| 1 |
dnl AMD64 SSE mpn_sec_tabselect. |
| 2 |
|
| 3 |
dnl Contributed to the GNU project by Torbjörn Granlund. |
| 4 |
|
| 5 |
dnl Copyright 2011-2013 Free Software Foundation, Inc. |
| 6 |
|
| 7 |
dnl This file is part of the GNU MP Library. |
| 8 |
dnl |
| 9 |
dnl The GNU MP Library is free software; you can redistribute it and/or modify |
| 10 |
dnl it under the terms of either: |
| 11 |
dnl |
| 12 |
dnl * the GNU Lesser General Public License as published by the Free |
| 13 |
dnl Software Foundation; either version 3 of the License, or (at your |
| 14 |
dnl option) any later version. |
| 15 |
dnl |
| 16 |
dnl or |
| 17 |
dnl |
| 18 |
dnl * the GNU General Public License as published by the Free Software |
| 19 |
dnl Foundation; either version 2 of the License, or (at your option) any |
| 20 |
dnl later version. |
| 21 |
dnl |
| 22 |
dnl or both in parallel, as here. |
| 23 |
dnl |
| 24 |
dnl The GNU MP Library is distributed in the hope that it will be useful, but |
| 25 |
dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY |
| 26 |
dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 27 |
dnl for more details. |
| 28 |
dnl |
| 29 |
dnl You should have received copies of the GNU General Public License and the |
| 30 |
dnl GNU Lesser General Public License along with the GNU MP Library. If not, |
| 31 |
dnl see https://www.gnu.org/licenses/. |
| 32 |
|
| 33 |
include(`../config.m4') |
| 34 |
|
| 35 |
|
| 36 |
C cycles/limb cycles/limb cycles/limb |
| 37 |
C ali,evn n unal,evn n other cases |
| 38 |
C AMD K8,K9 1.65 1.65 1.8 |
| 39 |
C AMD K10 0.78 0.78 0.85 |
| 40 |
C AMD bd1 0.80 0.91 1.25 |
| 41 |
C AMD bobcat 2.15 2.15 2.37 |
| 42 |
C Intel P4 2.5 2.5 2.95 |
| 43 |
C Intel core2 1.17 1.25 1.25 |
| 44 |
C Intel NHM 0.87 0.90 0.90 |
| 45 |
C Intel SBR 0.63 0.79 0.77 |
| 46 |
C Intel atom 4.3 4.3 4.3 slower than plain code |
| 47 |
C VIA nano 1.4 5.1 3.14 too alignment dependent |
| 48 |
|
| 49 |
C NOTES |
| 50 |
C * We only honour the least significant 32 bits of the `which' and `nents' |
| 51 |
C arguments to allow efficient code using just SSE2. We would need to |
| 52 |
C either use the SSE4_1 pcmpeqq, or find some other SSE2 sequence. |
| 53 |
C * We use movd for copying between xmm and plain registers, since old gas |
| 54 |
C rejects movq. But gas assembles movd as movq when given a 64-bit greg. |
| 55 |
|
| 56 |
define(`rp', `%rdi') |
| 57 |
define(`tp', `%rsi') |
| 58 |
define(`n', `%rdx') |
| 59 |
define(`nents', `%rcx') |
| 60 |
define(`which', `%r8') |
| 61 |
|
| 62 |
define(`i', `%r10') |
| 63 |
define(`j', `%r9') |
| 64 |
|
| 65 |
C rax rbx rcx rdx rdi rsi rbp r8 r9 r10 r11 r12 r13 r14 r15 |
| 66 |
C nents n rp tab which j i temp * * * * |
| 67 |
|
| 68 |
ABI_SUPPORT(DOS64) |
| 69 |
ABI_SUPPORT(STD64) |
| 70 |
|
| 71 |
ASM_START() |
| 72 |
TEXT |
| 73 |
ALIGN(16) |
| 74 |
PROLOGUE(mpn_sec_tabselect) |
| 75 |
FUNC_ENTRY(4) |
| 76 |
IFDOS(` mov 56(%rsp), %r8d ') |
| 77 |
|
| 78 |
movd which, %xmm8 |
| 79 |
pshufd $0, %xmm8, %xmm8 C 4 `which' copies |
| 80 |
mov $1, R32(%rax) |
| 81 |
movd %rax, %xmm9 |
| 82 |
pshufd $0, %xmm9, %xmm9 C 4 copies of 1 |
| 83 |
|
| 84 |
mov n, j |
| 85 |
add $-8, j |
| 86 |
js L(outer_end) |
| 87 |
|
| 88 |
L(outer_top): |
| 89 |
mov nents, i |
| 90 |
mov tp, %r11 |
| 91 |
pxor %xmm13, %xmm13 |
| 92 |
pxor %xmm4, %xmm4 |
| 93 |
pxor %xmm5, %xmm5 |
| 94 |
pxor %xmm6, %xmm6 |
| 95 |
pxor %xmm7, %xmm7 |
| 96 |
ALIGN(16) |
| 97 |
L(top): movdqa %xmm8, %xmm0 |
| 98 |
pcmpeqd %xmm13, %xmm0 |
| 99 |
paddd %xmm9, %xmm13 |
| 100 |
movdqu 0(tp), %xmm2 |
| 101 |
movdqu 16(tp), %xmm3 |
| 102 |
pand %xmm0, %xmm2 |
| 103 |
pand %xmm0, %xmm3 |
| 104 |
por %xmm2, %xmm4 |
| 105 |
por %xmm3, %xmm5 |
| 106 |
movdqu 32(tp), %xmm2 |
| 107 |
movdqu 48(tp), %xmm3 |
| 108 |
pand %xmm0, %xmm2 |
| 109 |
pand %xmm0, %xmm3 |
| 110 |
por %xmm2, %xmm6 |
| 111 |
por %xmm3, %xmm7 |
| 112 |
lea (tp,n,8), tp |
| 113 |
add $-1, i |
| 114 |
jne L(top) |
| 115 |
|
| 116 |
movdqu %xmm4, 0(rp) |
| 117 |
movdqu %xmm5, 16(rp) |
| 118 |
movdqu %xmm6, 32(rp) |
| 119 |
movdqu %xmm7, 48(rp) |
| 120 |
|
| 121 |
lea 64(%r11), tp |
| 122 |
lea 64(rp), rp |
| 123 |
add $-8, j |
| 124 |
jns L(outer_top) |
| 125 |
L(outer_end): |
| 126 |
|
| 127 |
test $4, R8(n) |
| 128 |
je L(b0xx) |
| 129 |
L(b1xx):mov nents, i |
| 130 |
mov tp, %r11 |
| 131 |
pxor %xmm13, %xmm13 |
| 132 |
pxor %xmm4, %xmm4 |
| 133 |
pxor %xmm5, %xmm5 |
| 134 |
ALIGN(16) |
| 135 |
L(tp4): movdqa %xmm8, %xmm0 |
| 136 |
pcmpeqd %xmm13, %xmm0 |
| 137 |
paddd %xmm9, %xmm13 |
| 138 |
movdqu 0(tp), %xmm2 |
| 139 |
movdqu 16(tp), %xmm3 |
| 140 |
pand %xmm0, %xmm2 |
| 141 |
pand %xmm0, %xmm3 |
| 142 |
por %xmm2, %xmm4 |
| 143 |
por %xmm3, %xmm5 |
| 144 |
lea (tp,n,8), tp |
| 145 |
add $-1, i |
| 146 |
jne L(tp4) |
| 147 |
movdqu %xmm4, 0(rp) |
| 148 |
movdqu %xmm5, 16(rp) |
| 149 |
lea 32(%r11), tp |
| 150 |
lea 32(rp), rp |
| 151 |
|
| 152 |
L(b0xx):test $2, R8(n) |
| 153 |
je L(b00x) |
| 154 |
L(b01x):mov nents, i |
| 155 |
mov tp, %r11 |
| 156 |
pxor %xmm13, %xmm13 |
| 157 |
pxor %xmm4, %xmm4 |
| 158 |
ALIGN(16) |
| 159 |
L(tp2): movdqa %xmm8, %xmm0 |
| 160 |
pcmpeqd %xmm13, %xmm0 |
| 161 |
paddd %xmm9, %xmm13 |
| 162 |
movdqu 0(tp), %xmm2 |
| 163 |
pand %xmm0, %xmm2 |
| 164 |
por %xmm2, %xmm4 |
| 165 |
lea (tp,n,8), tp |
| 166 |
add $-1, i |
| 167 |
jne L(tp2) |
| 168 |
movdqu %xmm4, 0(rp) |
| 169 |
lea 16(%r11), tp |
| 170 |
lea 16(rp), rp |
| 171 |
|
| 172 |
L(b00x):test $1, R8(n) |
| 173 |
je L(b000) |
| 174 |
L(b001):mov nents, i |
| 175 |
mov tp, %r11 |
| 176 |
pxor %xmm13, %xmm13 |
| 177 |
pxor %xmm4, %xmm4 |
| 178 |
ALIGN(16) |
| 179 |
L(tp1): movdqa %xmm8, %xmm0 |
| 180 |
pcmpeqd %xmm13, %xmm0 |
| 181 |
paddd %xmm9, %xmm13 |
| 182 |
movq 0(tp), %xmm2 |
| 183 |
pand %xmm0, %xmm2 |
| 184 |
por %xmm2, %xmm4 |
| 185 |
lea (tp,n,8), tp |
| 186 |
add $-1, i |
| 187 |
jne L(tp1) |
| 188 |
movq %xmm4, 0(rp) |
| 189 |
|
| 190 |
L(b000):FUNC_EXIT() |
| 191 |
ret |
| 192 |
EPILOGUE() |