1dnl IA-64 mpn_invert_limb -- Invert a normalized limb. 2 3dnl Contributed to the GNU project by Torbjorn Granlund and Kevin Ryde. 4 5dnl Copyright 2000, 2002, 2004 Free Software Foundation, Inc. 6 7dnl This file is part of the GNU MP Library. 8dnl 9dnl The GNU MP Library is free software; you can redistribute it and/or modify 10dnl it under the terms of either: 11dnl 12dnl * the GNU Lesser General Public License as published by the Free 13dnl Software Foundation; either version 3 of the License, or (at your 14dnl option) any later version. 15dnl 16dnl or 17dnl 18dnl * the GNU General Public License as published by the Free Software 19dnl Foundation; either version 2 of the License, or (at your option) any 20dnl later version. 21dnl 22dnl or both in parallel, as here. 23dnl 24dnl The GNU MP Library is distributed in the hope that it will be useful, but 25dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 26dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 27dnl for more details. 28dnl 29dnl You should have received copies of the GNU General Public License and the 30dnl GNU Lesser General Public License along with the GNU MP Library. If not, 31dnl see https://www.gnu.org/licenses/. 32 33include(`../config.m4') 34 35C INPUT PARAMETERS 36C d = r32 37 38C cycles 39C Itanium: 74 40C Itanium 2: 50+6 41 42C It should be possible to avoid the xmpy.hu and the following tests by 43C explicitly chopping in the last fma. That would save about 10 cycles. 44 45ASM_START() 46 .sdata 47 .align 16 48ifdef(`HAVE_DOUBLE_IEEE_LITTLE_ENDIAN',` 49.LC0: data4 0x00000000, 0x80000000, 0x0000403f, 0x00000000 C 2^64 50.LC1: data4 0x00000000, 0x80000000, 0x0000407f, 0x00000000 C 2^128 51 52',`ifdef(`HAVE_DOUBLE_IEEE_BIG_ENDIAN',` 53.LC0: data4 0x403f8000, 0x00000000, 0x00000000, 0x00000000 C 2^64 54.LC1: data4 0x407f8000, 0x00000000, 0x00000000, 0x00000000 C 2^128 55 56',`m4_error(`Oops, need to know float endianness 57')')') 58 59 60PROLOGUE(mpn_invert_limb) 61 C 00 62 addl r14 = @gprel(.LC0), gp 63 addl r15 = @gprel(.LC1), gp 64 setf.sig f7 = r32 65 add r9 = r32, r32 C check for d = 2^63 66 ;; C 01 67 ldfe f10 = [r14] C 2^64 68 ldfe f8 = [r15] C 2^128 69 cmp.eq p6, p0 = 0, r9 C check for d = 2^63 70 mov r8 = -1 C retval for 2^63 71 (p6) br.ret.spnt.many b0 72 ;; C 07 73 fmpy.s1 f11 = f7, f10 C f11 = d * 2^64 74 fnma.s1 f6 = f7, f10, f8 C f6 = 2^128 - d * 2^64 75 ;; C 11 76 frcpa.s1 f8, p6 = f6, f7 77 ;; C 15 78 (p6) fnma.s1 f9 = f7, f8, f1 79 (p6) fmpy.s1 f10 = f6, f8 80 ;; C 19 81 (p6) fmpy.s1 f11 = f9, f9 82 (p6) fma.s1 f10 = f9, f10, f10 83 ;; C 23 84 (p6) fma.s1 f8 = f9, f8, f8 85 (p6) fma.s1 f9 = f11, f10, f10 86 ;; C 27 87 (p6) fma.s1 f8 = f11, f8, f8 88 (p6) fnma.s1 f10 = f7, f9, f6 89 ;; C 31 90 (p6) fma.s1 f8 = f10, f8, f9 91 ;; C 35 92 fcvt.fxu.trunc.s1 f8 = f8 93 ;; C 39 94 getf.sig r8 = f8 95 xmpy.hu f10 = f8, f7 C di * d 96 ;; C 43 97 getf.sig r14 = f10 98 andcm r9 = -1, r32 C one's complement 99 ;; C 48 100 cmp.ltu p6, p0 = r9, r14 C got overflow? 101 ;; C 49 102 (p6) add r8 = -1, r8 C adjust di down 103 br.ret.sptk.many b0 104EPILOGUE() 105ASM_END() 106