1 //===-- lib/fp_lib.h - Floating-point utilities -------------------*- C -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is dual licensed under the MIT and the University of Illinois Open
6 // Source Licenses. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file is a configuration header for soft-float routines in compiler-rt.
11 // This file does not provide any part of the compiler-rt interface, but defines
12 // many useful constants and utility routines that are used in the
13 // implementation of the soft-float routines in compiler-rt.
15 // Assumes that float and double correspond to the IEEE-754 binary32 and
16 // binary64 types, respectively, and that integer endianness matches floating
17 // point endianness on the target platform.
19 //===----------------------------------------------------------------------===//
29 #if defined SINGLE_PRECISION
31 typedef uint32_t rep_t
;
32 typedef int32_t srep_t
;
34 #define REP_C UINT32_C
35 #define significandBits 23
37 static inline int rep_clz(rep_t a
) {
38 return __builtin_clz(a
);
41 // 32x32 --> 64 bit multiply
42 static inline void wideMultiply(rep_t a
, rep_t b
, rep_t
*hi
, rep_t
*lo
) {
43 const uint64_t product
= (uint64_t)a
*b
;
48 #elif defined DOUBLE_PRECISION
50 typedef uint64_t rep_t
;
51 typedef int64_t srep_t
;
53 #define REP_C UINT64_C
54 #define significandBits 52
56 static inline int rep_clz(rep_t a
) {
58 return __builtin_clzl(a
);
60 if (a
& REP_C(0xffffffff00000000))
61 return __builtin_clz(a
>> 32);
63 return 32 + __builtin_clz(a
& REP_C(0xffffffff));
67 #define loWord(a) (a & 0xffffffffU)
68 #define hiWord(a) (a >> 32)
70 // 64x64 -> 128 wide multiply for platforms that don't have such an operation;
71 // many 64-bit platforms have this operation, but they tend to have hardware
72 // floating-point, so we don't bother with a special case for them here.
73 static inline void wideMultiply(rep_t a
, rep_t b
, rep_t
*hi
, rep_t
*lo
) {
74 // Each of the component 32x32 -> 64 products
75 const uint64_t plolo
= loWord(a
) * loWord(b
);
76 const uint64_t plohi
= loWord(a
) * hiWord(b
);
77 const uint64_t philo
= hiWord(a
) * loWord(b
);
78 const uint64_t phihi
= hiWord(a
) * hiWord(b
);
79 // Sum terms that contribute to lo in a way that allows us to get the carry
80 const uint64_t r0
= loWord(plolo
);
81 const uint64_t r1
= hiWord(plolo
) + loWord(plohi
) + loWord(philo
);
82 *lo
= r0
+ (r1
<< 32);
83 // Sum terms contributing to hi with the carry from lo
84 *hi
= hiWord(plohi
) + hiWord(philo
) + hiWord(r1
) + phihi
;
88 #error Either SINGLE_PRECISION or DOUBLE_PRECISION must be defined.
91 #define typeWidth (sizeof(rep_t)*CHAR_BIT)
92 #define exponentBits (typeWidth - significandBits - 1)
93 #define maxExponent ((1 << exponentBits) - 1)
94 #define exponentBias (maxExponent >> 1)
96 #define implicitBit (REP_C(1) << significandBits)
97 #define significandMask (implicitBit - 1U)
98 #define signBit (REP_C(1) << (significandBits + exponentBits))
99 #define absMask (signBit - 1U)
100 #define exponentMask (absMask ^ significandMask)
101 #define oneRep ((rep_t)exponentBias << significandBits)
102 #define infRep exponentMask
103 #define quietBit (implicitBit >> 1)
104 #define qnanRep (exponentMask | quietBit)
106 static inline rep_t
toRep(fp_t x
) {
107 const union { fp_t f
; rep_t i
; } rep
= {.f
= x
};
111 static inline fp_t
fromRep(rep_t x
) {
112 const union { fp_t f
; rep_t i
; } rep
= {.i
= x
};
116 static inline int normalize(rep_t
*significand
) {
117 const int shift
= rep_clz(*significand
) - rep_clz(implicitBit
);
118 *significand
<<= shift
;
122 static inline void wideLeftShift(rep_t
*hi
, rep_t
*lo
, int count
) {
123 *hi
= *hi
<< count
| *lo
>> (typeWidth
- count
);
127 static inline void wideRightShiftWithSticky(rep_t
*hi
, rep_t
*lo
, unsigned int count
) {
128 if (count
< typeWidth
) {
129 const bool sticky
= *lo
<< (typeWidth
- count
);
130 *lo
= *hi
<< (typeWidth
- count
) | *lo
>> count
| sticky
;
133 else if (count
< 2*typeWidth
) {
134 const bool sticky
= *hi
<< (2*typeWidth
- count
) | *lo
;
135 *lo
= *hi
>> (count
- typeWidth
) | sticky
;
138 const bool sticky
= *hi
| *lo
;
144 #endif // FP_LIB_HEADER