1/*---------------------------------------------------------------------------+ 2 | poly.h | 3 | | 4 | Header file for the FPU-emu poly*.c source files. | 5 | | 6 | Copyright (C) 1994,1999 | 7 | W. Metzenthen, 22 Parker St, Ormond, Vic 3163, | 8 | Australia. E-mail billm@melbpc.org.au | 9 | | 10 | Declarations and definitions for functions operating on Xsig (12-byte | 11 | extended-significand) quantities. | 12 | | 13 +---------------------------------------------------------------------------*/ 14 15#ifndef _POLY_H 16#define _POLY_H 17 18/* This 12-byte structure is used to improve the accuracy of computation 19 of transcendental functions. 20 Intended to be used to get results better than 8-byte computation 21 allows. 9-byte would probably be sufficient. 22 */ 23typedef struct { 24 unsigned long lsw; 25 unsigned long midw; 26 unsigned long msw; 27} Xsig; 28 29asmlinkage void mul64(unsigned long long const *a, unsigned long long const *b, 30 unsigned long long *result); 31asmlinkage void polynomial_Xsig(Xsig *, const unsigned long long *x, 32 const unsigned long long terms[], const int n); 33 34asmlinkage void mul32_Xsig(Xsig *, const unsigned long mult); 35asmlinkage void mul64_Xsig(Xsig *, const unsigned long long *mult); 36asmlinkage void mul_Xsig_Xsig(Xsig *dest, const Xsig *mult); 37 38asmlinkage void shr_Xsig(Xsig *, const int n); 39asmlinkage int round_Xsig(Xsig *); 40asmlinkage int norm_Xsig(Xsig *); 41asmlinkage void div_Xsig(Xsig *x1, const Xsig *x2, const Xsig *dest); 42 43/* Macro to extract the most significant 32 bits from a long long */ 44#define LL_MSW(x) (((unsigned long *)&x)[1]) 45 46/* Macro to initialize an Xsig struct */ 47#define MK_XSIG(a,b,c) { c, b, a } 48 49/* Macro to access the 8 ms bytes of an Xsig as a long long */ 50#define XSIG_LL(x) (*(unsigned long long *)&x.midw) 51 52/* 53 Need to run gcc with optimizations on to get these to 54 actually be in-line. 55 */ 56 57/* Multiply two fixed-point 32 bit numbers, producing a 32 bit result. 58 The answer is the ms word of the product. */ 59/* Some versions of gcc make it difficult to stop eax from being clobbered. 60 Merely specifying that it is used doesn't work... 61 */ 62static inline unsigned long mul_32_32(const unsigned long arg1, 63 const unsigned long arg2) 64{ 65 int retval; 66 asm volatile ("mull %2; movl %%edx,%%eax":"=a" (retval) 67 :"0"(arg1), "g"(arg2) 68 :"dx"); 69 return retval; 70} 71 72/* Add the 12 byte Xsig x2 to Xsig dest, with no checks for overflow. */ 73static inline void add_Xsig_Xsig(Xsig *dest, const Xsig *x2) 74{ 75 asm volatile ("movl %1,%%edi; movl %2,%%esi;\n" 76 "movl (%%esi),%%eax; addl %%eax,(%%edi);\n" 77 "movl 4(%%esi),%%eax; adcl %%eax,4(%%edi);\n" 78 "movl 8(%%esi),%%eax; adcl %%eax,8(%%edi);\n":"=g" 79 (*dest):"g"(dest), "g"(x2) 80 :"ax", "si", "di"); 81} 82 83/* Add the 12 byte Xsig x2 to Xsig dest, adjust exp if overflow occurs. */ 84/* Note: the constraints in the asm statement didn't always work properly 85 with gcc 2.5.8. Changing from using edi to using ecx got around the 86 problem, but keep fingers crossed! */ 87static inline void add_two_Xsig(Xsig *dest, const Xsig *x2, long int *exp) 88{ 89 asm volatile ("movl %2,%%ecx; movl %3,%%esi;\n" 90 "movl (%%esi),%%eax; addl %%eax,(%%ecx);\n" 91 "movl 4(%%esi),%%eax; adcl %%eax,4(%%ecx);\n" 92 "movl 8(%%esi),%%eax; adcl %%eax,8(%%ecx);\n" 93 "jnc 0f;\n" 94 "rcrl 8(%%ecx); rcrl 4(%%ecx); rcrl (%%ecx)\n" 95 "movl %4,%%ecx; incl (%%ecx)\n" 96 "movl $1,%%eax; jmp 1f;\n" 97 "0: xorl %%eax,%%eax;\n" "1:\n":"=g" (*exp), "=g"(*dest) 98 :"g"(dest), "g"(x2), "g"(exp) 99 :"cx", "si", "ax"); 100} 101 102/* Negate (subtract from 1.0) the 12 byte Xsig */ 103/* This is faster in a loop on my 386 than using the "neg" instruction. */ 104static inline void negate_Xsig(Xsig *x) 105{ 106 asm volatile ("movl %1,%%esi;\n" 107 "xorl %%ecx,%%ecx;\n" 108 "movl %%ecx,%%eax; subl (%%esi),%%eax; movl %%eax,(%%esi);\n" 109 "movl %%ecx,%%eax; sbbl 4(%%esi),%%eax; movl %%eax,4(%%esi);\n" 110 "movl %%ecx,%%eax; sbbl 8(%%esi),%%eax; movl %%eax,8(%%esi);\n":"=g" 111 (*x):"g"(x):"si", "ax", "cx"); 112} 113 114#endif /* _POLY_H */ 115