Pipelined Floating Point Adder for Digital Computer

The present application discloses certain aspects of a computing system that is further described in the following U.S. patent applications filed concurrently with the present application: Evans et al., AN INTERFACE BETWEEN A SYSTEM CONTROL UNIT AND A SERVICE PROCESSING UNIT OF A DIGITAL COMPUTER, Ser. No. 07/306,325 filed Feb. 3, 1989; Arnold et al., METHOD AND APPARATUS FOR INTERFACING A SYSTEM CONTROL UNIT FOR A MULTIPROCESSOR SYSTEM WITH THE CENTRAL PROCESSING UNITS, Ser. No. 07/306,837 filed Feb. 3, 1989; Gagliardo et al., METHOD AND MEANS FOR INTERFACING A SYSTEM CONTROL UNIT FOR A MULTI-PROCESSOR SYSTEM WITH THE SYSTEM MAIN MEMORY, Ser. No. 07/306,326 filed Feb. 3, 1989; D. Fite et al., METHOD AND APPARATUS FOR RESOLVING A VARIABLE NUMBER OF POTENTIAL MEMORY ACCESS CONFLICTS IN A PIPELINED COMPUTER SYSTEM, Ser. No. 07/306,767; D. Fite et al., DECODING MULTIPLE SPECIFIERS IN A VARIABLE LENGTH INSTRUCTION ARCHITECTURE, Ser. No. 07/307,347 filed Feb. 3, 1989; D. Fite et al., VIRTUAL INSTRUCTION CACHE REFILL ALGORITHM, Ser. No. 07/306,831 filed Feb. 3, 1989; Murray et al., PIPELINE PROCESSING OF REGISTER AND REGISTER MODIFYING SPECIFIERS WITHIN THE SAME INSTRUCTION, Ser. No. 07/306,833 filed Feb. 3, 1989; Murray et al., MULTIPLE INSTRUCTION PREPROCESSING SYSTEM WITH DATA DEPENDENCY RESOLUTION FOR DIGITAL COMPUTERS, Ser. No. 07/306,773 filed Feb. 3, 1989; Murray et al., PREPROCESSING IMPLIED SPECIFIERS IN A PIPELINED PROCESSOR, Ser. No. 07/306,846 filed Feb. 3, 1989; D. Fite et al., BRANCH PREDICTION, Ser. No. 07/306,760 filed Feb. 3, 1989; Grundmann et al., SELF TIMED REGISTER FILE, Ser. No. 07/306,445 filed Feb. 3, 1989; Beaven et al., METHOD AND APPARATUS FOR DETECTING AND CORRECTING ERRORS IN A PIPELINED COMPUTER SYSTEM, Ser. No. 07/306,828 filed Feb. 3, 1989; Flynn et al., METHOD AND MEANS FOR ARBITRATING COMMUNICATION REQUESTS USING A SYSTEM CONTROL UNIT IN A MULTI-PROCESSOR SYSTEM, Ser. No. 07/306,871 filed Feb. 3, 1989; E. Fite et al., CONTROL OF MULTIPLE FUNCTION UNITS WITH PARALLEL OPERATION IN A MICROCODED EXECUTION UNIT, Ser. No. 07/306,832 filed Feb. 3, 1989; Webb, Jr. et al., PROCESSING OF MEMORY ACCESS EXCEPTIONS WITH PRE-FETCHED INSTRUCTIONS WITHIN THE INSTRUCTION PIPELINE OF A VIRTUAL MEMORY SYSTEM-BASED DIGITAL COMPUTER, Ser. No. 07/306,866 filed Feb. 3, 1989; Hetherington et al., METHOD AND APPARATUS FOR CONTROLLING THE CONVERSION OF VIRTUAL TO PHYSICAL MEMORY ADDRESSES IN A DIGITAL COMPUTER SYSTEM, Ser. No. 07/306,544 filed Feb. 3, 1989; Hetherington et al., WRITE BACK BUFFER WITH ERROR CORRECTING CAPABILITIES, Ser. No. 07/306,703 filed Feb. 3, 1989; Flynn et al., METHOD AND MEANS FOR ARBITRATING COMMUNICATION REQUESTS USING A SYSTEM CONTROL UNIT IN A MULTI-PROCESSING SYSTEM, Ser. No. 07/306,871 filed Feb. 3, 1989; Chinnasway et al., MODULAR CROSSBAR INTERCONNECTION NETWORK FOR DATA TRANSACTIONS BETWEEN SYSTEM UNITS IN A MULTI-PROCESSOR SYSTEM, Ser. No. 07/306,336 filed Feb. 3, 1989; Polzin et al., METHOD AND APPARATUS FOR INTERFACING A SYSTEM CONTROL UNIT FOR A MULTI-PROCESSOR SYSTEM WITH INPUT/OUTPUT UNITS, Ser. No. 07/306,862 filed Feb. 3, 1989; Gagliardo et al., MEMORY CONFIGURATION FOR USE WITH MEANS FOR INTERFACING A SYSTEM CONTROL UNIT FOR A MULTI-PROCESSOR SYSTEM WITH THE SYSTEM MAIN MEMORY, Ser. No. 07/306,404 filed Feb. 3, 1989; and Gagliardo et al., METHOD AND MEANS FOR ERROR CHECKING OF DRAM-CONTROL SIGNALS BETWEEN SYSTEM MODULES, Ser. No. 07/306,836 filed Feb. 3, 1989.

The present invention relates generally to floating point processors for use in digital computers and, more particularly, to an improved pipelined floating point adder/subtractor.

A floating-point number is a sequence of contiguous bits representing the fraction FRAC (or mantissa), the exponent EXP, and the sign S of a number N defined by the formula:

A typical 64-bit format for representing a floating-point number is shown in the following table: ##STR1## The fraction FRAC is expressed as a 53-bit positive fraction, with the binary point positioned to the left of the most significant bit. If the fraction FRAC is not zero, the most significant bit of FRAC must be 1, so this bit is not stored; this bit is referred to as the "hidden bit" and enables FRAC to be expressed in 52 bits rather than 53. One of the remaining twelve bits is used to express the sign S, and the other eleven bits are used to express the exponent EXP.

As is well known, the binary points of two floating-point numbers must be aligned before adding or subtracting the two numbers. This alignment is accomplished b comparing the relative magnitudes of the two exponents .vertline.EXP.sub.1 .vertline. and .vertline.EXP.sub.2 .vertline. and then shifting the fraction with the smaller exponent (EXP.sub.1 -EXP.sub.2) places to the right. The two fractions can then be added or subtracted, i.e., FRAC.sub.1 .+-.FRAC.sub.2, with the larger exponent serving as the exponent of the result. (In subtraction, the two's complement of the subtrahend is added to the minuend.) The resulting sum is then normalized by shifting the fraction to the left until the most significant bit is a 1, and decreasing the exponent accordingly. Finally, the result is rounded, e.g., by adding a rounding constant.

The steps described above work in a pipelined floating-point adder/subtractor if the exponents are different. If the exponents are of equal size, however, it is not known which number is smaller, which can be a problem in subtraction. One technique for subtracting numbers having exponents of equal size is to initially guess which number is smaller, but if the guess is wrong, an extra addition step is required to obtain the correct number. Another technique is to use two adders and perform the subtraction both ways, and then select the correct fraction in a subsequent step. This technique has the disadvantage of requiring two adders, and may result in extra pin requirements on integrated circuits, and extra loads on critical signals.

Another problem encountered in a pipelined floating-point adder/subtractor is the "sticky bit" problem, i.e., the loss of a one to the right of the least significant bit of a shifted number, due to truncation during alignment of that number. If there is no compensation for this loss, the addition/subtraction operation can produce an inaccurate result.

There is provided a system for subtracting two floating-point binary numbers by aligning the two fractions for subtraction, arbitrarily designating the fraction of one of the two floating-point numbers as the subtrahend, producing the complement of the designated fraction and adding that complement to the other fraction, normalizing the result, determining whether the result is negative and, if it is, producing the complement of the normalized result, selecting the larger of the exponents of the two floating-point numbers, and adjusting the value of the selected exponent in accordance with the normalization of the result.

There is also provided a system for producing a sticky bit signal by aligning the two fractions for subtraction by shifting one of the two fractions to the right, determining the number of consecutive zeros in the one fraction, prior to the shifting thereof, beginning at the least significant bit position, comparing (1) the number of positions the one fraction is shifted in the aligning step, with (2) the number of consecutive zeros in the one fraction, and producing a sticky bit signal when the number of consecutive zeros is less than the number of positions the one fraction is shifted in the aligning step, the sticky bit signal indicating the truncation of at least one set bit during the aligning step .

Other objects and advantages of the invention will become apparent upon reading the following detailed description and upon reference to the drawings in which:

FIG. 1 is a block diagram of a floating point adder embodying the present invention;

FIG. 2 is a schematic diagram of the fraction adder, normalization unit and rounding unit in the adder of FIG. 1;

FIG. 3 is a more detailed schematic diagram of the entire adder, shown in FIG. 1;

FIG. 4 is a schematic diagram of the exponent processing unit in the system of FIG. 3;

FIG. 5 is a schematic diagram of a shifter used in the fraction alignment units included in the system of FIG. 3;

FIG. 6 is a schematic diagram of one of the fraction alignment units included in the system of FIG. 3;

FIG. 7 is a schematic diagram of the trailing zero detector included in the circuit of FIG. 6; and

FIG. 8 is a truth table for the eight-bit priority encoders used in the circuit illustrated in FIG. 7.

While the invention is susceptible to various modifications and alternative forms, specific embodiments thereof have been shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that it is not intended to limit the invention to the particular forms disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the appended claims.

Turning now to the drawings and referring first to FIG. 1, two source operands SOURCE.sub.1 and SOURCE.sub.2, each of which comprises a 64-bit floating-point number, are supplied via latches 10 and 11 to a pair of fraction alignment units 12 and 13. The operand SOURCE.sub.1 comprises a 52-bit fraction FRAC.sub.1, an 11-bit exponent EXP.sub.1, and a sign bit S.sub.1. Similarly, the operand SOURCE.sub.2 comprises a 52-bit fraction FRAC.sub.2, an 11-bit exponent EXP.sub.2, and a sign bit S.sub.2. It will be understood that the illustrative system can also process floating-point numbers in other formats.

As explained previously, the binary points of the two source operands must be aligned before meaningful addition or subtraction can be performed. This alignment is effected in the two alignment units 12 and 13, which initially extract the fractions FRAC.sub.1 and FRAC.sub.2 from the source operands SOURCE.sub.1 and SOURCE.sub.2. The alignment is then accomplished by shifting the fraction with the smaller exponent a prescribed number of places to the right; this prescribed number is equal to the difference between the two exponents, as determined by an exponent processing unit 14.

The exponent processing unit 14 receives the two operands SOURCE.sub.1 and SOURCE.sub.2 from the latches 10 and 11, extracts the eleven exponent bits EXP.sub.1 and EXP.sub.2 from the respective operands, and produces a pair of 6-bit control signals for the two fraction alignment units 12 and 13. (In actual practice, each of the two alignment units 12 and 13 can contain its own exponent subtractor.) These 6-bit control signals determine which of the fractions FRAC.sub.1 and FRAC.sub.2, if any, is to be shifted in the alignment units 12 and 13, and the magnitude of the shift. The control signal supplied to the FRAC.sub.1 alignment unit 12 is zero if EXP.sub.1 is greater than EXP.sub.2, and is equal to the difference between the two exponents if EXP.sub.2 is greater than or equal to EXP.sub.1 ; thus, the fraction FRAC.sub.2 is shifted to the right by a number of positions equal to the difference between EXP.sub.1 and EXP.sub.2 if EXP.sub.2 is greater than EXP.sub.1. The control signal supplied to the FRAC.sub.2 alignment unit is zero if EXP.sub.2 is greater than or equal to EXP.sub.1, and is equal to the difference between the two exponents if EXP.sub.1 is greater than EXP.sub.2 ; thus, the fraction FRAC.sub.2 is shifted to the right by a number of positions equal to the difference between EXP.sub.1 and EXP.sub.2 if EXP.sub.1 is greater than EXP.sub.2. The net effect of these control signals, therefore, is to shift the fraction of the source operand having the smaller exponent the proper number of binary places to permit the fractions of the two source operands to be added or subtracted.

From the alignment units 12 and 13, the two aligned fractions FRAC.sub.1 and FRAC.sub.2 are passed to a fraction adder 15 through a pair of latches 16 and 17. As already mentioned, and as will be described in more detail below, the adder 15 can be used to either add or subtract the two fractions FRAC.sub.1 and FRAC.sub.2.

The sum FRAC.sub.S produced by the adder 15 is passed to a normalization unit 18 which shifts the sum FRAC.sub.S to the left by the requisite number of positions to normalize FRAC.sub.FRAC.sub.S. At the same time the normalization unit supplies a control signal to an exponent adjust unit 19, which receives the greater of the two exponents EXP.sub.1 and EXP.sub.2 via a latch 20. This greater exponent is referred to hereinafter as the sum exponent EXP.sub.S. The purpose of the exponent adjust unit 19 is to decrease EXP.sub.S by a number equal to the number of positions that FRAC.sub.S is shifted to the left in the normalization unit 18. Both the output FRACN of the normalization unit 18 and the output EXP.sub.N of the exponent adjust unit 19 are supplied to a rounding unit 21 which rounds the result by adding a rounding constant. The final fraction FRACF and the final exponent EXP.sub.F are then packed to form a result, which is supplied to a latch 22.

Referring now to FIG. 2, the operation of the adder 15, the normalization unit 18 and the rounding unit 21 will be described in more detail in connection of the schematic diagram of these three units in FIG. 2. For an addition operation in the fraction adder 15, the two 64-bit, aligned fractions FRAC.sub.1 and FRAC.sub.2 are simply summed in a conventional adder 30. For a subtraction operation, the control input SUBFRAC at a subtraction input terminal SUB is asserted to cause the complement of FRAC.sub.1 to be added to FRAC.sub.2, i.e., FRAC.sub.1 is subtracted from FRAC.sub.2. It is important to note that FRAC.sub.1 may always be arbitrarily designated as the subtrahend, regardless of whether FRAC.sub.1 is larger or smaller than FRAC.sub.2, because negative results can be correctly handled by the system. Consequently, floating-point numbers having equal exponents can be added or subtracted using a single adder, performing a single addition operation. The result produced by the adder 30 is passed over a bus 31 to the normalization unit 18.

As will be described in more detail below, the adder 30 also produces a control output signal -/+which indicates whether the number on bus 31 is positive or negative. This control signal -/+is asserted for a negative result.

In the normalization unit 18, the number on the bus 31 is received by both a leading-zeros detector 32 and a leading-ones detector 33. It will be recalled that normalizing a positive floating-point number requires shifting the fraction to the left to push off the leading zeros (in the most significant bit positions) until a one appears in the most significant bit position. In the case of negative numbers, a complementary normalizing operation is performed, i.e., the fraction is shifted to the left to push off the leading ones until a zero appears in the most significant bit position. Thus, in the system of FIG. 2 the leading-zeros detector 32 is used to determine the number of left shifts required to normalize positive numbers, and the leading-ones detector 33 does the same for negative numbers. In the special case where the number on the bus 31 is all ones, the leading-ones detector 33 gives an erroneous result, but this error is subsequently corrected by the ensuing overflow signal in the rounding unit 21, as will be described in more detail below.

The outputs of the two detectors 32 and 33 are passed to a multiplexer 34 which selects the output of one of the two detectors in response to the signal -/+. That is, the output of the leading-ones detector 33 is selected if the signal -/+is asserted, indicating that the number on the bus 31 is negative. If the signal -/+is not asserted, the multiplexer 34 selects the output of the leading-zeros detector 32. From the multiplexer 34, the selected detector output is supplied to a shifter 35 which receives the number from the bus 31 and shifts that number to the left by the number of positions dictated by the control signal from the multiplexer 34. The resulting normalized number FRACN is then passed over a bus 36 to the rounding unit 21. The shifter 35 will be described in more detail below in connection with FIG. 5.

The output of the multiplexer 34 is also supplied to the exponent adjust unit 19 shown in FIG. 1. In the exponent adjust unit 19, the exponent EXP.sub.S is decremented by one for each left shift of the fraction FRAC.sub.S in the shifter 35.

Returning to FIG. 2, the rounding unit 21 includes an adder 37 which adds a rounding constant K to the fraction FRACN received on the bus 36. The constant K is equal to one-half of the least significant bit in FRAC.sub.S. The resulting rounded number FRAC.sub.F is produced on an output bus 38 from the adder 37, as a 64-bit normalized, rounded, floating-point binary number.

It will be appreciated from the description thus far that in a subtraction operation, the fraction FRAC.sub.1 is always subtracted from the fraction FRAC.sub.2, regardless of whether FRAC.sub.1 is larger or smaller than FRAC.sub.2. Consequently, when FRAC.sub.1 is larger than FRAC.sub.2, the result of the subtraction operation will be negative, and the control signal -/+ will be asserted. The negative result will also be the complement of the true difference between FRAC.sub.1 and FRAC.sub.2. Thus, the signal -/+ is applied to a subtraction input terminal SUB of the adder 37 so that when a negative result is indicated by assertion of the signal -/+, the complement of FRAC.sub.N is added to the rounding constant K in the adder 37. The adder 37 also produces a carry-out signal C.sub.out which determines whether the adjusted value of EXP is increased by one or two, as will be described below in connection with FIG. 3.

Referring now to FIG. 3, the two latches 10 and 11 which receive the respective source operands SOURCE.sub.1 and SOURCE.sub.2, the two fraction alignment units 12 and 13, and the exponent processing unit 14, have already been described above in connection with FIG. 1. FIG. 3 also shows sticky-bit output signals SB.sub.1 and SB.sub.2 from the respective fraction alignment units 12 and 13, and the generation and processing of these sticky bit signals will be described below.

The outputs of the two fraction alignment units 12 and 13 are fed to the respective latches 16 and 17, and the clock inputs C to these latches then control when the two aligned fractions FRAC.sub.1 and FRAC.sub.2 are supplied to a fraction adder 40. While these numbers are being passed from the registers 16 and 17 to the adder 40, a leading zero is inserted into each number as the most significant bit, via lines 41 and 42. These added zeros are "overflow bits" which serve as placeholders to receive extra bits produced as a result of the addition performed in the fraction adder 40, i.e., to allow for a sum that is greater than or equal to one.

The number FRAC.sub.1 from the register 16 is passed through an exclusive OR gate 43, which also receives the control signal SUBFRAC which is asserted when the two numbers FRAC.sub.1 and FRAC.sub.2 are to be subtracted rather than added in the adder 40. The assertion of this signal SUBFRAC causes the number FRAC.sub.1 to be applied to the adder 40 in complement form rather than true form. Specifically, the gate 43 inverts the number FRAC.sub.2, and a one is supplied to the carry-in of the adder 40 by the signal SUBFRAC via inverter 44 and NOR gate 45, thereby producing the two's complement of FRAC.sub.1 for addition to FRAC.sub.2.

The signal SUBFRAC is produced by an exclusive OR gate 46 which receives three input signals: a subtraction signal SUB indicating that the source operand SOURCE.sub.1 is to be subtracted from the source operand SOURCE.sub.2, and the two sign bits S.sub.1 and S.sub.2 of the two source operands. This combination of input signals causes the output of the gate 46, i.e., the signal SUBFRAC.sub.1, to be asserted whenever (1) SUB is asserted and S.sub.1 and S.sub.2 are the same, or (2) SUB is not asserted but S.sub.1 and S.sub.2 are different. These are, of course, the conditions in which FRAC.sub.1 and FRAC.sub.2 must be processed as having opposite signs.

The subtraction signal SUBFRAC is not only supplied to the adder 40 (via inverter 44 and gate 45), but also to an AND gate 47 for generating the signal -/+. The second input to the AND gate 47 is the most significant bit in the output of the adder 40; when this bit is set, it indicates that the output of the adder 40 is a negative number. Thus, assertion of the signal -/+ indicates that the sum FRAC.sub.S produced by the fraction adder 40 must be converted to its complement form to obtain the desired arithmetic result. The manner in which this is accomplished will be described below.

The output of the fraction adder 40 is supplied via bus 48 to a shifter 49 (corresponding to the shifter 35 described above), and then on to a rounding adder 50 via bus 51. Between the shifter 49 and the adder 50, the normalized result FRAC.sub.N is passed through an exclusive OR gate 52 which receives the signal -/+ from the AND gate 47. The assertion of the signal -/+ causes the normalized result FRAC.sub.R on the bus 51 to be applied to the adder 50 in complement form rather than true form. More specifically, the gate 52 inverts the FRAC.sub.N, and a one is supplied to the carry-in of the adder 50 by the signal -/+ via an inverter 53 and a NOR gate 54, thereby producing the two's complement of FRAC.sub.N. The output of the rounding adder represents the final value FRAC.sub.F of the fraction of the floating-point number representing the actual difference between the two original source operands SOURCE.sub.1 and SOURCE.sub.2. This number FRAC.sub.F is supplied via bus 55 to a packing unit 56 and then a latch 57.

Because the hidden bit is a part of the true value of each of the fractions FRAC.sub.1 and FRAC.sub.2, this bit must be restored to each fraction before it is processed in the adder/subtractor. Thus, as illustrated in FIG. 3, the hidden bits are restored to the two fractions before they enter the respective alignment units 12 and 13. Then when the final fraction value FRAC.sub.F is produced by the rounding adder 50, the hidden bit is removed again before the final value is packed with the final exponent value and sign.

Turning next to the processing of the exponent number EXP selected by the exponent processing unit 14 and fed to the latch 20, the clock signal applied to the latch 20 determines when this number is passed to the exponent adjust unit 19 (see FIG. 1). In FIG. 3 the exponent adjust unit 19 includes an adder 60 in which the number EXP.sub.S is decremented by a number equal to the number of positions by which the number FRAC.sub.S is shifted in the normalization unit 49. The decrementing number is determined in the same manner described in connection with FIG. 2, using the detectors 32 and 33 and the multiplexer 34.

From the exponent-adjusting adder 60, the adjusted exponent number EXP.sub.N is passed through a pair of adders 61 and 62 which increase the value of the number EXP.sub.N by one and two, respectively. The value of EXP.sub.N must always be increased by one because of the addition of the overflow bit to each of the fractions FRAC.sub.1 and FRAC.sub.2 before they entered the adder 40. When an overflow is produced in the rounding adder 50, the value of EXP.sub.N must be increased by two; this also corrects the error introduced by the leading ones detector 33 in the special case discussed above, where the number fed to the detector 33 is all ones.

To select between the two adders 61 and 62, the outputs of the two adders are supplied to a multiplexer 63, whose select input receives the carry-out from the rounding adder 50. When this carry-out is asserted, the multiplexer 63 selects the output of the adder 62 rather than the adder 61, so that the number EXP.sub.N from the adder 60 is increased by two rather than one. The output of the multiplexer 63 is a number EXP.sub.F which is the final value of the exponent for the floating-point number representing the actual difference between the two original source operands SOURCE.sub.1 and SOURCE.sub.2.

The sign of the ultimate result of the system shown in FIG. 3 is determined by result sign logic 70. This logic receives as input signals the subtraction signal SUB, a signal from the exponent processing unit 14 indicating whether EXP.sub.1 is greater than EXP.sub.2, the two sign bits S.sub.1 and S.sub.2 of the two source operands, and the -/+ signal from the gate 47.

The absolute value of the operand SOURCE.sub.1 is greater than the absolute value of the operand SOURCE.sub.2 when either (1) EXP.sub.1 is greater than EXP.sub.2 or (2) EXP.sub.1 and EXP.sub.2 are equal and the signal -/+ is asserted. When the absolute value of SOURCE.sub.1 is greater than the absolute value of SOURCE.sub.2, and SOURCE.sub.1 is being added to SOURCE.sub.2, the sign SF of the result should be the sign S.sub.1 of SOURCE.sub.1. Thus, the sign bit S.sub.1 is passed through the logic 70 and becomes the sign bit SF of the ultimate result of the floating-point addition operation. When the absolute value of SOURCE.sub.1 is less than the value of the absolute value of SOURCE.sub.2, and SOURCE.sub.1 and SOURCE.sub.2 are being added, the logic 70 produces a result sign SF which is the same as the sign S2 of SOURCE.sub.2.

When SOURCE.sub.1 is being subtracted from SOURCE.sub.2, the signal SUB is asserted. Now if the absolute value of SOURCE.sub.1 is also greater than the absolute value of SOURCE.sub.2, the result sign SF is the opposite of S.sub.1. If the absolute value of SOURCE.sub.1 is smaller than the absolute value of SOURCE.sub.2, the result sign SF is the same as S.sub.2.

A preferred circuit for the exponent processing unit 14 is shown in FIG. 4. This circuit includes two exponent subtractors 80 and 81, each of which recieves the two exponents EXP.sub.1 and EXP.sub.2. The use of the two subtractors makes it possible to physically package one of the subtractors with each of the alignment units 12 and 13. In both subtractors, a one is added to the most significant position of each input to restore the bias bit. A positive output from either of the two subtractors 80 and 81 is an indication that the exponent supplied to that subtractor as a minuend is larger than the exponent supplied to that subtractor as a subtrahend. Consequently, the numerical value of the positive output of that adder represents the number of places that the fraction corresponding to the exponent supplied to that adder as a subtrahend should be shifted to the right for alignment purposes. In the case of the subtractor 80, this output is furnished by an AND gate 82 which receives the numerical output of the subtractor 80 as one input and the complement of the sign bit as a second input, via an inverter 83. The output of the AND gate 82 is supplied to the fraction alignment unit 12 to control the number of left shifts of the fraction FRAC.sub.1. If the output of the subtractor 80 is negative, the sign bit is a one which, when inverted by the inverter 83, disables the AND gate 82 so that the fraction FRAC1 is not shifted.

In the case of the subtractor 81, the output is converted to an input signal for the fraction alignment unit 13 by an AND gate 84 which receives the numerical output of the subtractor 81 as one input and the complement of the sign bit as a second input, via an inverter 85. When EXP.sub.1 is larger than EXP.sub.2, the output of the AND gate 84 controls the number of left shifts of the fraction FRAC.sub.2 in the alignment unit 13. If the output of the adder 81 is negative, the sign bit is a one which, when inverted by the inverter 85, disables the AND gate 84 so that the fraction FRAC.sub.2 is not shifted.

The circuit of FIG. 4 also includes a multiplexer 86 which receives the two exponent values EXP.sub.1 and EXP.sub.2. The sign bit from the output of the subtractor 80 operates the select line of the multiplexer 86 so that it selects the exponent EXP.sub.2 as the larger exponent when the sign bit from the subtractor 80 is positive, and selects the exponent EXP.sub.1 when the sign bit from the subtractor 80 is negative. The particular exponent appearing at the output of the multiplexer 86 is then the larger of the two exponents. This is the exponent value EXP.sub.S and, after any adjustments made in the adders 60 and 61 or 62, becomes the final exponent value EXP.sub.F.

A preferred circuit for one of the fraction alignment units 12 or 13 is illustrated in FIG. 5. This circuit receives the output of one of the AND gates 82 or 84 and supplies successive pairs of the six least significant bits of that output to three 4-bit-shift multiplexers 90, 91 and 92. The 64 bits of one of the fractions FRAC.sub.1 or FRAC.sub.2 are supplied to the first multiplexer 90 and, if there is a one present on any of the two control inputs to this multiplexer, the 64 bits are shifted accordingly.

Specifically, the multiplexer 90 can shift the fraction by 1, 2 or 3 positions; the multiplexer 91 by 4, 8 or 12 positions; and the multiplexer 92 by 16, 32 or 48 positions. For example, the presence of a one in the least significant bit position in the output of the exponent processing unit indicates that the two exponents EXP.sub.1 and EXP.sub.2 differ by at least one, and thus the fraction FRAC.sub.1 or FRAC.sub.2 is shifted by one position in the multiplexer 90. The shifted number is then passed on to the 4-bit-shift multiplexer 91 which shifts the number by bit positions according to the two control inputs received by that multiplexer. For example, the presence of a one in the more significant bit position in the two inputs to the multiplexer 91 indicates that the two exponents EXP.sub.1 and EXP.sub.2 still differ by at least eight. Consequently, the fraction FRAC.sub.1 or FRAC.sub.2 received by the multiplexer 91 is shifted by eight additional bit positions in response to this particular control input signal.

When the two exponents EXP.sub.1 and EXP.sub.2 are equal, both the AND gates 82 and 84 produce a zero output, and thus zeros are supplied to all six control inputs to the multiplexers 90-92. Consequently, the two fractions FRAC.sub.1 and FRAC.sub.2 are not shifted at all.

If ones are present on all six control inputs to the multiplexers 90-92, the fraction FRAC.sub.1 or FRAC.sub.2 will be shifted a total of 63 positions. Whenever one of the fractions FRAC.sub.1 or FRAC.sub.2 is shifted in one of the alignment units 12 or 13, there is a possibility that a one to the right of the least significant bit of the shifted fraction will be "thrown away" as a result of truncation of that fraction. When this occurs, the alignment unit 12 or 13 in which the shift is effected generates a "sticky bit" signal SB.sub.1 or SB.sub.2 which is used later to compensate for the loss of the bit by truncation. The circuitry for generating this signal SB.sub.1 or SB.sub.2 in the alignment units 12 and 13 is illustrated in FlG. 6. The fraction shifter 99 in this circuit is the shifter shown in FIG. 5.

In the alignment unit of FIG. 6, the same six bits supplied as control signals to the multiplexers 90-92 in the shifter, to control up to 63 bit shifts, are supplied to a comparator 100 as input A. Input B to the comparator 100 is derived from a trailing-zeros detector 101 which determines the number of consecutive zeros in the 64bit fraction, beginning at the least significant bit position. If input B is less than input A, then the number of bit shifts exceeds the number of trailing zeros, which means that a one has been lost in the truncation of the fraction being shifted. Consequently, the comparator 100 asserts an output signal which is the sticky bit signal SB.

The sticky bit signal SB1 from the alignment unit 12 for FRAC.sub.1 is supplied to the NOR gate 45 that applies the signal SUBFRAC to the carry-in of the fraction adder 40. When the signal SB.sub.1 is asserted, it blocks the application of a carry-in signal to the adder 40, thereby compensating for the truncation of a set bit in the alignment unit 12.

The sticky bit signal from the alignment unit 13 for FRAC.sub.2 is supplied to the NOR gate 54 that applies the signal -/+ to the carry-in of the rounding adder 50. When the signal SB.sub.2 is asserted, it blocks the application of a carry-in signal to the adder 50, thereby compensating for the truncation of a set bit in the alignment unit 13.

Preferred circuitry for the trailing zeros detector 101 is shown in FIG. 7. In this circuit, successive bytes of the 64-bit fraction are supplied to eight different 8-bit priority encoders 110-117, and to eight different OR gates 120-127. The outputs of the OR gates 120-127 are all supplied to an 8-bit priority encoder 130 and to an OR gate 131. Each of the encoders produces a 3-bit output representing the position of the first one in its input byte, according to the truth table in FIG. 8. These eight 3-bit outputs are all supplied to a multiplexer 132. The encoder 130 produces a 3-bit output representing the position of the first one in its 8-bit input, and this 3-bit output is supplied to the select input of the multiplexer 132. Thus, the encoder 130 selects the first of the encoders 110-117 to detect a one, beginning with the encoder 110 which receives the least significant byte, and the multiplexer passes the 3-bit output of that selected encoder. Then the output of the multiplexer combined with the output of the encoder 130, as illustrated in FIG. 7, form a 6-bit binary number identifying the bit position of the first one in the 64-bit fraction. This number also identifies the number of trailing zeros.

It will be appreciated that the trailing-zeros-detector circuit shown in FIG. 7 can also be used as a leading zeros detector by simply mirroring the 64-bit input.