1dnl ARM mpn_addmul_2. 2 3dnl Contributed to the GNU project by Torbjorn Granlund. 4 5dnl Copyright 2012 Free Software Foundation, Inc. 6 7dnl This file is part of the GNU MP Library. 8 9dnl The GNU MP Library is free software; you can redistribute it and/or modify 10dnl it under the terms of the GNU Lesser General Public License as published 11dnl by the Free Software Foundation; either version 3 of the License, or (at 12dnl your option) any later version. 13 14dnl The GNU MP Library is distributed in the hope that it will be useful, but 15dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 16dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public 17dnl License for more details. 18 19dnl You should have received a copy of the GNU Lesser General Public License 20dnl along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. 21 22include(`../config.m4') 23 24C cycles/limb 25C StrongARM: - 26C XScale - 27C Cortex-A8 ? 28C Cortex-A9 2.38 29C Cortex-A15 2.5 30 31C TODO 32C * Consider using more registers for the r[] loads, allowing better load-use 33C scheduling for a 6% speedup (on A9). Free: r10, r11, r14 34 35define(`rp',`r0') 36define(`up',`r1') 37define(`n', `r2') 38define(`vp',`r3') 39 40define(`v0',`r6') 41define(`v1',`r7') 42define(`u0',`r3') 43define(`u1',`r9') 44 45define(`cya',`r8') 46define(`cyb',`r12') 47 48 49ASM_START() 50PROLOGUE(mpn_addmul_2) 51 push { r4, r5, r6, r7, r8, r9 } 52 53 ldm vp, { v0, v1 } 54 mov cya, #0 55 mov cyb, #0 56 57 tst n, #1 58 beq L(evn) 59L(odd): ldr r5, [rp, #0] 60 ldr u0, [up, #0] 61 ldr r4, [rp, #4] 62 tst n, #2 63 beq L(fi1) 64L(fi3): sub up, up, #12 65 sub rp, rp, #16 66 b L(lo3) 67L(fi1): sub n, n, #1 68 sub up, up, #4 69 sub rp, rp, #8 70 b L(lo1) 71L(evn): ldr r4, [rp, #0] 72 ldr u1, [up, #0] 73 ldr r5, [rp, #4] 74 tst n, #2 75 bne L(fi2) 76L(fi0): sub up, up, #8 77 sub rp, rp, #12 78 b L(lo0) 79L(fi2): subs n, n, #2 80 sub rp, rp, #4 81 bls L(end) 82 83 ALIGN(16) 84L(top): ldr u0, [up, #4] 85 umaal r4, cya, u1, v0 86 str r4, [rp, #4] 87 ldr r4, [rp, #12] 88 umaal r5, cyb, u1, v1 89L(lo1): ldr u1, [up, #8] 90 umaal r5, cya, u0, v0 91 str r5, [rp, #8] 92 ldr r5, [rp, #16] 93 umaal r4, cyb, u0, v1 94L(lo0): ldr u0, [up, #12] 95 umaal r4, cya, u1, v0 96 str r4, [rp, #12] 97 ldr r4, [rp, #20] 98 umaal r5, cyb, u1, v1 99L(lo3): ldr u1, [up, #16]! 100 umaal r5, cya, u0, v0 101 str r5, [rp, #16]! 102 ldr r5, [rp, #8] 103 umaal r4, cyb, u0, v1 104 subs n, n, #4 105 bhi L(top) 106 107L(end): umaal r4, cya, u1, v0 108 ldr u0, [up, #4] 109 umaal r5, cyb, u1, v1 110 str r4, [rp, #4] 111 umaal r5, cya, u0, v0 112 umaal cya, cyb, u0, v1 113 str r5, [rp, #8] 114 str cya, [rp, #12] 115 mov r0, cyb 116 117 pop { r4, r5, r6, r7, r8, r9 } 118 bx r14 119EPILOGUE() 120