1dnl x86 fat binary entrypoints. 2 3dnl Contributed to the GNU project by Kevin Ryde (original x86_32 code) and 4dnl Torbjorn Granlund (port to x86_64) 5 6dnl Copyright 2003, 2009, 2011 Free Software Foundation, Inc. 7dnl 8dnl This file is part of the GNU MP Library. 9dnl 10dnl The GNU MP Library is free software; you can redistribute it and/or 11dnl modify it under the terms of the GNU Lesser General Public License as 12dnl published by the Free Software Foundation; either version 3 of the 13dnl License, or (at your option) any later version. 14dnl 15dnl The GNU MP Library is distributed in the hope that it will be useful, 16dnl but WITHOUT ANY WARRANTY; without even the implied warranty of 17dnl MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 18dnl Lesser General Public License for more details. 19dnl 20dnl You should have received a copy of the GNU Lesser General Public License 21dnl along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. 22 23include(`../config.m4') 24 25 26dnl Forcibly disable profiling. 27dnl 28dnl The entrypoints and inits are small enough not to worry about, the real 29dnl routines arrived at will have any profiling. Also, the way the code 30dnl here ends with a jump means we won't work properly with the 31dnl "instrument" profiling scheme anyway. 32 33define(`WANT_PROFILING',no) 34 35 36dnl We define PIC_OR_DARWIN as a helper symbol, the use it for suppressing 37dnl normal, fast call code, since that triggers problems on darwin. 38dnl 39dnl FIXME: There might be a more elegant solution, adding less overhead. 40 41ifdef(`DARWIN', 42`define(`PIC_OR_DARWIN')') 43ifdef(`PIC', 44`define(`PIC_OR_DARWIN')') 45 46 47 TEXT 48 49 50dnl Usage: FAT_ENTRY(name, offset) 51dnl 52dnl Emit a fat binary entrypoint function of the given name. This is the 53dnl normal entry for applications, eg. __gmpn_add_n. 54dnl 55dnl The code simply jumps through the function pointer in __gmpn_cpuvec at 56dnl the given "offset" (in bytes). 57dnl 58dnl For non-PIC, the jumps are 5 bytes each, aligning them to 8 should be 59dnl fine for all x86s. 60dnl 61dnl For PIC, the jumps are 20 bytes each, and are best aligned to 16 to 62dnl ensure at least the first two instructions don't cross a cache line 63dnl boundary. 64dnl 65dnl Note the extra `' ahead of PROLOGUE obscures it from the HAVE_NATIVE 66dnl grepping in configure, stopping that code trying to eval something with 67dnl $1 in it. 68 69define(FAT_ENTRY, 70m4_assert_numargs(2) 71` ALIGN(ifdef(`PIC',16,8)) 72`'PROLOGUE($1) 73ifdef(`PIC_OR_DARWIN', 74` LEA( GSYM_PREFIX`'__gmpn_cpuvec, %rax) 75 jmp *$2(%rax) 76',`dnl non-PIC 77 jmp *GSYM_PREFIX`'__gmpn_cpuvec+$2 78') 79EPILOGUE() 80') 81 82 83dnl FAT_ENTRY for each CPUVEC_FUNCS_LIST 84dnl 85 86define(`CPUVEC_offset',0) 87foreach(i, 88`FAT_ENTRY(MPN(i),CPUVEC_offset) 89define(`CPUVEC_offset',eval(CPUVEC_offset + 8))', 90CPUVEC_FUNCS_LIST) 91 92 93dnl Usage: FAT_INIT(name, offset) 94dnl 95dnl Emit a fat binary initializer function of the given name. These 96dnl functions are the initial values for the pointers in __gmpn_cpuvec. 97dnl 98dnl The code simply calls __gmpn_cpuvec_init, and then jumps back through 99dnl the __gmpn_cpuvec pointer, at the given "offset" (in bytes). 100dnl __gmpn_cpuvec_init will have stored the address of the selected 101dnl implementation there. 102dnl 103dnl Only one of these routines will be executed, and only once, since after 104dnl that all the __gmpn_cpuvec pointers go to real routines. So there's no 105dnl need for anything special here, just something small and simple. To 106dnl keep code size down, "fat_init" is a shared bit of code, arrived at 107dnl with the offset in %al. %al is used since the movb instruction is 2 108dnl bytes where %eax would be 4. 109dnl 110dnl Note having `PROLOGUE in FAT_INIT obscures that PROLOGUE from the 111dnl HAVE_NATIVE grepping in configure, preventing that code trying to eval 112dnl something with $1 in it. 113dnl 114dnl We need to preserve parameter registers over the __gmpn_cpuvec_init call 115 116define(FAT_INIT, 117m4_assert_numargs(2) 118`PROLOGUE($1) 119 mov $`'$2, %al 120 jmp L(fat_init) 121EPILOGUE() 122') 123 124L(fat_init): 125 C al __gmpn_cpuvec byte offset 126 127 movzbl %al, %eax 128 push %rdi 129 push %rsi 130 push %rdx 131 push %rcx 132 push %r8 133 push %r9 134 push %rax 135 CALL( __gmpn_cpuvec_init) 136 pop %rax 137 pop %r9 138 pop %r8 139 pop %rcx 140 pop %rdx 141 pop %rsi 142 pop %rdi 143ifdef(`PIC_OR_DARWIN',` 144 LEA( GSYM_PREFIX`'__gmpn_cpuvec, %r10) 145 jmp *(%r10,%rax) 146',`dnl non-PIC 147 jmp *GSYM_PREFIX`'__gmpn_cpuvec(%rax) 148') 149 150dnl FAT_INIT for each CPUVEC_FUNCS_LIST 151dnl 152 153define(`CPUVEC_offset',0) 154foreach(i, 155`FAT_INIT(MPN(i`'_init),CPUVEC_offset) 156define(`CPUVEC_offset',eval(CPUVEC_offset + 8))', 157CPUVEC_FUNCS_LIST) 158 159 160 161C long __gmpn_cpuid (char dst[12], int id); 162C 163C This is called only once, so just something simple and compact is fine. 164 165 166PROLOGUE(__gmpn_cpuid) 167 mov %rbx, %r8 168 mov %esi, %eax 169 cpuid 170 mov %ebx, (%rdi) 171 mov %edx, 4(%rdi) 172 mov %ecx, 8(%rdi) 173 mov %r8, %rbx 174 ret 175EPILOGUE() 176