|
Packit |
6c4009 |
/* Function exp 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_d_exp_data.h"
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (_ZGVbN2v_exp_sse4)
|
|
Packit |
6c4009 |
/*
|
|
Packit |
6c4009 |
ALGORITHM DESCRIPTION:
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
Argument representation:
|
|
Packit |
6c4009 |
N = rint(X*2^k/ln2) = 2^k*M+j
|
|
Packit |
6c4009 |
X = N*ln2/2^k + r = M*ln2 + ln2*(j/2^k) + r
|
|
Packit |
6c4009 |
then -ln2/2^(k+1) < r < ln2/2^(k+1)
|
|
Packit |
6c4009 |
Alternatively:
|
|
Packit |
6c4009 |
N = trunc(X*2^k/ln2)
|
|
Packit |
6c4009 |
then 0 < r < ln2/2^k
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
Result calculation:
|
|
Packit |
6c4009 |
exp(X) = exp(M*ln2 + ln2*(j/2^k) + r)
|
|
Packit |
6c4009 |
= 2^M * 2^(j/2^k) * exp(r)
|
|
Packit |
6c4009 |
2^M is calculated by bit manipulation
|
|
Packit |
6c4009 |
2^(j/2^k) is stored in table
|
|
Packit |
6c4009 |
exp(r) is approximated by polynomial.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The table lookup is skipped if k = 0. */
|
|
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 |
movaps %xmm0, %xmm3
|
|
Packit |
6c4009 |
movq __svml_dexp_data@GOTPCREL(%rip), %r8
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* iAbsX = (int)(lX>>32), lX = *(longlong*)&X */
|
|
Packit |
6c4009 |
pshufd $221, %xmm3, %xmm7
|
|
Packit |
6c4009 |
movups __dbInvLn2(%r8), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* dK = X*dbInvLn2 */
|
|
Packit |
6c4009 |
mulpd %xmm3, %xmm0
|
|
Packit |
6c4009 |
movq __iAbsMask(%r8), %xmm5
|
|
Packit |
6c4009 |
movq __iDomainRange(%r8), %xmm6
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* iAbsX = iAbsX&iAbsMask */
|
|
Packit |
6c4009 |
pand %xmm5, %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* iRangeMask = (iAbsX>iDomainRange) */
|
|
Packit |
6c4009 |
pcmpgtd %xmm6, %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Mask = iRangeMask?1:0, set mask for overflow/underflow */
|
|
Packit |
6c4009 |
movmskps %xmm7, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* dN = rint(X*2^k/Ln2) */
|
|
Packit |
6c4009 |
xorps %xmm7, %xmm7
|
|
Packit |
6c4009 |
movups __dbLn2hi(%r8), %xmm5
|
|
Packit |
6c4009 |
movups __dbLn2lo(%r8), %xmm6
|
|
Packit |
6c4009 |
roundpd $0, %xmm0, %xmm7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* dR = X - dN*dbLn2hi, dbLn2hi is 52-8-k hi bits of ln2/2^k */
|
|
Packit |
6c4009 |
mulpd %xmm7, %xmm5
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* dR = dR - dN*dbLn2lo, dbLn2lo is 40..94 bits of lo part of ln2/2^k */
|
|
Packit |
6c4009 |
mulpd %xmm6, %xmm7
|
|
Packit |
6c4009 |
movups __dbShifter(%r8), %xmm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* dM = X*dbInvLn2+dbShifter */
|
|
Packit |
6c4009 |
addpd %xmm0, %xmm4
|
|
Packit |
6c4009 |
movaps %xmm3, %xmm0
|
|
Packit |
6c4009 |
subpd %xmm5, %xmm0
|
|
Packit |
6c4009 |
subpd %xmm7, %xmm0
|
|
Packit |
6c4009 |
movups __dPC2(%r8), %xmm5
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* exp(r) = b0+r*(b0+r*(b1+r*b2)) */
|
|
Packit |
6c4009 |
mulpd %xmm0, %xmm5
|
|
Packit |
6c4009 |
addpd __dPC1(%r8), %xmm5
|
|
Packit |
6c4009 |
mulpd %xmm0, %xmm5
|
|
Packit |
6c4009 |
movups __dPC0(%r8), %xmm6
|
|
Packit |
6c4009 |
addpd %xmm6, %xmm5
|
|
Packit |
6c4009 |
mulpd %xmm5, %xmm0
|
|
Packit |
6c4009 |
movdqu __lIndexMask(%r8), %xmm2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* lIndex = (*(longlong*)&dM)&lIndexMask, lIndex is the lower K bits of lM */
|
|
Packit |
6c4009 |
movdqa %xmm2, %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* lM = (*(longlong*)&dM)&(~lIndexMask) */
|
|
Packit |
6c4009 |
pandn %xmm4, %xmm2
|
|
Packit |
6c4009 |
pand %xmm4, %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* lM = lM<<(52-K), 2^M */
|
|
Packit |
6c4009 |
psllq $42, %xmm2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* table lookup for dT[j] = 2^(j/2^k) */
|
|
Packit |
6c4009 |
movd %xmm1, %edx
|
|
Packit |
6c4009 |
pextrw $4, %xmm1, %ecx
|
|
Packit |
6c4009 |
addpd %xmm0, %xmm6
|
|
Packit |
6c4009 |
shll $3, %edx
|
|
Packit |
6c4009 |
shll $3, %ecx
|
|
Packit |
6c4009 |
movq (%r8,%rdx), %xmm0
|
|
Packit |
6c4009 |
andl $3, %eax
|
|
Packit |
6c4009 |
movhpd (%r8,%rcx), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* 2^(j/2^k) * exp(r) */
|
|
Packit |
6c4009 |
mulpd %xmm6, %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* multiply by 2^M through integer add */
|
|
Packit |
6c4009 |
paddq %xmm2, %xmm0
|
|
Packit |
6c4009 |
jne .LBL_1_3
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_2:
|
|
Packit |
6c4009 |
cfi_remember_state
|
|
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 %xmm3, 192(%rsp)
|
|
Packit |
6c4009 |
movups %xmm0, 256(%rsp)
|
|
Packit |
6c4009 |
je .LBL_1_2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
xorb %cl, %cl
|
|
Packit |
6c4009 |
xorl %edx, %edx
|
|
Packit |
6c4009 |
movups %xmm8, 112(%rsp)
|
|
Packit |
6c4009 |
movups %xmm9, 96(%rsp)
|
|
Packit |
6c4009 |
movups %xmm10, 80(%rsp)
|
|
Packit |
6c4009 |
movups %xmm11, 64(%rsp)
|
|
Packit |
6c4009 |
movups %xmm12, 48(%rsp)
|
|
Packit |
6c4009 |
movups %xmm13, 32(%rsp)
|
|
Packit |
6c4009 |
movups %xmm14, 16(%rsp)
|
|
Packit |
6c4009 |
movups %xmm15, (%rsp)
|
|
Packit |
6c4009 |
movq %rsi, 136(%rsp)
|
|
Packit |
6c4009 |
movq %rdi, 128(%rsp)
|
|
Packit |
6c4009 |
movq %r12, 168(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (12, 168)
|
|
Packit |
6c4009 |
movb %cl, %r12b
|
|
Packit |
6c4009 |
movq %r13, 160(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (13, 160)
|
|
Packit |
6c4009 |
movl %eax, %r13d
|
|
Packit |
6c4009 |
movq %r14, 152(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (14, 152)
|
|
Packit |
6c4009 |
movl %edx, %r14d
|
|
Packit |
6c4009 |
movq %r15, 144(%rsp)
|
|
Packit |
6c4009 |
cfi_offset_rel_rsp (15, 144)
|
|
Packit |
6c4009 |
cfi_remember_state
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_6:
|
|
Packit |
6c4009 |
btl %r14d, %r13d
|
|
Packit |
6c4009 |
jc .LBL_1_12
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_7:
|
|
Packit |
6c4009 |
lea 1(%r14), %esi
|
|
Packit |
6c4009 |
btl %esi, %r13d
|
|
Packit |
6c4009 |
jc .LBL_1_10
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_8:
|
|
Packit |
6c4009 |
incb %r12b
|
|
Packit |
6c4009 |
addl $2, %r14d
|
|
Packit |
6c4009 |
cmpb $16, %r12b
|
|
Packit |
6c4009 |
jb .LBL_1_6
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movups 112(%rsp), %xmm8
|
|
Packit |
6c4009 |
movups 96(%rsp), %xmm9
|
|
Packit |
6c4009 |
movups 80(%rsp), %xmm10
|
|
Packit |
6c4009 |
movups 64(%rsp), %xmm11
|
|
Packit |
6c4009 |
movups 48(%rsp), %xmm12
|
|
Packit |
6c4009 |
movups 32(%rsp), %xmm13
|
|
Packit |
6c4009 |
movups 16(%rsp), %xmm14
|
|
Packit |
6c4009 |
movups (%rsp), %xmm15
|
|
Packit |
6c4009 |
movq 136(%rsp), %rsi
|
|
Packit |
6c4009 |
movq 128(%rsp), %rdi
|
|
Packit |
6c4009 |
movq 168(%rsp), %r12
|
|
Packit |
6c4009 |
cfi_restore (%r12)
|
|
Packit |
6c4009 |
movq 160(%rsp), %r13
|
|
Packit |
6c4009 |
cfi_restore (%r13)
|
|
Packit |
6c4009 |
movq 152(%rsp), %r14
|
|
Packit |
6c4009 |
cfi_restore (%r14)
|
|
Packit |
6c4009 |
movq 144(%rsp), %r15
|
|
Packit |
6c4009 |
cfi_restore (%r15)
|
|
Packit |
6c4009 |
movups 256(%rsp), %xmm0
|
|
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 |
shlq $4, %r15
|
|
Packit |
6c4009 |
movsd 200(%rsp,%r15), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(__exp_finite)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movsd %xmm0, 264(%rsp,%r15)
|
|
Packit |
6c4009 |
jmp .LBL_1_8
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.LBL_1_12:
|
|
Packit |
6c4009 |
movzbl %r12b, %r15d
|
|
Packit |
6c4009 |
shlq $4, %r15
|
|
Packit |
6c4009 |
movsd 192(%rsp,%r15), %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
call JUMPTARGET(__exp_finite)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movsd %xmm0, 256(%rsp,%r15)
|
|
Packit |
6c4009 |
jmp .LBL_1_7
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
END (_ZGVbN2v_exp_sse4)
|