divsf3.c revision 214152
1214152Sed//===-- lib/divsf3.c - Single-precision division ------------------*- C -*-===// 2214152Sed// 3214152Sed// The LLVM Compiler Infrastructure 4214152Sed// 5214152Sed// This file is distributed under the University of Illinois Open Source 6214152Sed// License. See LICENSE.TXT for details. 7214152Sed// 8214152Sed//===----------------------------------------------------------------------===// 9214152Sed// 10214152Sed// This file implements single-precision soft-float division 11214152Sed// with the IEEE-754 default rounding (to nearest, ties to even). 12214152Sed// 13214152Sed// For simplicity, this implementation currently flushes denormals to zero. 14214152Sed// It should be a fairly straightforward exercise to implement gradual 15214152Sed// underflow with correct rounding. 16214152Sed// 17214152Sed//===----------------------------------------------------------------------===// 18214152Sed 19214152Sed#define SINGLE_PRECISION 20214152Sed#include "fp_lib.h" 21214152Sed 22214152Sedfp_t __divsf3(fp_t a, fp_t b) { 23214152Sed 24214152Sed const unsigned int aExponent = toRep(a) >> significandBits & maxExponent; 25214152Sed const unsigned int bExponent = toRep(b) >> significandBits & maxExponent; 26214152Sed const rep_t quotientSign = (toRep(a) ^ toRep(b)) & signBit; 27214152Sed 28214152Sed rep_t aSignificand = toRep(a) & significandMask; 29214152Sed rep_t bSignificand = toRep(b) & significandMask; 30214152Sed int scale = 0; 31214152Sed 32214152Sed // Detect if a or b is zero, denormal, infinity, or NaN. 33214152Sed if (aExponent-1U >= maxExponent-1U || bExponent-1U >= maxExponent-1U) { 34214152Sed 35214152Sed const rep_t aAbs = toRep(a) & absMask; 36214152Sed const rep_t bAbs = toRep(b) & absMask; 37214152Sed 38214152Sed // NaN / anything = qNaN 39214152Sed if (aAbs > infRep) return fromRep(toRep(a) | quietBit); 40214152Sed // anything / NaN = qNaN 41214152Sed if (bAbs > infRep) return fromRep(toRep(b) | quietBit); 42214152Sed 43214152Sed if (aAbs == infRep) { 44214152Sed // infinity / infinity = NaN 45214152Sed if (bAbs == infRep) return fromRep(qnanRep); 46214152Sed // infinity / anything else = +/- infinity 47214152Sed else return fromRep(aAbs | quotientSign); 48214152Sed } 49214152Sed 50214152Sed // anything else / infinity = +/- 0 51214152Sed if (bAbs == infRep) return fromRep(quotientSign); 52214152Sed 53214152Sed if (!aAbs) { 54214152Sed // zero / zero = NaN 55214152Sed if (!bAbs) return fromRep(qnanRep); 56214152Sed // zero / anything else = +/- zero 57214152Sed else return fromRep(quotientSign); 58214152Sed } 59214152Sed // anything else / zero = +/- infinity 60214152Sed if (!bAbs) return fromRep(infRep | quotientSign); 61214152Sed 62214152Sed // one or both of a or b is denormal, the other (if applicable) is a 63214152Sed // normal number. Renormalize one or both of a and b, and set scale to 64214152Sed // include the necessary exponent adjustment. 65214152Sed if (aAbs < implicitBit) scale += normalize(&aSignificand); 66214152Sed if (bAbs < implicitBit) scale -= normalize(&bSignificand); 67214152Sed } 68214152Sed 69214152Sed // Or in the implicit significand bit. (If we fell through from the 70214152Sed // denormal path it was already set by normalize( ), but setting it twice 71214152Sed // won't hurt anything.) 72214152Sed aSignificand |= implicitBit; 73214152Sed bSignificand |= implicitBit; 74214152Sed int quotientExponent = aExponent - bExponent + scale; 75214152Sed 76214152Sed // Align the significand of b as a Q31 fixed-point number in the range 77214152Sed // [1, 2.0) and get a Q32 approximate reciprocal using a small minimax 78214152Sed // polynomial approximation: reciprocal = 3/4 + 1/sqrt(2) - b/2. This 79214152Sed // is accurate to about 3.5 binary digits. 80214152Sed uint32_t q31b = bSignificand << 8; 81214152Sed uint32_t reciprocal = UINT32_C(0x7504f333) - q31b; 82214152Sed 83214152Sed // Now refine the reciprocal estimate using a Newton-Raphson iteration: 84214152Sed // 85214152Sed // x1 = x0 * (2 - x0 * b) 86214152Sed // 87214152Sed // This doubles the number of correct binary digits in the approximation 88214152Sed // with each iteration, so after three iterations, we have about 28 binary 89214152Sed // digits of accuracy. 90214152Sed uint32_t correction; 91214152Sed correction = -((uint64_t)reciprocal * q31b >> 32); 92214152Sed reciprocal = (uint64_t)reciprocal * correction >> 31; 93214152Sed correction = -((uint64_t)reciprocal * q31b >> 32); 94214152Sed reciprocal = (uint64_t)reciprocal * correction >> 31; 95214152Sed correction = -((uint64_t)reciprocal * q31b >> 32); 96214152Sed reciprocal = (uint64_t)reciprocal * correction >> 31; 97214152Sed 98214152Sed // Exhaustive testing shows that the error in reciprocal after three steps 99214152Sed // is in the interval [-0x1.f58108p-31, 0x1.d0e48cp-29], in line with our 100214152Sed // expectations. We bump the reciprocal by a tiny value to force the error 101214152Sed // to be strictly positive (in the range [0x1.4fdfp-37,0x1.287246p-29], to 102214152Sed // be specific). This also causes 1/1 to give a sensible approximation 103214152Sed // instead of zero (due to overflow). 104214152Sed reciprocal -= 2; 105214152Sed 106214152Sed // The numerical reciprocal is accurate to within 2^-28, lies in the 107214152Sed // interval [0x1.000000eep-1, 0x1.fffffffcp-1], and is strictly smaller 108214152Sed // than the true reciprocal of b. Multiplying a by this reciprocal thus 109214152Sed // gives a numerical q = a/b in Q24 with the following properties: 110214152Sed // 111214152Sed // 1. q < a/b 112214152Sed // 2. q is in the interval [0x1.000000eep-1, 0x1.fffffffcp0) 113214152Sed // 3. the error in q is at most 2^-24 + 2^-27 -- the 2^24 term comes 114214152Sed // from the fact that we truncate the product, and the 2^27 term 115214152Sed // is the error in the reciprocal of b scaled by the maximum 116214152Sed // possible value of a. As a consequence of this error bound, 117214152Sed // either q or nextafter(q) is the correctly rounded 118214152Sed rep_t quotient = (uint64_t)reciprocal*(aSignificand << 1) >> 32; 119214152Sed 120214152Sed // Two cases: quotient is in [0.5, 1.0) or quotient is in [1.0, 2.0). 121214152Sed // In either case, we are going to compute a residual of the form 122214152Sed // 123214152Sed // r = a - q*b 124214152Sed // 125214152Sed // We know from the construction of q that r satisfies: 126214152Sed // 127214152Sed // 0 <= r < ulp(q)*b 128214152Sed // 129214152Sed // if r is greater than 1/2 ulp(q)*b, then q rounds up. Otherwise, we 130214152Sed // already have the correct result. The exact halfway case cannot occur. 131214152Sed // We also take this time to right shift quotient if it falls in the [1,2) 132214152Sed // range and adjust the exponent accordingly. 133214152Sed rep_t residual; 134214152Sed if (quotient < (implicitBit << 1)) { 135214152Sed residual = (aSignificand << 24) - quotient * bSignificand; 136214152Sed quotientExponent--; 137214152Sed } else { 138214152Sed quotient >>= 1; 139214152Sed residual = (aSignificand << 23) - quotient * bSignificand; 140214152Sed } 141214152Sed 142214152Sed const int writtenExponent = quotientExponent + exponentBias; 143214152Sed 144214152Sed if (writtenExponent >= maxExponent) { 145214152Sed // If we have overflowed the exponent, return infinity. 146214152Sed return fromRep(infRep | quotientSign); 147214152Sed } 148214152Sed 149214152Sed else if (writtenExponent < 1) { 150214152Sed // Flush denormals to zero. In the future, it would be nice to add 151214152Sed // code to round them correctly. 152214152Sed return fromRep(quotientSign); 153214152Sed } 154214152Sed 155214152Sed else { 156214152Sed const bool round = (residual << 1) > bSignificand; 157214152Sed // Clear the implicit bit 158214152Sed rep_t absResult = quotient & significandMask; 159214152Sed // Insert the exponent 160214152Sed absResult |= (rep_t)writtenExponent << significandBits; 161214152Sed // Round 162214152Sed absResult += round; 163214152Sed // Insert the sign and return 164214152Sed return fromRep(absResult | quotientSign); 165214152Sed } 166214152Sed} 167