overby.me
this repo has no description
1
2
3//
4// Basic types used internally in xmmLibm
5//
6
7#ifndef __XMMLIBM_PREFIX_H__
8#define __XMMLIBM_PREFIX_H__
9
10#include <stdint.h>
11#include <emmintrin.h>
12
13
14//The types we use
15typedef float xFloat __attribute__ ((vector_size (16)));
16typedef double xDouble __attribute__ ((vector_size (16)));
17typedef int32_t xSInt32 __attribute__ ((vector_size (16)));
18typedef uint32_t xUInt32 __attribute__ ((vector_size (16)));
19typedef int64_t xSInt64 __attribute__ ((vector_size (16)));
20typedef uint64_t xUInt64 __attribute__ ((vector_size (16)));
21typedef double DoubleConstant_sd;
22typedef xDouble DoubleConstant_pd;
23
24typedef union
25{
26 uint64_t u;
27 double d;
28}hexdouble;
29
30typedef union
31{
32 uint32_t u;
33 float f;
34}hexfloat;
35
36#if defined( __i386__ ) || defined( __x86_64__ )
37 #include <xmmintrin.h>
38 #include <emmintrin.h>
39
40 #define DEFAULT_MXCSR 0x1F80
41 #define INVALID_FLAG 0x0001
42 #define DENORMAL_FLAG 0x0002
43 #define DIVIDE_0_FLAG 0x0004
44 #define OVERFLOW_FLAG 0x0008
45 #define UNDERFLOW_FLAG 0x0010
46 #define INEXACT_FLAG 0x0020
47 #define ALL_FLAGS 0x3F
48 #define DEFAULT_MASK ( DEFAULT_MXCSR | ALL_FLAGS )
49 #define INVALID_MASK 0x0080
50 #define DENORMAL_MASK 0x0100
51 #define DIVIDE_0_MASK 0x0200
52 #define OVERFLOW_MASK 0x0400
53 #define UNDERFLOW_MASK 0x0800
54 #define INEXACT_MASK 0x1000
55 #define ALL_MASKS 0x1F80
56
57 #define ROUND_TO_NEAREST 0x0000
58 #define ROUND_TO_ZERO 0x6000
59 #define ROUND_TO_INFINITY 0x4000
60 #define ROUND_TO_NEG_INFINITY 0x2000
61 #define ROUND_MASK ROUND_TO_ZERO
62
63 static const xDouble minusZeroD = {-0.0, -0.0};
64 static const xFloat minusZeroF = {-0.0f, -0.0f, -0.0f, -0.0f };
65
66 //Macros to handle conversions between scalar types to vector types
67#if __APPLE_CC__-0 < 5342
68 #define DOUBLE_2_XDOUBLE( _d ) _mm_load_sd( &(_d) )
69 #define FLOAT_2_XFLOAT( _f ) _mm_load_ss( &(_f) )
70 #define XDOUBLE_2_DOUBLE( _xd ) ({ double _d; _mm_store_sd( &_d, _xd ); /*return*/ _d; })
71 #define XFLOAT_2_FLOAT( _xf ) ({ float _f; _mm_store_ss( &_f, _xf ); /*return*/ _f; })
72 #define FLOAT_2_XDOUBLE( _f ) ({ xFloat _xf = _mm_load_ss( &(_f) ); xDouble _xd = _mm_cvtss_sd( (xDouble) _xf, _xf ); /*return*/ _xd; })
73 #define XDOUBLE_2_FLOAT( _xd ) ({ float _f; xDouble _d = _xd; xFloat _xf = _mm_cvtsd_ss( (xFloat) _d, _d ); _mm_store_ss( &_f, _xf ); /* return*/ _f; })
74#else
75 #define DOUBLE_2_XDOUBLE( _d ) _mm_set_sd( _d )
76 #define FLOAT_2_XFLOAT( _f ) _mm_set_ss( _f )
77 #define XDOUBLE_2_DOUBLE( _xd ) _mm_cvtsd_f64( _xd )
78 #define XFLOAT_2_FLOAT( _xf ) _mm_cvtss_f32( _xf )
79 #define FLOAT_2_XDOUBLE( _f ) ({ xFloat _xf = FLOAT_2_XFLOAT(_f); xDouble _xd = _mm_cvtss_sd( (xDouble) _xf, _xf ); /*return*/ _xd; })
80 #define XDOUBLE_2_FLOAT( _xd ) ({ float _f; xDouble _d = _xd; xFloat _xf = _mm_cvtsd_ss( (xFloat) _d, _d ); _f = XFLOAT_2_FLOAT( _xf); /* return*/ _f; })
81#endif
82
83 //Macros to do constant expansion
84 #define GET_CONSTANT_sd( _p ) _p
85 #define GET_CONSTANT_pd( _p ) { _p, _p }
86
87 //Macros to extend the set of intrinsics to allow for memory operands
88 #define _mm_cmplt_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "cmpltsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
89 #define _mm_cmple_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "cmplesd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
90 #define _mm_cmpgt_sdm( _xd, _m64 ) ({ register xDouble _r; asm( "movsd %1, %0" : "=x" (_r) : "m" (*(_m64)) ); _r = (xDouble) _mm_cmplt_sd( _r, _xd ); /*return*/ _r; })
91 #define _mm_cmpge_sdm( _xd, _m64 ) ({ register xDouble _r; asm( "movsd %1, %0" : "=x" (_r) : "m" (*(_m64)) ); _r = (xDouble) _mm_cmple_sd( _r, _xd ); /*return*/ _r; })
92 #define _mm_cmpeq_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "cmpeqsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
93 #define _mm_cmpne_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "cmpneqsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
94 #define _mm_add_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "addsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
95 #define _mm_sub_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "subsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
96 #define _mm_mul_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "mulsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
97 #define _mm_max_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "maxsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
98 #define _mm_min_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "minsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
99 #define _mm_div_sdm( _xd, _m64 ) ({ register xDouble _r = _xd; asm( "divsd %1, %0" : "+x" (_r) : "m" (*(_m64)) ); /*return*/ _r; })
100 #define _mm_sel_pd( _xd1, _xd2, _mask) ({ xDouble _m = _mask; _m = _mm_or_pd( _mm_and_pd(_xd2, _m ), _mm_andnot_pd( _m, _xd1 ) ); /* return */ _m; })
101
102 #define _mm_ucomieq_sdm( _xd, _m64 ) ({ bool _r; asm("ucomisd %1, %2\n\tsete %b0" : "=q" (_r) : "m" (*(_m64)), "X" (_xd)); /*return */ _r; })
103 #define _mm_ucomieq_ssm( _xf, _m32 ) ({ bool _r; asm("ucomiss %1, %2\n\tsete %b0" : "=q" (_r) : "m" (*(_m32)), "X" (_xf)); /*return */ _r; })
104
105 #define _mm_cmplt_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "cmpltss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
106 #define _mm_cmple_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "cmpless %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
107 #define _mm_cmpgt_ssm( _xf, _m32 ) ({ register xFloat _r; asm( "movss %1, %0" : "=x" (_r) : "m" (*(_m32)) ); _r = (xFloat) __builtin_ia32_cmpltss( _r, _xf ); /*return*/ _r; })
108 #define _mm_cmpge_ssm( _xf, _m32 ) ({ register xFloat _r; asm( "movss %1, %0" : "=x" (_r) : "m" (*(_m32)) ); _r = (xFloat) __builtin_ia32_cmpless( _r, _xf ); /*return*/ _r; })
109 #define _mm_cmpeq_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "cmpeqss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
110 #define _mm_cmpne_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "cmpneqss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
111 #define _mm_add_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "addss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
112 #define _mm_sub_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "subss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
113 #define _mm_mul_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "mulss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
114 #define _mm_max_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "maxss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
115 #define _mm_min_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "minss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
116 #define _mm_div_ssm( _xf, _m32 ) ({ register xFloat _r = _xf; asm( "divss %1, %0" : "+x" (_r) : "m" (*(_m32)) ); /*return*/ _r; })
117 #define _mm_sel_ps( _xf1, _xf2, _mask) ({ xFloat _m = _mask; _m = _mm_or_ps( _mm_and_ps(_xf2, _m ), _mm_andnot_ps( _m, _xf1 ) ); /* return */ _m; })
118
119
120 #define _mm_istrue_sd( _test ) (1 == ( _mm_movemask_pd( _test ) & 1 ))
121 #define _mm_isfalse_sd( _test ) (0 == ( _mm_movemask_pd( _test ) & 1 ))
122
123 #define _mm_istrue_ss( _test ) (1 == ( _mm_movemask_ps( _test ) & 1 ))
124 #define _mm_isfalse_ss( _test ) (0 == ( _mm_movemask_ps( _test ) & 1 ))
125
126 //fast scalb. Doesn't check for overflow
127 #define twoToTheM( _i ) ({ xDouble _xi = (xDouble) _mm_slli_epi64( _mm_cvtsi32_si128( (_i) + 1023 ), 52 ); /*return*/ _xi; })
128
129 //work around for broken _mm_sqrt_sd intrinsic in gcc-4.0
130 #define _MM_SQRT_SD( x ) __builtin_ia32_sqrtsd( x )
131
132 #define _mm_cvtsi64_si128( x ) ({ int64_t _x = x; xSInt64 _r; asm volatile( "movq %1, %0" : "=x" (_r) : "m" (*&_x)); _r })
133
134 #define REQUIRED_ADD_sd( _x, _y ) ({ xDouble __x = _x; xDouble __y = _y; asm volatile( "addsd %1, %0" : "+x" (__x ) : "x" (__y)); /*return*/ __x; })
135 #define REQUIRED_ADD_sdm( _x, _m64 ) ({ xDouble __x = _x; asm volatile( "addsd %1, %0" : "+x" (__x ) : "m" (*(_m64))); /*return*/ __x; })
136 #define REQUIRED_MULTIPLY_sd( _x, _y ) ({ xDouble __x = _x; xDouble __y = _y; asm volatile( "mulsd %1, %0" : "+x" (__x ) : "x" (__y)); /*return*/ __x; })
137 #define REQUIRED_MULTIPLY_sdm( _x, _m64 ) ({ xDouble __x = _x; xDouble __y = _y; asm volatile( "mulsd %1, %0" : "+x" (__x ) : "m" (*(_m64))); /*return*/ __x; })
138 #define REQUIRED_DIVIDE_sd( _x, _y ) ({ xDouble __x = _x; xDouble __y = _y; asm volatile( "divsd %1, %0" : "+x" (__x ) : "x" (__y)); /*return*/ __x; })
139 #define REQUIRED_DIVIDE_sdm( _x, _m64 ) ({ xDouble __x = _x; asm volatile( "divsd %1, %0" : "+x" (__x ) : "m" (*(_m64))); /*return*/ __x; })
140
141 #define REQUIRED_ADD_ss( _x, _y ) ({ xFloat __x = _x; xFloat __y = _y; asm volatile( "addss %1, %0" : "+x" (__x ) : "x" (__y)); /*return*/ __x; })
142 #define REQUIRED_ADD_ssm( _x, _m32 ) ({ xFloat __x = _x; asm volatile( "addss %1, %0" : "+x" (__x ) : "m" (*(_m32))); /*return*/ __x; })
143 #define REQUIRED_MULTIPLY_ss( _x, _y ) ({ xFloat __x = _x; xFloat __y = _y; asm volatile( "mulss %1, %0" : "+x" (__x ) : "x" (__y)); /*return*/ __x; })
144 #define REQUIRED_MULTIPLY_ssm( _x, _m32 ) ({ xFloat __x = _x; xFloat __y = _y; asm volatile( "mulss %1, %0" : "+x" (__x ) : "m" (*(_m32))); /*return*/ __x; })
145 #define REQUIRED_DIVIDE_ss( _x, _y ) ({ xFloat __x = _x; xFloat __y = _y; asm volatile( "divss %1, %0" : "+x" (__x ) : "x" (__y)); /*return*/ __x; })
146 #define REQUIRED_DIVIDE_ssm( _x, _m32 ) ({ xFloat __x = _x; asm volatile( "divss %1, %0" : "+x" (__x ) : "m" (*(_m32))); /*return*/ __x; })
147
148
149 #define SET_INVALID_FLAG() { __m128d __x = _mm_setzero_pd(); __asm__ __volatile__( "pcmpeqd %0, %0 \n\t cmpltsd %0, %0" : "+x" (__x) ); }
150
151#if defined( __SSE3__ )
152 #define CVTTLD_SI64( _ld ) ({ long double _x = _ld; int64_t _r; asm( "fldt %1 \n\t fisttpll %0" : "=m" (*&_r): "m" (*&_x) ); /*return*/ _r; })
153#else
154 #warning SSE3 disabled
155 static inline int64_t CVTTLD_SI64( long double ld ){ return (int64_t) ld; }
156#endif
157 #define CVTLD_SI64( _ld ) ({ long double _x = _ld; int64_t _r; asm( "fldt %1 \n\t fistpll %0" : "=m" (*&_r): "m" (*&_x) ); /*return*/ _r; })
158
159 #define EXPECT_TRUE( _a ) __builtin_expect( (_a), 1 )
160 #define EXPECT_FALSE( _a ) __builtin_expect( (_a), 0 )
161
162
163#endif
164
165#if defined( XMMLIBM_DEBUG )
166 #warning always inline disabled
167 #define ALWAYS_INLINE
168#else
169 #define ALWAYS_INLINE __attribute__ ((always_inline))
170#endif
171
172#endif /* __XMMLIBM_PREFIX_H__ */