Commit | Line | Data |
---|---|---|
b2441318 | 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
1da177e4 LT |
2 | /*---------------------------------------------------------------------------+ |
3 | | poly.h | | |
4 | | | | |
5 | | Header file for the FPU-emu poly*.c source files. | | |
6 | | | | |
7 | | Copyright (C) 1994,1999 | | |
8 | | W. Metzenthen, 22 Parker St, Ormond, Vic 3163, | | |
9 | | Australia. E-mail billm@melbpc.org.au | | |
10 | | | | |
11 | | Declarations and definitions for functions operating on Xsig (12-byte | | |
12 | | extended-significand) quantities. | | |
13 | | | | |
14 | +---------------------------------------------------------------------------*/ | |
15 | ||
16 | #ifndef _POLY_H | |
17 | #define _POLY_H | |
18 | ||
19 | /* This 12-byte structure is used to improve the accuracy of computation | |
20 | of transcendental functions. | |
21 | Intended to be used to get results better than 8-byte computation | |
22 | allows. 9-byte would probably be sufficient. | |
23 | */ | |
24 | typedef struct { | |
3d0d14f9 IM |
25 | unsigned long lsw; |
26 | unsigned long midw; | |
27 | unsigned long msw; | |
1da177e4 LT |
28 | } Xsig; |
29 | ||
30 | asmlinkage void mul64(unsigned long long const *a, unsigned long long const *b, | |
31 | unsigned long long *result); | |
32 | asmlinkage void polynomial_Xsig(Xsig *, const unsigned long long *x, | |
33 | const unsigned long long terms[], const int n); | |
34 | ||
35 | asmlinkage void mul32_Xsig(Xsig *, const unsigned long mult); | |
36 | asmlinkage void mul64_Xsig(Xsig *, const unsigned long long *mult); | |
e8d591dc | 37 | asmlinkage void mul_Xsig_Xsig(Xsig *dest, const Xsig *mult); |
1da177e4 LT |
38 | |
39 | asmlinkage void shr_Xsig(Xsig *, const int n); | |
40 | asmlinkage int round_Xsig(Xsig *); | |
41 | asmlinkage int norm_Xsig(Xsig *); | |
e8d591dc | 42 | asmlinkage void div_Xsig(Xsig *x1, const Xsig *x2, const Xsig *dest); |
1da177e4 LT |
43 | |
44 | /* Macro to extract the most significant 32 bits from a long long */ | |
45 | #define LL_MSW(x) (((unsigned long *)&x)[1]) | |
46 | ||
47 | /* Macro to initialize an Xsig struct */ | |
48 | #define MK_XSIG(a,b,c) { c, b, a } | |
49 | ||
50 | /* Macro to access the 8 ms bytes of an Xsig as a long long */ | |
51 | #define XSIG_LL(x) (*(unsigned long long *)&x.midw) | |
52 | ||
1da177e4 LT |
53 | /* |
54 | Need to run gcc with optimizations on to get these to | |
55 | actually be in-line. | |
56 | */ | |
57 | ||
58 | /* Multiply two fixed-point 32 bit numbers, producing a 32 bit result. | |
59 | The answer is the ms word of the product. */ | |
60 | /* Some versions of gcc make it difficult to stop eax from being clobbered. | |
61 | Merely specifying that it is used doesn't work... | |
62 | */ | |
63 | static inline unsigned long mul_32_32(const unsigned long arg1, | |
64 | const unsigned long arg2) | |
65 | { | |
3d0d14f9 IM |
66 | int retval; |
67 | asm volatile ("mull %2; movl %%edx,%%eax":"=a" (retval) | |
68 | :"0"(arg1), "g"(arg2) | |
69 | :"dx"); | |
70 | return retval; | |
1da177e4 LT |
71 | } |
72 | ||
1da177e4 | 73 | /* Add the 12 byte Xsig x2 to Xsig dest, with no checks for overflow. */ |
e8d591dc | 74 | static inline void add_Xsig_Xsig(Xsig *dest, const Xsig *x2) |
1da177e4 | 75 | { |
3d0d14f9 IM |
76 | asm volatile ("movl %1,%%edi; movl %2,%%esi;\n" |
77 | "movl (%%esi),%%eax; addl %%eax,(%%edi);\n" | |
78 | "movl 4(%%esi),%%eax; adcl %%eax,4(%%edi);\n" | |
79 | "movl 8(%%esi),%%eax; adcl %%eax,8(%%edi);\n":"=g" | |
80 | (*dest):"g"(dest), "g"(x2) | |
81 | :"ax", "si", "di"); | |
1da177e4 LT |
82 | } |
83 | ||
1da177e4 LT |
84 | /* Add the 12 byte Xsig x2 to Xsig dest, adjust exp if overflow occurs. */ |
85 | /* Note: the constraints in the asm statement didn't always work properly | |
86 | with gcc 2.5.8. Changing from using edi to using ecx got around the | |
87 | problem, but keep fingers crossed! */ | |
e8d591dc | 88 | static inline void add_two_Xsig(Xsig *dest, const Xsig *x2, long int *exp) |
1da177e4 | 89 | { |
3d0d14f9 IM |
90 | asm volatile ("movl %2,%%ecx; movl %3,%%esi;\n" |
91 | "movl (%%esi),%%eax; addl %%eax,(%%ecx);\n" | |
92 | "movl 4(%%esi),%%eax; adcl %%eax,4(%%ecx);\n" | |
93 | "movl 8(%%esi),%%eax; adcl %%eax,8(%%ecx);\n" | |
94 | "jnc 0f;\n" | |
95 | "rcrl 8(%%ecx); rcrl 4(%%ecx); rcrl (%%ecx)\n" | |
96 | "movl %4,%%ecx; incl (%%ecx)\n" | |
97 | "movl $1,%%eax; jmp 1f;\n" | |
98 | "0: xorl %%eax,%%eax;\n" "1:\n":"=g" (*exp), "=g"(*dest) | |
99 | :"g"(dest), "g"(x2), "g"(exp) | |
100 | :"cx", "si", "ax"); | |
1da177e4 LT |
101 | } |
102 | ||
1da177e4 LT |
103 | /* Negate (subtract from 1.0) the 12 byte Xsig */ |
104 | /* This is faster in a loop on my 386 than using the "neg" instruction. */ | |
e8d591dc | 105 | static inline void negate_Xsig(Xsig *x) |
1da177e4 | 106 | { |
3d0d14f9 IM |
107 | asm volatile ("movl %1,%%esi;\n" |
108 | "xorl %%ecx,%%ecx;\n" | |
109 | "movl %%ecx,%%eax; subl (%%esi),%%eax; movl %%eax,(%%esi);\n" | |
110 | "movl %%ecx,%%eax; sbbl 4(%%esi),%%eax; movl %%eax,4(%%esi);\n" | |
111 | "movl %%ecx,%%eax; sbbl 8(%%esi),%%eax; movl %%eax,8(%%esi);\n":"=g" | |
112 | (*x):"g"(x):"si", "ax", "cx"); | |
1da177e4 LT |
113 | } |
114 | ||
115 | #endif /* _POLY_H */ |