|
Packit |
6c4009 |
/* Function sincos vectorized in AVX ISA as wrapper to SSE4 ISA version.
|
|
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_d_wrapper_impl.h"
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (_ZGVcN4vl8l8_sincos)
|
|
Packit |
6c4009 |
WRAPPER_IMPL_AVX_fFF _ZGVbN2vl8l8_sincos
|
|
Packit |
6c4009 |
END (_ZGVcN4vl8l8_sincos)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* AVX ISA version as wrapper to SSE ISA version (for vector
|
|
Packit |
6c4009 |
function declared with #pragma omp declare simd notinbranch). */
|
|
Packit |
6c4009 |
.macro WRAPPER_IMPL_AVX_fFF_vvv callee
|
|
Packit |
6c4009 |
#ifndef __ILP32__
|
|
Packit |
6c4009 |
pushq %rbp
|
|
Packit |
6c4009 |
movq %rsp, %rbp
|
|
Packit |
6c4009 |
andq $-32, %rsp
|
|
Packit |
6c4009 |
subq $160, %rsp
|
|
Packit |
6c4009 |
vmovupd %ymm0, 64(%rsp)
|
|
Packit |
6c4009 |
lea (%rsp), %rdi
|
|
Packit |
6c4009 |
vmovdqu %xmm1, 96(%rdi)
|
|
Packit |
6c4009 |
vmovdqu %xmm2, 112(%rdi)
|
|
Packit |
6c4009 |
vmovdqu %xmm3, 128(%rdi)
|
|
Packit |
6c4009 |
vmovdqu %xmm4, 144(%rdi)
|
|
Packit |
6c4009 |
lea 32(%rsp), %rsi
|
|
Packit |
6c4009 |
vzeroupper
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
vmovdqu 80(%rsp), %xmm0
|
|
Packit |
6c4009 |
lea 16(%rsp), %rdi
|
|
Packit |
6c4009 |
lea 48(%rsp), %rsi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
movq 96(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 104(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 112(%rsp), %r8
|
|
Packit |
6c4009 |
movq 120(%rsp), %r10
|
|
Packit |
6c4009 |
movq (%rsp), %rax
|
|
Packit |
6c4009 |
movq 8(%rsp), %rcx
|
|
Packit |
6c4009 |
movq 16(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 24(%rsp), %r9
|
|
Packit |
6c4009 |
movq %rax, (%rdx)
|
|
Packit |
6c4009 |
movq %rcx, (%rsi)
|
|
Packit |
6c4009 |
movq 128(%rsp), %rax
|
|
Packit |
6c4009 |
movq 136(%rsp), %rcx
|
|
Packit |
6c4009 |
movq %rdi, (%r8)
|
|
Packit |
6c4009 |
movq %r9, (%r10)
|
|
Packit |
6c4009 |
movq 144(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 152(%rsp), %r9
|
|
Packit |
6c4009 |
movq 32(%rsp), %r11
|
|
Packit |
6c4009 |
movq 40(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 48(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 56(%rsp), %r8
|
|
Packit |
6c4009 |
movq %r11, (%rax)
|
|
Packit |
6c4009 |
movq %rdx, (%rcx)
|
|
Packit |
6c4009 |
movq %rsi, (%rdi)
|
|
Packit |
6c4009 |
movq %r8, (%r9)
|
|
Packit |
6c4009 |
movq %rbp, %rsp
|
|
Packit |
6c4009 |
popq %rbp
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
#else
|
|
Packit |
6c4009 |
leal 8(%rsp), %r10d
|
|
Packit |
6c4009 |
.cfi_def_cfa 10, 0
|
|
Packit |
6c4009 |
andl $-32, %esp
|
|
Packit |
6c4009 |
pushq -8(%r10d)
|
|
Packit |
6c4009 |
pushq %rbp
|
|
Packit |
6c4009 |
.cfi_escape 0x10,0x6,0x2,0x76,0
|
|
Packit |
6c4009 |
movl %esp, %ebp
|
|
Packit |
6c4009 |
pushq %r12
|
|
Packit |
6c4009 |
leal -80(%rbp), %esi
|
|
Packit |
6c4009 |
pushq %r10
|
|
Packit |
6c4009 |
.cfi_escape 0xf,0x3,0x76,0x70,0x6
|
|
Packit |
6c4009 |
.cfi_escape 0x10,0xc,0x2,0x76,0x78
|
|
Packit |
6c4009 |
leal -112(%rbp), %edi
|
|
Packit |
6c4009 |
movq %rsi, %r12
|
|
Packit |
6c4009 |
pushq %rbx
|
|
Packit |
6c4009 |
.cfi_escape 0x10,0x3,0x2,0x76,0x68
|
|
Packit |
6c4009 |
movq %rdi, %rbx
|
|
Packit |
6c4009 |
subl $152, %esp
|
|
Packit |
6c4009 |
vmovaps %xmm1, -128(%ebp)
|
|
Packit |
6c4009 |
vmovaps %xmm2, -144(%ebp)
|
|
Packit |
6c4009 |
vmovapd %ymm0, -176(%ebp)
|
|
Packit |
6c4009 |
vzeroupper
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
leal 16(%r12), %esi
|
|
Packit |
6c4009 |
vmovupd -160(%ebp), %xmm0
|
|
Packit |
6c4009 |
leal 16(%rbx), %edi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
movq -128(%ebp), %rax
|
|
Packit |
6c4009 |
vmovsd -112(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovdqa -128(%ebp), %xmm5
|
|
Packit |
6c4009 |
vmovdqa -144(%ebp), %xmm1
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
vmovsd -104(%ebp), %xmm0
|
|
Packit |
6c4009 |
vpextrd $1, %xmm5, %eax
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movq -120(%ebp), %rax
|
|
Packit |
6c4009 |
vmovsd -96(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
vmovsd -88(%ebp), %xmm0
|
|
Packit |
6c4009 |
vpextrd $3, %xmm5, %eax
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movq -144(%ebp), %rax
|
|
Packit |
6c4009 |
vmovsd -80(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
vmovsd -72(%ebp), %xmm0
|
|
Packit |
6c4009 |
vpextrd $1, %xmm1, %eax
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
movq -136(%ebp), %rax
|
|
Packit |
6c4009 |
vmovsd -64(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
vmovsd -56(%ebp), %xmm0
|
|
Packit |
6c4009 |
vpextrd $3, %xmm1, %eax
|
|
Packit |
6c4009 |
vmovsd %xmm0, (%eax)
|
|
Packit |
6c4009 |
addl $152, %esp
|
|
Packit |
6c4009 |
popq %rbx
|
|
Packit |
6c4009 |
popq %r10
|
|
Packit |
6c4009 |
.cfi_def_cfa 10, 0
|
|
Packit |
6c4009 |
popq %r12
|
|
Packit |
6c4009 |
popq %rbp
|
|
Packit |
6c4009 |
leal -8(%r10), %esp
|
|
Packit |
6c4009 |
.cfi_def_cfa 7, 8
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
#endif
|
|
Packit |
6c4009 |
.endm
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
ENTRY (_ZGVcN4vvv_sincos)
|
|
Packit |
6c4009 |
WRAPPER_IMPL_AVX_fFF_vvv _ZGVbN2vl8l8_sincos
|
|
Packit |
6c4009 |
END (_ZGVcN4vvv_sincos)
|