fmpyfadd.c

这个linux源代码是很全面的~基本完整了~使用c编译的~由于时间问题我没有亲自测试~但就算用来做参考资料也是非常好的

	/*
	 * Now we are ready to perform the add portion of the operation.
	 *
	 * The exponents need to be kept as integers for now, since the
	 * multiply result might not fit into the exponent field.  We
	 * can't overflow or underflow because of this yet, since the
	 * add could bring the final result back into range.
	 */
	add_exponent = Sgl_exponent(opnd3);

	/*
	 * Check for denormalized or zero add operand.
	 */
	if (add_exponent == 0) {
		/* check for zero */
		if (Sgl_iszero_mantissa(opnd3)) {
			/* right is zero */
			/* Left can't be zero and must be result.
			 *
			 * The final result is now in tmpres and mpy_exponent,
			 * and needs to be rounded and squeezed back into
			 * double precision format from double extended.
			 */
			result_exponent = mpy_exponent;
			Sglext_copy(tmpresp1,tmpresp2,resultp1,resultp2);
			sign_save = Sgl_signextendedsign(resultp1);/*save sign*/
			goto round;
		}

		/* 
		 * Neither are zeroes.  
		 * Adjust exponent and normalize add operand.
		 */
		sign_save = Sgl_signextendedsign(opnd3);	/* save sign */
		Sgl_clear_signexponent(opnd3);
		Sgl_leftshiftby1(opnd3);
		Sgl_normalize(opnd3,add_exponent);
		Sgl_set_sign(opnd3,sign_save);		/* restore sign */
	} else {
		Sgl_clear_exponent_set_hidden(opnd3);
	}
	/*
	 * Copy opnd3 to the double extended variable called right.
	 */
	Sgl_copyto_sglext(opnd3,rightp1,rightp2);

	/*
	 * A zero "save" helps discover equal operands (for later),
	 * and is used in swapping operands (if needed).
	 */
	Sglext_xortointp1(tmpresp1,rightp1,/*to*/save);

	/*
	 * Compare magnitude of operands.
	 */
	Sglext_copytoint_exponentmantissa(tmpresp1,signlessleft1);
	Sglext_copytoint_exponentmantissa(rightp1,signlessright1);
	if (mpy_exponent < add_exponent || mpy_exponent == add_exponent &&
	    Sglext_ismagnitudeless(signlessleft1,signlessright1)) {
		/*
		 * Set the left operand to the larger one by XOR swap.
		 * First finish the first word "save".
		 */
		Sglext_xorfromintp1(save,rightp1,/*to*/rightp1);
		Sglext_xorfromintp1(save,tmpresp1,/*to*/tmpresp1);
		Sglext_swap_lower(tmpresp2,rightp2);
		/* also setup exponents used in rest of routine */
		diff_exponent = add_exponent - mpy_exponent;
		result_exponent = add_exponent;
	} else {
		/* also setup exponents used in rest of routine */
		diff_exponent = mpy_exponent - add_exponent;
		result_exponent = mpy_exponent;
	}
	/* Invariant: left is not smaller than right. */

	/*
	 * Special case alignment of operands that would force alignment
	 * beyond the extent of the extension.  A further optimization
	 * could special case this but only reduces the path length for
	 * this infrequent case.
	 */
	if (diff_exponent > SGLEXT_THRESHOLD) {
		diff_exponent = SGLEXT_THRESHOLD;
	}

	/* Align right operand by shifting it to the right */
	Sglext_clear_sign(rightp1);
	Sglext_right_align(rightp1,rightp2,/*shifted by*/diff_exponent);
	
	/* Treat sum and difference of the operands separately. */
	if ((int)save < 0) {
		/*
		 * Difference of the two operands.  Overflow can occur if the
		 * multiply overflowed.  A borrow can occur out of the hidden
		 * bit and force a post normalization phase.
		 */
		Sglext_subtract(tmpresp1,tmpresp2, rightp1,rightp2,
			resultp1,resultp2);
		sign_save = Sgl_signextendedsign(resultp1);
		if (Sgl_iszero_hidden(resultp1)) {
			/* Handle normalization */
		/* A straight foward algorithm would now shift the
		 * result and extension left until the hidden bit
		 * becomes one.  Not all of the extension bits need
		 * participate in the shift.  Only the two most 
		 * significant bits (round and guard) are needed.
		 * If only a single shift is needed then the guard
		 * bit becomes a significant low order bit and the
		 * extension must participate in the rounding.
		 * If more than a single shift is needed, then all
		 * bits to the right of the guard bit are zeros, 
		 * and the guard bit may or may not be zero. */
			Sglext_leftshiftby1(resultp1,resultp2);

			/* Need to check for a zero result.  The sign and
			 * exponent fields have already been zeroed.  The more
			 * efficient test of the full object can be used.
			 */
			 if (Sglext_iszero(resultp1,resultp2)) {
				/* Must have been "x-x" or "x+(-x)". */
				if (Is_rounding_mode(ROUNDMINUS))
					Sgl_setone_sign(resultp1);
				Sgl_copytoptr(resultp1,dstptr);
				return(NOEXCEPTION);
			}
			result_exponent--;

			/* Look to see if normalization is finished. */
			if (Sgl_isone_hidden(resultp1)) {
				/* No further normalization is needed */
				goto round;
			}

			/* Discover first one bit to determine shift amount.
			 * Use a modified binary search.  We have already
			 * shifted the result one position right and still
			 * not found a one so the remainder of the extension
			 * must be zero and simplifies rounding. */
			/* Scan bytes */
			while (Sgl_iszero_hiddenhigh7mantissa(resultp1)) {
				Sglext_leftshiftby8(resultp1,resultp2);
				result_exponent -= 8;
			}
			/* Now narrow it down to the nibble */
			if (Sgl_iszero_hiddenhigh3mantissa(resultp1)) {
				/* The lower nibble contains the
				 * normalizing one */
				Sglext_leftshiftby4(resultp1,resultp2);
				result_exponent -= 4;
			}
			/* Select case where first bit is set (already
			 * normalized) otherwise select the proper shift. */
			jumpsize = Sgl_hiddenhigh3mantissa(resultp1);
			if (jumpsize <= 7) switch(jumpsize) {
			case 1:
				Sglext_leftshiftby3(resultp1,resultp2);
				result_exponent -= 3;
				break;
			case 2:
			case 3:
				Sglext_leftshiftby2(resultp1,resultp2);
				result_exponent -= 2;
				break;
			case 4:
			case 5:
			case 6:
			case 7:
				Sglext_leftshiftby1(resultp1,resultp2);
				result_exponent -= 1;
				break;
			}
		} /* end if (hidden...)... */
	/* Fall through and round */
	} /* end if (save < 0)... */
	else {
		/* Add magnitudes */
		Sglext_addition(tmpresp1,tmpresp2,
			rightp1,rightp2, /*to*/resultp1,resultp2);
		sign_save = Sgl_signextendedsign(resultp1);
		if (Sgl_isone_hiddenoverflow(resultp1)) {
	    		/* Prenormalization required. */
	    		Sglext_arithrightshiftby1(resultp1,resultp2);
	    		result_exponent++;
		} /* end if hiddenoverflow... */
	} /* end else ...add magnitudes... */

	/* Round the result.  If the extension and lower two words are
	 * all zeros, then the result is exact.  Otherwise round in the
	 * correct direction.  Underflow is possible. If a postnormalization
	 * is necessary, then the mantissa is all zeros so no shift is needed.
	 */
  round:
	if (result_exponent <= 0 && !Is_underflowtrap_enabled()) {
		Sglext_denormalize(resultp1,resultp2,result_exponent,is_tiny);
	}
	Sgl_set_sign(resultp1,/*using*/sign_save);
	if (Sglext_isnotzero_mantissap2(resultp2)) {
		inexact = TRUE;
		switch(Rounding_mode()) {
		case ROUNDNEAREST: /* The default. */
			if (Sglext_isone_highp2(resultp2)) {
				/* at least 1/2 ulp */
				if (Sglext_isnotzero_low31p2(resultp2) ||
				    Sglext_isone_lowp1(resultp1)) {
					/* either exactly half way and odd or
					 * more than 1/2ulp */
					Sgl_increment(resultp1);
				}
			}
	    		break;

		case ROUNDPLUS:
	    		if (Sgl_iszero_sign(resultp1)) {
				/* Round up positive results */
				Sgl_increment(resultp1);
			}
			break;
	    
		case ROUNDMINUS:
	    		if (Sgl_isone_sign(resultp1)) {
				/* Round down negative results */
				Sgl_increment(resultp1);
			}
	    
		case ROUNDZERO:;
			/* truncate is simple */
		} /* end switch... */
		if (Sgl_isone_hiddenoverflow(resultp1)) result_exponent++;
	}
	if (result_exponent >= SGL_INFINITY_EXPONENT) {
		/* Overflow */
		if (Is_overflowtrap_enabled()) {
                        /*
                         * Adjust bias of result
                         */
                        Sgl_setwrapped_exponent(resultp1,result_exponent,ovfl);
                        Sgl_copytoptr(resultp1,dstptr);
                        if (inexact)
                            if (Is_inexacttrap_enabled())
                                return (OPC_2E_OVERFLOWEXCEPTION |
					OPC_2E_INEXACTEXCEPTION);
                            else Set_inexactflag();
                        return (OPC_2E_OVERFLOWEXCEPTION);
		}
		inexact = TRUE;
		Set_overflowflag();
		Sgl_setoverflow(resultp1);
	} else if (result_exponent <= 0) {	/* underflow case */
		if (Is_underflowtrap_enabled()) {
                        /*
                         * Adjust bias of result
                         */
                	Sgl_setwrapped_exponent(resultp1,result_exponent,unfl);
			Sgl_copytoptr(resultp1,dstptr);
                        if (inexact)
                            if (Is_inexacttrap_enabled())
                                return (OPC_2E_UNDERFLOWEXCEPTION |
					OPC_2E_INEXACTEXCEPTION);
                            else Set_inexactflag();
	    		return(OPC_2E_UNDERFLOWEXCEPTION);
		}
		else if (inexact && is_tiny) Set_underflowflag();
	}
	else Sgl_set_exponent(resultp1,result_exponent);
	Sgl_copytoptr(resultp1,dstptr);
	if (inexact) 
		if (Is_inexacttrap_enabled()) return(OPC_2E_INEXACTEXCEPTION);
		else Set_inexactflag();
    	return(NOEXCEPTION);
}

/*
 *  Single Floating-point Multiply Negate Fused Add
 */

sgl_fmpynfadd(src1ptr,src2ptr,src3ptr,status,dstptr)

sgl_floating_point *src1ptr, *src2ptr, *src3ptr, *dstptr;
unsigned int *status;
{
	unsigned int opnd1, opnd2, opnd3;
	register unsigned int tmpresp1, tmpresp2;
	unsigned int rightp1, rightp2;
	unsigned int resultp1, resultp2 = 0;
	register int mpy_exponent, add_exponent, count;
	boolean inexact = FALSE, is_tiny = FALSE;

	unsigned int signlessleft1, signlessright1, save;
	register int result_exponent, diff_exponent;
	int sign_save, jumpsize;
	
	Sgl_copyfromptr(src1ptr,opnd1);
	Sgl_copyfromptr(src2ptr,opnd2);
	Sgl_copyfromptr(src3ptr,opnd3);

	/* 
	 * set sign bit of result of multiply
	 */
	if (Sgl_sign(opnd1) ^ Sgl_sign(opnd2)) 
		Sgl_setzero(resultp1);
	else 
		Sgl_setnegativezero(resultp1); 

	/*
	 * Generate multiply exponent 
	 */
	mpy_exponent = Sgl_exponent(opnd1) + Sgl_exponent(opnd2) - SGL_BIAS;

	/*
	 * check first operand for NaN's or infinity
	 */
	if (Sgl_isinfinity_exponent(opnd1)) {
		if (Sgl_iszero_mantissa(opnd1)) {
			if (Sgl_isnotnan(opnd2) && Sgl_isnotnan(opnd3)) {
				if (Sgl_iszero_exponentmantissa(opnd2)) {
					/* 
					 * invalid since operands are infinity 
					 * and zero 
					 */
					if (Is_invalidtrap_enabled())
						return(OPC_2E_INVALIDEXCEPTION);
					Set_invalidflag();
					Sgl_makequietnan(resultp1);
					Sgl_copytoptr(resultp1,dstptr);
					return(NOEXCEPTION);
				}
				/*
				 * Check third operand for infinity with a
				 *  sign opposite of the multiply result
				 */
				if (Sgl_isinfinity(opnd3) &&
				    (Sgl_sign(resultp1) ^ Sgl_sign(opnd3))) {
					/* 
					 * invalid since attempting a magnitude
					 * subtraction of infinities
					 */
					if (Is_invalidtrap_enabled())
						return(OPC_2E_INVALIDEXCEPTION);
					Set_invalidflag();
					Sgl_makequietnan(resultp1);
					Sgl_copytoptr(resultp1,dstptr);
					return(NOEXCEPTION);
				}

				/*
			 	 * return infinity
			 	 */
				Sgl_setinfinity_exponentmantissa(resultp1);
				Sgl_copytoptr(resultp1,dstptr);
				return(NOEXCEPTION);
			}
		}
		else {
			/*
		 	 * is NaN; signaling or quiet?
		 	 */
			if (Sgl_isone_signaling(opnd1)) {
				/* trap if INVALIDTRAP enabled */
				if (Is_invalidtrap_enabled()) 
			    		return(OPC_2E_INVALIDEXCEPTION);
				/* make NaN quiet */
				Set_invalidflag();
				Sgl_set_quiet(opnd1);
			}
			/* 
			 * is second operand a signaling NaN? 
			 */
			else if (Sgl_is_signalingnan(opnd2)) {
				/* trap if INVALIDTRAP enabled */
				if (Is_invalidtrap_enabled())
			    		return(OPC_2E_INVALIDEXCEPTION);
				/* make NaN quiet */
				Set_invalidflag();
				Sgl_set_quiet(opnd2);
				Sgl_copytoptr(opnd2,dstptr);
				return(NOEXCEPTION);
			}
			/* 
			 * is third operand a signaling NaN? 
			 */
			else if (Sgl_is_signalingnan(opnd3)) {
				/* trap if INVALIDTRAP enabled */
				if (Is_invalidtrap_enabled())
			    		return(OPC_2E_INVALIDEXCEPTION);
				/* make NaN quiet */
				Set_invalidflag();
				Sgl_set_quiet(opnd3);
				Sgl_copytoptr(opnd3,dstptr);
				return(NOEXCEPTION);
			}
			/*
		 	 * return quiet NaN
		 	 */
			Sgl_copytoptr(opnd1,dstptr);
			return(NOEXCEPTION);
		}
	}

	/*
	 * check second operand for NaN's or infinity
	 */
	if (Sgl_isinfinity_exponent(opnd2)) {
		if (Sgl_iszero_mantissa(opnd2)) {
			if (Sgl_isnotnan(opnd3)) {
				if (Sgl_iszero_exponentmantissa(opnd1)) {
					/* 
					 * invalid since multiply operands are
					 * zero & infinity
					 */
					if (Is_invalidtrap_enabled())
						return(OPC_2E_INVALIDEXCEPTION);
					Set_invalidflag();
					Sgl_makequietnan(opnd2);
					Sgl_copytoptr(opnd2,dstptr);
					return(NOEXCEPTION);
				}

				/*
				 * Check third operand for infinity with a
				 *  sign opposite of the multiply result
				 */
				if (Sgl_isinfinity(opnd3) &&
				    (Sgl_sign(resultp1) ^ Sgl_sign(opnd3))) {
					/* 
					 * invalid since attempting a magnitude
					 * subtraction of infinities
					 */
					if (Is_invalidtrap_enabled())
				       		return(OPC_2E_INVALIDEXCEPTION);
				       	Set_invalidflag();
				       	Sgl_makequietnan(resultp1);
					Sgl_copytoptr