|
Packit |
6c4009 |
/* Function sincosf vectorized with AVX-512. Wrapper to AVX2 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 (_ZGVeN16vl4l4_sincosf)
|
|
Packit |
6c4009 |
WRAPPER_IMPL_AVX512_fFF _ZGVdN8vl4l4_sincosf
|
|
Packit |
6c4009 |
END (_ZGVeN16vl4l4_sincosf)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* AVX512 ISA version as wrapper to AVX2 ISA version (for vector
|
|
Packit |
6c4009 |
function declared with #pragma omp declare simd notinbranch). */
|
|
Packit |
6c4009 |
.macro WRAPPER_IMPL_AVX512_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 $-64, %rsp
|
|
Packit |
6c4009 |
subq $448, %rsp
|
|
Packit |
6c4009 |
vmovups %zmm0, 384(%rsp)
|
|
Packit |
6c4009 |
lea (%rsp), %rdi
|
|
Packit |
6c4009 |
vmovups %zmm1, 128(%rdi)
|
|
Packit |
6c4009 |
vmovups %zmm2, 192(%rdi)
|
|
Packit |
6c4009 |
vmovups %zmm3, 256(%rdi)
|
|
Packit |
6c4009 |
vmovups %zmm4, 320(%rdi)
|
|
Packit |
6c4009 |
lea 64(%rsp), %rsi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
vmovdqu 416(%rsp), %ymm0
|
|
Packit |
6c4009 |
lea 32(%rsp), %rdi
|
|
Packit |
6c4009 |
lea 96(%rsp), %rsi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
movq 128(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 136(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 144(%rsp), %r8
|
|
Packit |
6c4009 |
movq 152(%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 160(%rsp), %rax
|
|
Packit |
6c4009 |
movq 168(%rsp), %rcx
|
|
Packit |
6c4009 |
movl %edi, (%r8)
|
|
Packit |
6c4009 |
movl %r9d, (%r10)
|
|
Packit |
6c4009 |
movq 176(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 184(%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 192(%rsp), %r11
|
|
Packit |
6c4009 |
movq 200(%rsp), %rdx
|
|
Packit |
6c4009 |
movl %esi, (%rdi)
|
|
Packit |
6c4009 |
movl %r8d, (%r9)
|
|
Packit |
6c4009 |
movq 208(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 216(%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 224(%rsp), %r10
|
|
Packit |
6c4009 |
movq 232(%rsp), %rax
|
|
Packit |
6c4009 |
movl %ecx, (%rsi)
|
|
Packit |
6c4009 |
movl %edi, (%r8)
|
|
Packit |
6c4009 |
movq 240(%rsp), %rcx
|
|
Packit |
6c4009 |
movq 248(%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 |
movq 256(%rsp), %r9
|
|
Packit |
6c4009 |
movq 264(%rsp), %r11
|
|
Packit |
6c4009 |
movl %edx, (%rcx)
|
|
Packit |
6c4009 |
movl %esi, (%rdi)
|
|
Packit |
6c4009 |
movq 272(%rsp), %rdx
|
|
Packit |
6c4009 |
movq 280(%rsp), %rsi
|
|
Packit |
6c4009 |
movl 64(%rsp), %r8d
|
|
Packit |
6c4009 |
movl 68(%rsp), %r10d
|
|
Packit |
6c4009 |
movl 72(%rsp), %eax
|
|
Packit |
6c4009 |
movl 76(%rsp), %ecx
|
|
Packit |
6c4009 |
movl %r8d, (%r9)
|
|
Packit |
6c4009 |
movl %r10d, (%r11)
|
|
Packit |
6c4009 |
movq 288(%rsp), %r8
|
|
Packit |
6c4009 |
movq 296(%rsp), %r10
|
|
Packit |
6c4009 |
movl %eax, (%rdx)
|
|
Packit |
6c4009 |
movl %ecx, (%rsi)
|
|
Packit |
6c4009 |
movq 304(%rsp), %rax
|
|
Packit |
6c4009 |
movq 312(%rsp), %rcx
|
|
Packit |
6c4009 |
movl 80(%rsp), %edi
|
|
Packit |
6c4009 |
movl 84(%rsp), %r9d
|
|
Packit |
6c4009 |
movl 88(%rsp), %r11d
|
|
Packit |
6c4009 |
movl 92(%rsp), %edx
|
|
Packit |
6c4009 |
movl %edi, (%r8)
|
|
Packit |
6c4009 |
movl %r9d, (%r10)
|
|
Packit |
6c4009 |
movq 320(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 328(%rsp), %r9
|
|
Packit |
6c4009 |
movl %r11d, (%rax)
|
|
Packit |
6c4009 |
movl %edx, (%rcx)
|
|
Packit |
6c4009 |
movq 336(%rsp), %r11
|
|
Packit |
6c4009 |
movq 344(%rsp), %rdx
|
|
Packit |
6c4009 |
movl 96(%rsp), %esi
|
|
Packit |
6c4009 |
movl 100(%rsp), %r8d
|
|
Packit |
6c4009 |
movl 104(%rsp), %r10d
|
|
Packit |
6c4009 |
movl 108(%rsp), %eax
|
|
Packit |
6c4009 |
movl %esi, (%rdi)
|
|
Packit |
6c4009 |
movl %r8d, (%r9)
|
|
Packit |
6c4009 |
movq 352(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 360(%rsp), %r8
|
|
Packit |
6c4009 |
movl %r10d, (%r11)
|
|
Packit |
6c4009 |
movl %eax, (%rdx)
|
|
Packit |
6c4009 |
movq 368(%rsp), %r10
|
|
Packit |
6c4009 |
movq 376(%rsp), %rax
|
|
Packit |
6c4009 |
movl 112(%rsp), %ecx
|
|
Packit |
6c4009 |
movl 116(%rsp), %edi
|
|
Packit |
6c4009 |
movl 120(%rsp), %r9d
|
|
Packit |
6c4009 |
movl 124(%rsp), %r11d
|
|
Packit |
6c4009 |
movl %ecx, (%rsi)
|
|
Packit |
6c4009 |
movl %edi, (%r8)
|
|
Packit |
6c4009 |
movl %r9d, (%r10)
|
|
Packit |
6c4009 |
movl %r11d, (%rax)
|
|
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 $-64, %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 -112(%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 -176(%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 $344, %esp
|
|
Packit |
6c4009 |
vmovdqa64 %zmm1, -240(%ebp)
|
|
Packit |
6c4009 |
vmovdqa64 %zmm2, -304(%ebp)
|
|
Packit |
6c4009 |
vmovaps %zmm0, -368(%ebp)
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
leal 32(%r12), %esi
|
|
Packit |
6c4009 |
vmovups -336(%ebp), %ymm0
|
|
Packit |
6c4009 |
leal 32(%rbx), %edi
|
|
Packit |
6c4009 |
call HIDDEN_JUMPTARGET(\callee)
|
|
Packit |
6c4009 |
movl -240(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -176(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -236(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -172(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -232(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -168(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -228(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -164(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -224(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -160(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -220(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -156(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -216(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -152(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -212(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -148(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -208(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -144(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -204(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -140(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -200(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -136(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -196(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -132(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -192(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -128(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -188(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -124(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -184(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -120(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -180(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -116(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -304(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -112(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -300(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -108(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -296(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -104(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -292(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -100(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -288(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -96(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -284(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -92(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -280(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -88(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -276(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -84(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -272(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -80(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -268(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -76(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -264(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -72(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -260(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -68(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -256(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -64(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -252(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -60(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -248(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -56(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
movl -244(%ebp), %eax
|
|
Packit |
6c4009 |
vmovss -52(%ebp), %xmm0
|
|
Packit |
6c4009 |
vmovss %xmm0, (%eax)
|
|
Packit |
6c4009 |
addl $344, %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 (_ZGVeN16vvv_sincosf)
|
|
Packit |
6c4009 |
WRAPPER_IMPL_AVX512_fFF_vvv _ZGVdN8vl4l4_sincosf
|
|
Packit |
6c4009 |
END (_ZGVeN16vvv_sincosf)
|