|
Packit |
6c4009 |
/* Function sincosf vectorized with AVX2, wrapper 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_s_wrapper_impl.h"
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (_ZGVdN8vl4l4_sincosf)
|
|
Packit |
6c4009 |
WRAPPER_IMPL_AVX_fFF _ZGVbN4vl4l4_sincosf
|
|
Packit |
6c4009 |
END (_ZGVdN8vl4l4_sincosf)
|
|
Packit |
6c4009 |
libmvec_hidden_def (_ZGVdN8vl4l4_sincosf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* AVX2 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_AVX2_fFF_vvv callee
|
|
Packit |
6c4009 |
#ifndef __ILP32__
|
|
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 $-32, %rsp
|
|
Packit |
6c4009 |
subq $224, %rsp
|
|
Packit |
6c4009 |
vmovups %ymm0, 192(%rsp)
|
|
Packit |
6c4009 |
lea (%rsp), %rdi
|
|
Packit |
6c4009 |
vmovdqu %ymm1, 64(%rdi)
|
|
Packit |
6c4009 |
vmovdqu %ymm2, 96(%rdi)
|
|
Packit |
6c4009 |
vmovdqu %ymm3, 128(%rdi)
|
|
Packit |
6c4009 |
vmovdqu %ymm4, 160(%rdi)
|
|
Packit |
6c4009 |
lea 32(%rsp), %rsi
|
|
Packit |
6c4009 |
vzeroupper
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
vmovups 208(%rsp), %xmm0
|
|
Packit |
6c4009 |
lea 16(%rsp), %rdi
|
|
Packit |
6c4009 |
lea 48(%rsp), %rsi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
movq 64(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 72(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 80(%rsp), %r8
|
|
Packit |
6c4009 |
movq 88(%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 96(%rsp), %rax
|
|
Packit |
6c4009 |
movq 104(%rsp), %rcx
|
|
Packit |
6c4009 |
movl %edi, (%r8)
|
|
Packit |
6c4009 |
movl %r9d, (%r10)
|
|
Packit |
6c4009 |
movq 112(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 120(%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 |
movq 128(%rsp), %r11
|
|
Packit |
6c4009 |
movq 136(%rsp), %rdx
|
|
Packit |
6c4009 |
movl %esi, (%rdi)
|
|
Packit |
6c4009 |
movl %r8d, (%r9)
|
|
Packit |
6c4009 |
movq 144(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 152(%rsp), %r8
|
|
Packit |
6c4009 |
movl 32(%rsp), %r10d
|
|
Packit |
6c4009 |
movl 36(%rsp), %eax
|
|
Packit |
6c4009 |
movl 40(%rsp), %ecx
|
|
Packit |
6c4009 |
movl 44(%rsp), %edi
|
|
Packit |
6c4009 |
movl %r10d, (%r11)
|
|
Packit |
6c4009 |
movl %eax, (%rdx)
|
|
Packit |
6c4009 |
movq 160(%rsp), %r10
|
|
Packit |
6c4009 |
movq 168(%rsp), %rax
|
|
Packit |
6c4009 |
movl %ecx, (%rsi)
|
|
Packit |
6c4009 |
movl %edi, (%r8)
|
|
Packit |
6c4009 |
movq 176(%rsp), %rcx
|
|
Packit |
6c4009 |
movq 184(%rsp), %rdi
|
|
Packit |
6c4009 |
movl 48(%rsp), %r9d
|
|
Packit |
6c4009 |
movl 52(%rsp), %r11d
|
|
Packit |
6c4009 |
movl 56(%rsp), %edx
|
|
Packit |
6c4009 |
movl 60(%rsp), %esi
|
|
Packit |
6c4009 |
movl %r9d, (%r10)
|
|
Packit |
6c4009 |
movl %r11d, (%rax)
|
|
Packit |
6c4009 |
movl %edx, (%rcx)
|
|
Packit |
6c4009 |
movl %esi, (%rdi)
|
|
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 |
#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 $184, %esp
|
|
Packit |
6c4009 |
vmovdqa %ymm1, -144(%ebp)
|
|
Packit |
6c4009 |
vmovdqa %ymm2, -176(%ebp)
|
|
Packit |
6c4009 |
vmovaps %ymm0, -208(%ebp)
|
|
Packit |
6c4009 |
vzeroupper
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
leal 16(%r12), %esi
|
|
Packit |
6c4009 |
vmovups -192(%ebp), %xmm0
|
|
Packit |
6c4009 |
leal 16(%rbx), %edi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
movl -144(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -112(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -140(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -108(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -136(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -104(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -132(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -100(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -128(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -96(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -124(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -92(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -120(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -88(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -116(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -84(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -176(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -80(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -172(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -76(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -168(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -72(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -164(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -68(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -160(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -64(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -156(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -60(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -152(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -56(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -148(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -52(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
addl $184, %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 (_ZGVdN8vvv_sincosf)
|
|
Packit |
6c4009 |
WRAPPER_IMPL_AVX2_fFF_vvv _ZGVbN4vl4l4_sincosf
|
|
Packit |
6c4009 |
END (_ZGVdN8vvv_sincosf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#ifndef USE_MULTIARCH
|
|
Packit |
6c4009 |
libmvec_hidden_def (_ZGVdN8vvv_sincosf)
|
|
Packit |
6c4009 |
#endif
|