lshiftc.asm revision 1.1
1 1.1 mrg dnl PowerPC-32 mpn_lshiftc. 2 1.1 mrg 3 1.1 mrg dnl Copyright 1995, 1998, 2000, 2002, 2003, 2004, 2005, 2010 Free Software 4 1.1 mrg dnl Foundation, Inc. 5 1.1 mrg 6 1.1 mrg dnl This file is part of the GNU MP Library. 7 1.1 mrg 8 1.1 mrg dnl The GNU MP Library is free software; you can redistribute it and/or modify 9 1.1 mrg dnl it under the terms of the GNU Lesser General Public License as published 10 1.1 mrg dnl by the Free Software Foundation; either version 3 of the License, or (at 11 1.1 mrg dnl your option) any later version. 12 1.1 mrg 13 1.1 mrg dnl The GNU MP Library is distributed in the hope that it will be useful, but 14 1.1 mrg dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 15 1.1 mrg dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public 16 1.1 mrg dnl License for more details. 17 1.1 mrg 18 1.1 mrg dnl You should have received a copy of the GNU Lesser General Public License 19 1.1 mrg dnl along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. 20 1.1 mrg 21 1.1 mrg include(`../config.m4') 22 1.1 mrg 23 1.1 mrg C cycles/limb 24 1.1 mrg C 603e: ? 25 1.1 mrg C 604e: 3.0 26 1.1 mrg C 75x (G3): 3.0 27 1.1 mrg C 7400,7410 (G4): 3.0 28 1.1 mrg C 7445,7455 (G4+): 2.5 29 1.1 mrg C 7447,7457 (G4+): 2.25 30 1.1 mrg C power4/ppc970: 2.5 31 1.1 mrg C power5: 2.5 32 1.1 mrg 33 1.1 mrg C INPUT PARAMETERS 34 1.1 mrg C rp r3 35 1.1 mrg C up r4 36 1.1 mrg C n r5 37 1.1 mrg C cnt r6 38 1.1 mrg 39 1.1 mrg ASM_START() 40 1.1 mrg PROLOGUE(mpn_lshiftc) 41 1.1 mrg cmpwi cr0, r5, 30 C more than 30 limbs? 42 1.1 mrg slwi r0, r5, 2 43 1.1 mrg add r4, r4, r0 C make r4 point at end of s1 44 1.1 mrg add r7, r3, r0 C make r7 point at end of res 45 1.1 mrg bgt L(BIG) C branch if more than 12 limbs 46 1.1 mrg 47 1.1 mrg mtctr r5 C copy size into CTR 48 1.1 mrg subfic r8, r6, 32 49 1.1 mrg lwzu r11, -4(r4) C load first s1 limb 50 1.1 mrg srw r3, r11, r8 C compute function return value 51 1.1 mrg bdz L(end1) 52 1.1 mrg 53 1.1 mrg L(oop): lwzu r10, -4(r4) 54 1.1 mrg slw r9, r11, r6 55 1.1 mrg srw r12, r10, r8 56 1.1 mrg nor r9, r9, r12 57 1.1 mrg stwu r9, -4(r7) 58 1.1 mrg bdz L(end2) 59 1.1 mrg lwzu r11, -4(r4) 60 1.1 mrg slw r9, r10, r6 61 1.1 mrg srw r12, r11, r8 62 1.1 mrg nor r9, r9, r12 63 1.1 mrg stwu r9, -4(r7) 64 1.1 mrg bdnz L(oop) 65 1.1 mrg 66 1.1 mrg L(end1): 67 1.1 mrg slw r0, r11, r6 68 1.1 mrg nor r0, r0, r0 69 1.1 mrg stw r0, -4(r7) 70 1.1 mrg blr 71 1.1 mrg L(end2): 72 1.1 mrg slw r0, r10, r6 73 1.1 mrg nor r0, r0, r0 74 1.1 mrg stw r0, -4(r7) 75 1.1 mrg blr 76 1.1 mrg 77 1.1 mrg L(BIG): 78 1.1 mrg stmw r24, -32(r1) C save registers we are supposed to preserve 79 1.1 mrg lwzu r9, -4(r4) 80 1.1 mrg subfic r8, r6, 32 81 1.1 mrg srw r3, r9, r8 C compute function return value 82 1.1 mrg slw r0, r9, r6 83 1.1 mrg addi r5, r5, -1 84 1.1 mrg 85 1.1 mrg andi. r10, r5, 3 C count for spill loop 86 1.1 mrg beq L(e) 87 1.1 mrg mtctr r10 88 1.1 mrg lwzu r28, -4(r4) 89 1.1 mrg bdz L(xe0) 90 1.1 mrg 91 1.1 mrg L(loop0): 92 1.1 mrg slw r12, r28, r6 93 1.1 mrg srw r24, r28, r8 94 1.1 mrg lwzu r28, -4(r4) 95 1.1 mrg nor r24, r0, r24 96 1.1 mrg stwu r24, -4(r7) 97 1.1 mrg mr r0, r12 98 1.1 mrg bdnz L(loop0) C taken at most once! 99 1.1 mrg 100 1.1 mrg L(xe0): slw r12, r28, r6 101 1.1 mrg srw r24, r28, r8 102 1.1 mrg nor r24, r0, r24 103 1.1 mrg stwu r24, -4(r7) 104 1.1 mrg mr r0, r12 105 1.1 mrg 106 1.1 mrg L(e): srwi r5, r5, 2 C count for unrolled loop 107 1.1 mrg addi r5, r5, -1 108 1.1 mrg mtctr r5 109 1.1 mrg lwz r28, -4(r4) 110 1.1 mrg lwz r29, -8(r4) 111 1.1 mrg lwz r30, -12(r4) 112 1.1 mrg lwzu r31, -16(r4) 113 1.1 mrg 114 1.1 mrg L(loopU): 115 1.1 mrg slw r9, r28, r6 116 1.1 mrg srw r24, r28, r8 117 1.1 mrg lwz r28, -4(r4) 118 1.1 mrg slw r10, r29, r6 119 1.1 mrg srw r25, r29, r8 120 1.1 mrg lwz r29, -8(r4) 121 1.1 mrg slw r11, r30, r6 122 1.1 mrg srw r26, r30, r8 123 1.1 mrg lwz r30, -12(r4) 124 1.1 mrg slw r12, r31, r6 125 1.1 mrg srw r27, r31, r8 126 1.1 mrg lwzu r31, -16(r4) 127 1.1 mrg nor r24, r0, r24 128 1.1 mrg stw r24, -4(r7) 129 1.1 mrg nor r25, r9, r25 130 1.1 mrg stw r25, -8(r7) 131 1.1 mrg nor r26, r10, r26 132 1.1 mrg stw r26, -12(r7) 133 1.1 mrg nor r27, r11, r27 134 1.1 mrg stwu r27, -16(r7) 135 1.1 mrg mr r0, r12 136 1.1 mrg bdnz L(loopU) 137 1.1 mrg 138 1.1 mrg slw r9, r28, r6 139 1.1 mrg srw r24, r28, r8 140 1.1 mrg slw r10, r29, r6 141 1.1 mrg srw r25, r29, r8 142 1.1 mrg slw r11, r30, r6 143 1.1 mrg srw r26, r30, r8 144 1.1 mrg slw r12, r31, r6 145 1.1 mrg srw r27, r31, r8 146 1.1 mrg nor r24, r0, r24 147 1.1 mrg stw r24, -4(r7) 148 1.1 mrg nor r25, r9, r25 149 1.1 mrg stw r25, -8(r7) 150 1.1 mrg nor r26, r10, r26 151 1.1 mrg stw r26, -12(r7) 152 1.1 mrg nor r27, r11, r27 153 1.1 mrg stw r27, -16(r7) 154 1.1 mrg nor r12, r12, r12 155 1.1 mrg stw r12, -20(r7) 156 1.1 mrg lmw r24, -32(r1) C restore registers 157 1.1 mrg blr 158 1.1 mrg EPILOGUE() 159
Indexes created Thu Apr 30 00:23:01 UTC 2026