movss %xmm0, 196(%rsp,%r15,8)
|
|
Packit |
6c4009 |
movss 132(%rsp,%r15,8), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(cosf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movss %xmm0, 260(%rsp,%r15,8)
|
|
Packit |
6c4009 |
jmp .LBL_1_8
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_13:
|
|
Packit |
6c4009 |
movzbl %r12b, %r15d
|
|
Packit |
6c4009 |
movss 128(%rsp,%r15,8), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(sinf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movss %xmm0, 192(%rsp,%r15,8)
|
|
Packit |
6c4009 |
movss 128(%rsp,%r15,8), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(cosf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movss %xmm0, 256(%rsp,%r15,8)
|
|
Packit |
6c4009 |
jmp .LBL_1_7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
END (_ZGVbN4vl4l4_sincosf_sse4)
|
|
Packit |
6c4009 |
libmvec_hidden_def(_ZGVbN4vl4l4_sincosf_sse4)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* vvv version implemented with wrapper to vl4l4 variant. */
|
|
Packit |
6c4009 |
ENTRY (_ZGVbN4vvv_sincosf_sse4)
|
|
Packit |
6c4009 |
#ifndef __ILP32__
|
|
Packit |
6c4009 |
subq $104, %rsp
|
|
Packit |
6c4009 |
.cfi_def_cfa_offset 112
|
|
Packit |
6c4009 |
movdqu %xmm1, 32(%rsp)
|
|
Packit |
6c4009 |
lea (%rsp), %rdi
|
|
Packit |
6c4009 |
movdqu %xmm2, 48(%rdi)
|
|
Packit |
6c4009 |
lea 16(%rsp), %rsi
|
|
Packit |
6c4009 |
movdqu %xmm3, 48(%rsi)
|
|
Packit |
6c4009 |
movdqu %xmm4, 64(%rsi)
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(_ZGVbN4vl4l4_sincosf_sse4)
|
|
Packit |
6c4009 |
movq 32(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 40(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 48(%rsp), %r8
|
|
Packit |
6c4009 |
movq 56(%rsp), %r10
|
|
Packit |
6c4009 |
movl (%rsp), %eax
|
|
Packit |
6c4009 |
movl 4(%rsp), %ecx
|
|
Packit |
6c4009 |
movl 8(%rsp), %edi
|
|
Packit |
6c4009 |
movl 12(%rsp), %r9d
|
|
Packit |
6c4009 |
movl %eax, (%rdx)
|
|
Packit |
6c4009 |
movl %ecx, (%rsi)
|
|
Packit |
6c4009 |
movq 64(%rsp), %rax
|
|
Packit |
6c4009 |
movq 72(%rsp), %rcx
|
|
Packit |
6c4009 |
movl %edi, (%r8)
|
|
Packit |
6c4009 |
movl %r9d, (%r10)
|
|
Packit |
6c4009 |
movq 80(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 88(%rsp), %r9
|
|
Packit |
6c4009 |
movl 16(%rsp), %r11d
|
|
Packit |
6c4009 |
movl 20(%rsp), %edx
|
|
Packit |
6c4009 |
movl 24(%rsp), %esi
|
|
Packit |
6c4009 |
movl 28(%rsp), %r8d
|
|
Packit |
6c4009 |
movl %r11d, (%rax)
|
|
Packit |
6c4009 |
movl %edx, (%rcx)
|
|
Packit |
6c4009 |
movl %esi, (%rdi)
|
|
Packit |
6c4009 |
movl %r8d, (%r9)
|
|
Packit |
6c4009 |
addq $104, %rsp
|
|
Packit |
6c4009 |
.cfi_def_cfa_offset 8
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
subl $72, %esp
|
|
Packit |
6c4009 |
.cfi_def_cfa_offset 80
|
|
Packit |
6c4009 |
leal 48(%rsp), %esi
|
|
Packit |
6c4009 |
movaps %xmm1, 16(%esp)
|
|
Packit |
6c4009 |
leal 32(%rsp), %edi
|
|
Packit |
6c4009 |
movaps %xmm2, (%esp)
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(_ZGVbN4vl4l4_sincosf_sse4)
|
|
Packit |
6c4009 |
movl 16(%esp), %eax
|
|
Packit |
6c4009 |
movss 32(%esp), %xmm0
|
|
Packit |
6c4009 |
movss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl 20(%esp), %eax
|
|
Packit |
6c4009 |
movss 36(%esp), %xmm0
|
|
Packit |
6c4009 |
movss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl 24(%esp), %eax
|
|
Packit |
6c4009 |
movss 40(%esp), %xmm0
|
|
Packit |
6c4009 |
movss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl 28(%esp), %eax
|
|
Packit |
6c4009 |
movss 44(%esp), %xmm0
|
|
Packit |
6c4009 |
movss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl (%esp), %eax
|
|
Packit |
6c4009 |
movss 48(%esp), %xmm0
|
|
Packit |
6c4009 |
movss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl 4(%esp), %eax
|
|
Packit |
6c4009 |
movss 52(%esp), %xmm0
|
|
Packit |
6c4009 |
movss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl 8(%esp), %eax
|
|
Packit |
6c4009 |
movss 56(%esp), %xmm0
|
|
Packit |
6c4009 |
movss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl 12(%esp), %eax
|
|
Packit |
6c4009 |
movss 60(%esp), %xmm0
|
|
Packit |
6c4009 |
movss %xmm0, (%eax)
|
|
Packit |
6c4009 |
addl $72, %esp
|
|
Packit |
6c4009 |
.cfi_def_cfa_offset 8
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
END (_ZGVbN4vvv_sincosf_sse4)
|