15821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//===-- lib/mulsf3.c - Single-precision multiplication ------------*- C -*-===// 25821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 35821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// The LLVM Compiler Infrastructure 45821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 55821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// This file is dual licensed under the MIT and the University of Illinois Open 65821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// Source Licenses. See LICENSE.TXT for details. 75821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 85821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//===----------------------------------------------------------------------===// 95821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 105821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// This file implements single-precision soft-float multiplication 115821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// with the IEEE-754 default rounding (to nearest, ties to even). 125821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// 135821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//===----------------------------------------------------------------------===// 145821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 155821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#define SINGLE_PRECISION 165821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "fp_lib.h" 175821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 185821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)ARM_EABI_FNALIAS(fmul, mulsf3) 195821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 205821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)COMPILER_RT_ABI fp_t 215821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)__mulsf3(fp_t a, fp_t b) { 225821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 235821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const unsigned int aExponent = toRep(a) >> significandBits & maxExponent; 245821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const unsigned int bExponent = toRep(b) >> significandBits & maxExponent; 255821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const rep_t productSign = (toRep(a) ^ toRep(b)) & signBit; 265821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 275821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) rep_t aSignificand = toRep(a) & significandMask; 285821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) rep_t bSignificand = toRep(b) & significandMask; 295821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) int scale = 0; 305821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 315821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // Detect if a or b is zero, denormal, infinity, or NaN. 325821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (aExponent-1U >= maxExponent-1U || bExponent-1U >= maxExponent-1U) { 335821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 345821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const rep_t aAbs = toRep(a) & absMask; 355821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) const rep_t bAbs = toRep(b) & absMask; 365821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 375821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // NaN * anything = qNaN 385821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (aAbs > infRep) return fromRep(toRep(a) | quietBit); 395821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // anything * NaN = qNaN 405821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (bAbs > infRep) return fromRep(toRep(b) | quietBit); 415821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 425821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (aAbs == infRep) { 435821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // infinity * non-zero = +/- infinity 445821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (bAbs) return fromRep(aAbs | productSign); 455821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // infinity * zero = NaN 465821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) else return fromRep(qnanRep); 475821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) } 485821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 495821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (bAbs == infRep) { 505821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // non-zero * infinity = +/- infinity 515821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (aAbs) return fromRep(bAbs | productSign); 525821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // zero * infinity = NaN 535821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) else return fromRep(qnanRep); 545821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) } 555821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 565821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // zero * anything = +/- zero 575821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (!aAbs) return fromRep(productSign); 585821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) // anything * zero = +/- zero 595821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) if (!bAbs) return fromRep(productSign); 605821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles) 61 // one or both of a or b is denormal, the other (if applicable) is a 62 // normal number. Renormalize one or both of a and b, and set scale to 63 // include the necessary exponent adjustment. 64 if (aAbs < implicitBit) scale += normalize(&aSignificand); 65 if (bAbs < implicitBit) scale += normalize(&bSignificand); 66 } 67 68 // Or in the implicit significand bit. (If we fell through from the 69 // denormal path it was already set by normalize( ), but setting it twice 70 // won't hurt anything.) 71 aSignificand |= implicitBit; 72 bSignificand |= implicitBit; 73 74 // Get the significand of a*b. Before multiplying the significands, shift 75 // one of them left to left-align it in the field. Thus, the product will 76 // have (exponentBits + 2) integral digits, all but two of which must be 77 // zero. Normalizing this result is just a conditional left-shift by one 78 // and bumping the exponent accordingly. 79 rep_t productHi, productLo; 80 wideMultiply(aSignificand, bSignificand << exponentBits, 81 &productHi, &productLo); 82 83 int productExponent = aExponent + bExponent - exponentBias + scale; 84 85 // Normalize the significand, adjust exponent if needed. 86 if (productHi & implicitBit) productExponent++; 87 else wideLeftShift(&productHi, &productLo, 1); 88 89 // If we have overflowed the type, return +/- infinity. 90 if (productExponent >= maxExponent) return fromRep(infRep | productSign); 91 92 if (productExponent <= 0) { 93 // Result is denormal before rounding, the exponent is zero and we 94 // need to shift the significand. 95 wideRightShiftWithSticky(&productHi, &productLo, 1U - (unsigned)productExponent); 96 } 97 98 else { 99 // Result is normal before rounding; insert the exponent. 100 productHi &= significandMask; 101 productHi |= (rep_t)productExponent << significandBits; 102 } 103 104 // Insert the sign of the result: 105 productHi |= productSign; 106 107 // Final rounding. The final result may overflow to infinity, or underflow 108 // to zero, but those are the correct results in those cases. 109 if (productLo > signBit) productHi++; 110 if (productLo == signBit) productHi += productHi & 1; 111 return fromRep(productHi); 112} 113