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

					/* either exactly half way and odd or
					 * more than 1/2ulp */
					Dbl_increment(resultp1,resultp2);
				}
			}
	    		break;

		case ROUNDPLUS:
	    		if (Dbl_iszero_sign(resultp1)) {
				/* Round up positive results */
				Dbl_increment(resultp1,resultp2);
			}
			break;
	    
		case ROUNDMINUS:
	    		if (Dbl_isone_sign(resultp1)) {
				/* Round down negative results */
				Dbl_increment(resultp1,resultp2);
			}
	    
		case ROUNDZERO:;
			/* truncate is simple */
		} /* end switch... */
		if (Dbl_isone_hiddenoverflow(resultp1)) result_exponent++;
	}
	if (result_exponent >= DBL_INFINITY_EXPONENT) {
		/* Overflow */
		if (Is_overflowtrap_enabled()) {
                        /*
                         * Adjust bias of result
                         */
                        Dbl_setwrapped_exponent(resultp1,result_exponent,ovfl);
                        Dbl_copytoptr(resultp1,resultp2,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();
		Dbl_setoverflow(resultp1,resultp2);
	} else if (result_exponent <= 0) {	/* underflow case */
		if (Is_underflowtrap_enabled()) {
                        /*
                         * Adjust bias of result
                         */
                	Dbl_setwrapped_exponent(resultp1,result_exponent,unfl);
			Dbl_copytoptr(resultp1,resultp2,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 Dbl_set_exponent(resultp1,result_exponent);
	Dbl_copytoptr(resultp1,resultp2,dstptr);
	if (inexact) 
		if (Is_inexacttrap_enabled()) return(OPC_2E_INEXACTEXCEPTION);
		else Set_inexactflag();
    	return(NOEXCEPTION);
}

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

sgl_fmpyfadd(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_setnegativezero(resultp1); 
	else Sgl_setzero(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(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(opnd2)) {
				/* trap if INVALIDTRAP enabled */
				if (Is_invalidtrap_enabled())
					return(OPC_2E_INVALIDEXCEPTION);
				/* make NaN quiet */
				Set_invalidflag();
				Sgl_set_quiet(opnd2);
			}
			/* 
			 * 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(opnd2,dstptr);
			return(NOEXCEPTION);
		}
	}

	/*
	 * check third operand for NaN's or infinity
	 */
	if (Sgl_isinfinity_exponent(opnd3)) {
		if (Sgl_iszero_mantissa(opnd3)) {
			/* return infinity */
			Sgl_copytoptr(opnd3,dstptr);
			return(NOEXCEPTION);
		} else {
			/*
			 * is NaN; signaling or quiet?
			 */
			if (Sgl_isone_signaling(opnd3)) {
				/* trap if INVALIDTRAP enabled */
				if (Is_invalidtrap_enabled())
					return(OPC_2E_INVALIDEXCEPTION);
				/* make NaN quiet */
				Set_invalidflag();
				Sgl_set_quiet(opnd3);
			}
			/*
			 * return quiet NaN
 			 */
			Sgl_copytoptr(opnd3,dstptr);
			return(NOEXCEPTION);
		}
    	}

	/*
	 * Generate multiply mantissa
	 */
	if (Sgl_isnotzero_exponent(opnd1)) {
		/* set hidden bit */
		Sgl_clear_signexponent_set_hidden(opnd1);
	}
	else {
		/* check for zero */
		if (Sgl_iszero_mantissa(opnd1)) {
			/*
			 * Perform the add opnd3 with zero here.
			 */
			if (Sgl_iszero_exponentmantissa(opnd3)) {
				if (Is_rounding_mode(ROUNDMINUS)) {
					Sgl_or_signs(opnd3,resultp1);
				} else {
					Sgl_and_signs(opnd3,resultp1);
				}
			}
			/*
			 * Now let's check for trapped underflow case.
			 */
			else if (Sgl_iszero_exponent(opnd3) &&
			         Is_underflowtrap_enabled()) {
                    		/* need to normalize results mantissa */
                    		sign_save = Sgl_signextendedsign(opnd3);
				result_exponent = 0;
                    		Sgl_leftshiftby1(opnd3);
                    		Sgl_normalize(opnd3,result_exponent);
                    		Sgl_set_sign(opnd3,/*using*/sign_save);
                    		Sgl_setwrapped_exponent(opnd3,result_exponent,
							unfl);
                    		Sgl_copytoptr(opnd3,dstptr);
                    		/* inexact = FALSE */
                    		return(OPC_2E_UNDERFLOWEXCEPTION);
			}
			Sgl_copytoptr(opnd3,dstptr);
			return(NOEXCEPTION);
		}
		/* is denormalized, adjust exponent */
		Sgl_clear_signexponent(opnd1);
		Sgl_leftshiftby1(opnd1);
		Sgl_normalize(opnd1,mpy_exponent);
	}
	/* opnd2 needs to have hidden bit set with msb in hidden bit */
	if (Sgl_isnotzero_exponent(opnd2)) {
		Sgl_clear_signexponent_set_hidden(opnd2);
	}
	else {
		/* check for zero */
		if (Sgl_iszero_mantissa(opnd2)) {
			/*
			 * Perform the add opnd3 with zero here.
			 */
			if (Sgl_iszero_exponentmantissa(opnd3)) {
				if (Is_rounding_mode(ROUNDMINUS)) {
					Sgl_or_signs(opnd3,resultp1);
				} else {
					Sgl_and_signs(opnd3,resultp1);
				}
			}
			/*
			 * Now let's check for trapped underflow case.
			 */
			else if (Sgl_iszero_exponent(opnd3) &&
			    Is_underflowtrap_enabled()) {
                    		/* need to normalize results mantissa */
                    		sign_save = Sgl_signextendedsign(opnd3);
				result_exponent = 0;
                    		Sgl_leftshiftby1(opnd3);
                    		Sgl_normalize(opnd3,result_exponent);
                    		Sgl_set_sign(opnd3,/*using*/sign_save);
                    		Sgl_setwrapped_exponent(opnd3,result_exponent,
							unfl);
                    		Sgl_copytoptr(opnd3,dstptr);
                    		/* inexact = FALSE */
                    		return(OPC_2E_UNDERFLOWEXCEPTION);
			}
			Sgl_copytoptr(opnd3,dstptr);
			return(NOEXCEPTION);
		}
		/* is denormalized; want to normalize */
		Sgl_clear_signexponent(opnd2);
		Sgl_leftshiftby1(opnd2);
		Sgl_normalize(opnd2,mpy_exponent);
	}

	/* Multiply the first two source mantissas together */

	/* 
	 * The intermediate result will be kept in tmpres,
	 * which needs enough room for 106 bits of mantissa,
	 * so lets call it a Double extended.
	 */
	Sglext_setzero(tmpresp1,tmpresp2);

	/* 
	 * Four bits at a time are inspected in each loop, and a 
	 * simple shift and add multiply algorithm is used. 
	 */ 
	for (count = SGL_P-1; count >= 0; count -= 4) {
		Sglext_rightshiftby4(tmpresp1,tmpresp2);
		if (Sbit28(opnd1)) {
	 		/* Twoword_add should be an ADD followed by 2 ADDC's */
			Twoword_add(tmpresp1, tmpresp2, opnd2<<3, 0);
		}
		if (Sbit29(opnd1)) {
			Twoword_add(tmpresp1, tmpresp2, opnd2<<2, 0);
		}
		if (Sbit30(opnd1)) {
			Twoword_add(tmpresp1, tmpresp2, opnd2<<1, 0);
		}
		if (Sbit31(opnd1)) {
			Twoword_add(tmpresp1, tmpresp2, opnd2, 0);
		}
		Sgl_rightshiftby4(opnd1);
	}
	if (Is_sexthiddenoverflow(tmpresp1)) {
		/* result mantissa >= 2 (mantissa overflow) */
		mpy_exponent++;
		Sglext_rightshiftby4(tmpresp1,tmpresp2);
	} else {
		Sglext_rightshiftby3(tmpresp1,tmpresp2);
	}

	/*
	 * Restore the sign of the mpy result which was saved in resultp1.
	 * The exponent will continue to be kept in mpy_exponent.
	 */
	Sglext_set_sign(tmpresp1,Sgl_sign(resultp1));

	/* 
	 * No rounding is required, since the result of the multiply
	 * is exact in the extended format.
	 */