fp_lib.h revision 9ad441ffec97db647fee3725b3424284fb913e14
1//===-- lib/fp_lib.h - Floating-point utilities -------------------*- C -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is dual licensed under the MIT and the University of Illinois Open 6// Source Licenses. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 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. 14// 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. 18// 19//===----------------------------------------------------------------------===// 20 21#ifndef FP_LIB_HEADER 22#define FP_LIB_HEADER 23 24#include <stdint.h> 25#include <stdbool.h> 26#include <limits.h> 27 28#if defined SINGLE_PRECISION 29 30typedef uint32_t rep_t; 31typedef int32_t srep_t; 32typedef float fp_t; 33#define REP_C UINT32_C 34#define significandBits 23 35 36static inline int rep_clz(rep_t a) { 37 return __builtin_clz(a); 38} 39 40// 32x32 --> 64 bit multiply 41static inline void wideMultiply(rep_t a, rep_t b, rep_t *hi, rep_t *lo) { 42 const uint64_t product = (uint64_t)a*b; 43 *hi = product >> 32; 44 *lo = product; 45} 46 47#elif defined DOUBLE_PRECISION 48 49typedef uint64_t rep_t; 50typedef int64_t srep_t; 51typedef double fp_t; 52#define REP_C UINT64_C 53#define significandBits 52 54 55static inline int rep_clz(rep_t a) { 56#if defined __LP64__ 57 return __builtin_clzl(a); 58#else 59 if (a & REP_C(0xffffffff00000000)) 60 return __builtin_clz(a >> 32); 61 else 62 return 32 + __builtin_clz(a & REP_C(0xffffffff)); 63#endif 64} 65 66#define loWord(a) (a & 0xffffffffU) 67#define hiWord(a) (a >> 32) 68 69// 64x64 -> 128 wide multiply for platforms that don't have such an operation; 70// many 64-bit platforms have this operation, but they tend to have hardware 71// floating-point, so we don't bother with a special case for them here. 72static inline void wideMultiply(rep_t a, rep_t b, rep_t *hi, rep_t *lo) { 73 // Each of the component 32x32 -> 64 products 74 const uint64_t plolo = loWord(a) * loWord(b); 75 const uint64_t plohi = loWord(a) * hiWord(b); 76 const uint64_t philo = hiWord(a) * loWord(b); 77 const uint64_t phihi = hiWord(a) * hiWord(b); 78 // Sum terms that contribute to lo in a way that allows us to get the carry 79 const uint64_t r0 = loWord(plolo); 80 const uint64_t r1 = hiWord(plolo) + loWord(plohi) + loWord(philo); 81 *lo = r0 + (r1 << 32); 82 // Sum terms contributing to hi with the carry from lo 83 *hi = hiWord(plohi) + hiWord(philo) + hiWord(r1) + phihi; 84} 85 86#else 87#error Either SINGLE_PRECISION or DOUBLE_PRECISION must be defined. 88#endif 89 90#define typeWidth (sizeof(rep_t)*CHAR_BIT) 91#define exponentBits (typeWidth - significandBits - 1) 92#define maxExponent ((1 << exponentBits) - 1) 93#define exponentBias (maxExponent >> 1) 94 95#define implicitBit (REP_C(1) << significandBits) 96#define significandMask (implicitBit - 1U) 97#define signBit (REP_C(1) << (significandBits + exponentBits)) 98#define absMask (signBit - 1U) 99#define exponentMask (absMask ^ significandMask) 100#define oneRep ((rep_t)exponentBias << significandBits) 101#define infRep exponentMask 102#define quietBit (implicitBit >> 1) 103#define qnanRep (exponentMask | quietBit) 104 105static inline rep_t toRep(fp_t x) { 106 const union { fp_t f; rep_t i; } rep = {.f = x}; 107 return rep.i; 108} 109 110static inline fp_t fromRep(rep_t x) { 111 const union { fp_t f; rep_t i; } rep = {.i = x}; 112 return rep.f; 113} 114 115static inline int normalize(rep_t *significand) { 116 const int shift = rep_clz(*significand) - rep_clz(implicitBit); 117 *significand <<= shift; 118 return 1 - shift; 119} 120 121static inline void wideLeftShift(rep_t *hi, rep_t *lo, int count) { 122 *hi = *hi << count | *lo >> (typeWidth - count); 123 *lo = *lo << count; 124} 125 126static inline void wideRightShiftWithSticky(rep_t *hi, rep_t *lo, int count) { 127 if (count < typeWidth) { 128 const bool sticky = *lo << (typeWidth - count); 129 *lo = *hi << (typeWidth - count) | *lo >> count | sticky; 130 *hi = *hi >> count; 131 } 132 else if (count < 2*typeWidth) { 133 const bool sticky = *hi << (2*typeWidth - count) | *lo; 134 *lo = *hi >> (count - typeWidth) | sticky; 135 *hi = 0; 136 } else { 137 const bool sticky = *hi | *lo; 138 *lo = sticky; 139 *hi = 0; 140 } 141} 142 143#endif // FP_LIB_HEADER 144