| 1 | //===----- lib/fp_add_impl.inc - floaing point addition -----------*- C -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file implements soft-float addition with the IEEE-754 default rounding |
| 10 | // (to nearest, ties to even). |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "fp_lib.h" |
| 15 | #include "fp_mode.h" |
| 16 | |
| 17 | static __inline fp_t __addXf3__(fp_t a, fp_t b) { |
| 18 | rep_t aRep = toRep(x: a); |
| 19 | rep_t bRep = toRep(x: b); |
| 20 | const rep_t aAbs = aRep & absMask; |
| 21 | const rep_t bAbs = bRep & absMask; |
| 22 | |
| 23 | // Detect if a or b is zero, infinity, or NaN. |
| 24 | if (aAbs - REP_C(1) >= infRep - REP_C(1) || |
| 25 | bAbs - REP_C(1) >= infRep - REP_C(1)) { |
| 26 | // NaN + anything = qNaN |
| 27 | if (aAbs > infRep) |
| 28 | return fromRep(x: toRep(x: a) | quietBit); |
| 29 | // anything + NaN = qNaN |
| 30 | if (bAbs > infRep) |
| 31 | return fromRep(x: toRep(x: b) | quietBit); |
| 32 | |
| 33 | if (aAbs == infRep) { |
| 34 | // +/-infinity + -/+infinity = qNaN |
| 35 | if ((toRep(x: a) ^ toRep(x: b)) == signBit) |
| 36 | return fromRep(qnanRep); |
| 37 | // +/-infinity + anything remaining = +/- infinity |
| 38 | else |
| 39 | return a; |
| 40 | } |
| 41 | |
| 42 | // anything remaining + +/-infinity = +/-infinity |
| 43 | if (bAbs == infRep) |
| 44 | return b; |
| 45 | |
| 46 | // zero + anything = anything |
| 47 | if (!aAbs) { |
| 48 | // We need to get the sign right for zero + zero. |
| 49 | if (!bAbs) |
| 50 | return fromRep(x: toRep(x: a) & toRep(x: b)); |
| 51 | else |
| 52 | return b; |
| 53 | } |
| 54 | |
| 55 | // anything + zero = anything |
| 56 | if (!bAbs) |
| 57 | return a; |
| 58 | } |
| 59 | |
| 60 | // Swap a and b if necessary so that a has the larger absolute value. |
| 61 | if (bAbs > aAbs) { |
| 62 | const rep_t temp = aRep; |
| 63 | aRep = bRep; |
| 64 | bRep = temp; |
| 65 | } |
| 66 | |
| 67 | // Extract the exponent and significand from the (possibly swapped) a and b. |
| 68 | int aExponent = aRep >> significandBits & maxExponent; |
| 69 | int bExponent = bRep >> significandBits & maxExponent; |
| 70 | rep_t aSignificand = aRep & significandMask; |
| 71 | rep_t bSignificand = bRep & significandMask; |
| 72 | |
| 73 | // Normalize any denormals, and adjust the exponent accordingly. |
| 74 | if (aExponent == 0) |
| 75 | aExponent = normalize(significand: &aSignificand); |
| 76 | if (bExponent == 0) |
| 77 | bExponent = normalize(significand: &bSignificand); |
| 78 | |
| 79 | // The sign of the result is the sign of the larger operand, a. If they |
| 80 | // have opposite signs, we are performing a subtraction. Otherwise, we |
| 81 | // perform addition. |
| 82 | const rep_t resultSign = aRep & signBit; |
| 83 | const bool subtraction = (aRep ^ bRep) & signBit; |
| 84 | |
| 85 | // Shift the significands to give us round, guard and sticky, and set the |
| 86 | // implicit significand bit. If we fell through from the denormal path it |
| 87 | // was already set by normalize( ), but setting it twice won't hurt |
| 88 | // anything. |
| 89 | aSignificand = (aSignificand | implicitBit) << 3; |
| 90 | bSignificand = (bSignificand | implicitBit) << 3; |
| 91 | |
| 92 | // Shift the significand of b by the difference in exponents, with a sticky |
| 93 | // bottom bit to get rounding correct. |
| 94 | const unsigned int align = (unsigned int)(aExponent - bExponent); |
| 95 | if (align) { |
| 96 | if (align < typeWidth) { |
| 97 | const bool sticky = (bSignificand << (typeWidth - align)) != 0; |
| 98 | bSignificand = bSignificand >> align | sticky; |
| 99 | } else { |
| 100 | bSignificand = 1; // Set the sticky bit. b is known to be non-zero. |
| 101 | } |
| 102 | } |
| 103 | if (subtraction) { |
| 104 | aSignificand -= bSignificand; |
| 105 | // If a == -b, return +zero. |
| 106 | if (aSignificand == 0) |
| 107 | return fromRep(x: 0); |
| 108 | |
| 109 | // If partial cancellation occured, we need to left-shift the result |
| 110 | // and adjust the exponent. |
| 111 | if (aSignificand < implicitBit << 3) { |
| 112 | const int shift = rep_clz(a: aSignificand) - rep_clz(implicitBit << 3); |
| 113 | aSignificand <<= shift; |
| 114 | aExponent -= shift; |
| 115 | } |
| 116 | } else /* addition */ { |
| 117 | aSignificand += bSignificand; |
| 118 | |
| 119 | // If the addition carried up, we need to right-shift the result and |
| 120 | // adjust the exponent. |
| 121 | if (aSignificand & implicitBit << 4) { |
| 122 | const bool sticky = aSignificand & 1; |
| 123 | aSignificand = aSignificand >> 1 | sticky; |
| 124 | aExponent += 1; |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | // If we have overflowed the type, return +/- infinity. |
| 129 | if (aExponent >= maxExponent) |
| 130 | return fromRep(infRep | resultSign); |
| 131 | |
| 132 | if (aExponent <= 0) { |
| 133 | // The result is denormal before rounding. The exponent is zero and we |
| 134 | // need to shift the significand. |
| 135 | const int shift = 1 - aExponent; |
| 136 | const bool sticky = (aSignificand << (typeWidth - shift)) != 0; |
| 137 | aSignificand = aSignificand >> shift | sticky; |
| 138 | aExponent = 0; |
| 139 | } |
| 140 | |
| 141 | // Low three bits are round, guard, and sticky. |
| 142 | const int roundGuardSticky = aSignificand & 0x7; |
| 143 | |
| 144 | // Shift the significand into place, and mask off the implicit bit. |
| 145 | rep_t result = aSignificand >> 3 & significandMask; |
| 146 | |
| 147 | // Insert the exponent and sign. |
| 148 | result |= (rep_t)aExponent << significandBits; |
| 149 | result |= resultSign; |
| 150 | |
| 151 | // Perform the final rounding. The result may overflow to infinity, but |
| 152 | // that is the correct result in that case. |
| 153 | switch (__fe_getround()) { |
| 154 | case CRT_FE_TONEAREST: |
| 155 | if (roundGuardSticky > 0x4) |
| 156 | result++; |
| 157 | if (roundGuardSticky == 0x4) |
| 158 | result += result & 1; |
| 159 | break; |
| 160 | case CRT_FE_DOWNWARD: |
| 161 | if (resultSign && roundGuardSticky) result++; |
| 162 | break; |
| 163 | case CRT_FE_UPWARD: |
| 164 | if (!resultSign && roundGuardSticky) result++; |
| 165 | break; |
| 166 | case CRT_FE_TOWARDZERO: |
| 167 | break; |
| 168 | } |
| 169 | if (roundGuardSticky) |
| 170 | __fe_raise_inexact(); |
| 171 | return fromRep(x: result); |
| 172 | } |
| 173 | |