1dnl ARM mpn_mod_1_1p 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 7 29C Cortex-A15 6 30 31define(`ap', `r0') 32define(`n', `r1') 33define(`d', `r2') 34define(`cps',`r3') 35 36ASM_START() 37PROLOGUE(mpn_mod_1_1p) 38 push {r4-r10} 39 add r0, r0, r1, asl #2 40 ldr r5, [r0, #-4]! 41 ldr r12, [r0, #-4]! 42 subs r1, r1, #2 43 ble L(4) 44 ldr r8, [r3, #12] 45 mov r4, r12 46 mov r10, r5 47 umull r7, r5, r10, r8 48 sub r1, r1, #1 49 b L(mid) 50 51L(top): adds r12, r6, r7 52 adcs r10, r4, r5 53 sub r1, r1, #1 54 mov r6, #0 55 movcs r6, r8 56 umull r7, r5, r10, r8 57 adds r4, r12, r6 58 subcs r4, r4, r2 59L(mid): ldr r6, [r0, #-4]! 60 teq r1, #0 61 bne L(top) 62 63 adds r12, r6, r7 64 adcs r5, r4, r5 65 subcs r5, r5, r2 66L(4): ldr r1, [r3, #4] 67 cmp r1, #0 68 beq L(7) 69 ldr r4, [r3, #8] 70 umull r0, r6, r5, r4 71 adds r12, r0, r12 72 addcs r6, r6, #1 73 rsb r0, r1, #32 74 mov r0, r12, lsr r0 75 orr r5, r0, r6, asl r1 76 mov r12, r12, asl r1 77 b L(8) 78L(7): cmp r5, r2 79 subcs r5, r5, r2 80L(8): ldr r0, [r3, #0] 81 umull r4, r3, r5, r0 82 add r5, r5, #1 83 adds r0, r4, r12 84 adc r5, r3, r5 85 mul r5, r2, r5 86 sub r12, r12, r5 87 cmp r12, r0 88 addhi r12, r12, r2 89 cmp r2, r12 90 subls r12, r12, r2 91 mov r0, r12, lsr r1 92 pop {r4-r10} 93 bx r14 94EPILOGUE() 95 96PROLOGUE(mpn_mod_1_1p_cps) 97 stmfd sp!, {r4, r5, r6, r14} 98 mov r5, r0 99 clz r4, r1 100 mov r0, r1, asl r4 101 rsb r6, r0, #0 102 bl mpn_invert_limb 103 str r0, [r5, #0] 104 str r4, [r5, #4] 105 cmp r4, #0 106 beq L(2) 107 rsb r1, r4, #32 108 mov r3, #1 109 mov r3, r3, asl r4 110 orr r3, r3, r0, lsr r1 111 mul r3, r6, r3 112 mov r4, r3, lsr r4 113 str r4, [r5, #8] 114L(2): mul r0, r6, r0 115 str r0, [r5, #12] 116 ldmfd sp!, {r4, r5, r6, pc} 117EPILOGUE() 118