1dnl ARM64 mpn_lshiftc. 2 3dnl Copyright 2013, 2014, 2017 Free Software Foundation, Inc. 4 5dnl This file is part of the GNU MP Library. 6 7dnl The GNU MP Library is free software; you can redistribute it and/or modify 8dnl it under the terms of the GNU Lesser General Public License as published 9dnl by the Free Software Foundation; either version 3 of the License, or (at 10dnl your option) any later version. 11 12dnl The GNU MP Library is distributed in the hope that it will be useful, but 13dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 14dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public 15dnl License for more details. 16 17dnl You should have received a copy of the GNU Lesser General Public License 18dnl along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. 19 20include(`../config.m4') 21 22C cycles/limb assumed optimal c/l 23C Cortex-A53 3.5-4.0 3.25 24C Cortex-A57 2.0 2.0 25C X-Gene 2.67 2.5 26 27C TODO 28C * The feed-in code used 1 ldr for odd sized and 2 ldr for even sizes. These 29C numbers should be 1 and 0, respectively. The str in wind-down should also 30C go. 31C * Using extr and with 63 separate loops we might reach 1.5 c/l on A57. 32C * A53's speed depends on alignment, tune/speed -w1 gives 3.5, -w0 gives 4.0. 33 34changecom(blah) 35 36define(`rp_arg', `x0') 37define(`up', `x1') 38define(`n', `x2') 39define(`cnt', `x3') 40 41define(`rp', `x16') 42 43define(`tnc',`x8') 44 45define(`PSHIFT', lsl) 46define(`NSHIFT', lsr) 47 48ASM_START() 49PROLOGUE(mpn_lshiftc) 50 add rp, rp_arg, n, lsl #3 51 add up, up, n, lsl #3 52 sub tnc, xzr, cnt 53 lsr x18, n, #2 54 tbz n, #0, L(bx0) 55 56L(bx1): ldr x4, [up,#-8] 57 tbnz n, #1, L(b11) 58 59L(b01): NSHIFT x0, x4, tnc 60 PSHIFT x2, x4, cnt 61 cbnz x18, L(gt1) 62 mvn x2, x2 63 str x2, [rp,#-8] 64 ret 65L(gt1): ldp x4, x5, [up,#-24] 66 sub up, up, #8 67 add rp, rp, #16 68 b L(lo2) 69 70L(b11): NSHIFT x0, x4, tnc 71 PSHIFT x2, x4, cnt 72 ldp x6, x7, [up,#-24]! 73 b L(lo3) 74 75L(bx0): ldp x4, x5, [up,#-16] 76 tbz n, #1, L(b00) 77 78L(b10): NSHIFT x0, x5, tnc 79 PSHIFT x13, x5, cnt 80 NSHIFT x10, x4, tnc 81 PSHIFT x2, x4, cnt 82 cbnz x18, L(gt2) 83 eon x10, x10, x13 84 mvn x2, x2 85 stp x2, x10, [rp,#-16] 86 ret 87L(gt2): ldp x4, x5, [up,#-32] 88 eon x10, x10, x13 89 str x10, [rp,#-8] 90 sub up, up, #16 91 add rp, rp, #8 92 b L(lo2) 93 94L(b00): NSHIFT x0, x5, tnc 95 PSHIFT x13, x5, cnt 96 NSHIFT x10, x4, tnc 97 PSHIFT x2, x4, cnt 98 ldp x6, x7, [up,#-32]! 99 eon x10, x10, x13 100 str x10, [rp,#-8]! 101 b L(lo0) 102 103 ALIGN(16) 104L(top): ldp x4, x5, [up,#-16] 105 eon x10, x10, x13 106 eon x11, x12, x2 107 stp x10, x11, [rp,#-16] 108 PSHIFT x2, x6, cnt 109L(lo2): NSHIFT x10, x4, tnc 110 PSHIFT x13, x5, cnt 111 NSHIFT x12, x5, tnc 112 ldp x6, x7, [up,#-32]! 113 eon x10, x10, x13 114 eon x11, x12, x2 115 stp x10, x11, [rp,#-32]! 116 PSHIFT x2, x4, cnt 117L(lo0): sub x18, x18, #1 118L(lo3): NSHIFT x10, x6, tnc 119 PSHIFT x13, x7, cnt 120 NSHIFT x12, x7, tnc 121 cbnz x18, L(top) 122 123L(end): eon x10, x10, x13 124 eon x11, x12, x2 125 PSHIFT x2, x6, cnt 126 stp x10, x11, [rp,#-16] 127 mvn x2, x2 128 str x2, [rp,#-24] 129 ret 130EPILOGUE() 131