# # Copyright (c) 1985 Regents of the University of California. # # Use and reproduction of this software are granted in accordance with # the terms and conditions specified in the Berkeley Software License # Agreement (in particular, this entails acknowledgement of the programs' # source, and inclusion of this notice) with the additional understanding # that all recipients should regard themselves as participants in an # ongoing research project and hence should feel obligated to report # their experiences (good or bad) with these elementary function codes, # using "sendbug 4bsd-bugs@BERKELEY", to the authors. # # @(#)argred.s 1.1 (Berkeley) 8/21/85 # libm$argred implements Bob Corbett's argument reduction and # libm$sincos implements Peter Tang's double precision sin/cos. # # Note: The two entry points libm$argred and libm$sincos are meant # to be used only by _sin, _cos and _tan. # # method: true range reduction to [-pi/4,pi/4], P. Tang & B. Corbett # S. McDonald, April 4, 1985 # .globl libm$argred .globl libm$sincos .text .align 1 libm$argred: # # Compare the argument with the largest possible that can # be reduced by table lookup. r3 := |x| will be used in table_lookup . # movd r0,r3 bgeq abs1 mnegd r3,r3 abs1: cmpd r3,$0d+4.55530934770520019583e+01 blss small_arg jsb trigred rsb small_arg: jsb table_lookup rsb # # At this point, # r0 contains the quadrant number, 0, 1, 2, or 3; # r2/r1 contains the reduced argument as a D-format number; # r3 contains a F-format extension to the reduced argument; # r4 contains a 0 or 1 corresponding to a sin or cos entry. # libm$sincos: # # Compensate for a cosine entry by adding one to the quadrant number. # addl2 r4,r0 # # Polyd clobbers r5-r0 ; save X in r7/r6 . # This can be avoided by rewriting trigred . # movd r1,r6 # # Likewise, save alpha in r8 . # This can be avoided by rewriting trigred . # movf r3,r8 # # Odd or even quadrant? cosine if odd, sine otherwise. # Save floor(quadrant/2) in r9 ; it determines the final sign. # rotl $-1,r0,r9 blss cosine sine: muld2 r1,r1 # Xsq = X * X polyd r1,$7,sin_coef # Q = P(Xsq) , of deg 7 mulf3 $0f3.0,r8,r4 # beta = 3 * alpha mulf2 r0,r4 # beta = Q * beta addf2 r8,r4 # beta = alpha + beta muld2 r6,r0 # S(X) = X * Q # cvtfd r4,r4 ... r5 = 0 after a polyd. addd2 r4,r0 # S(X) = beta + S(X) addd2 r6,r0 # S(X) = X + S(X) brb done cosine: muld2 r6,r6 # Xsq = X * X beql zero_arg mulf2 r1,r8 # beta = X * alpha polyd r6,$7,cos_coef # Q = P'(Xsq) , of deg 7 subd3 r0,r8,r0 # beta = beta - Q subw2 $0x80,r6 # Xsq = Xsq / 2 addd2 r0,r6 # Xsq = Xsq + beta zero_arg: subd3 r6,$0d1.0,r0 # C(X) = 1 - Xsq done: blbc r9,even mnegd r0,r0 even: rsb .data .align 2 sin_coef: .double 0d-7.53080332264191085773e-13 # s7 = 2^-29 -1.a7f2504ffc49f8.. .double 0d+1.60573519267703489121e-10 # s6 = 2^-21 1.611adaede473c8.. .double 0d-2.50520965150706067211e-08 # s5 = 2^-1a -1.ae644921ed8382.. .double 0d+2.75573191800593885716e-06 # s4 = 2^-13 1.71de3a4b884278.. .double 0d-1.98412698411850507950e-04 # s3 = 2^-0d -1.a01a01a0125e7d.. .double 0d+8.33333333333325688985e-03 # s2 = 2^-07 1.11111111110e50 .double 0d-1.66666666666666664354e-01 # s1 = 2^-03 -1.55555555555554 .double 0d+0.00000000000000000000e+00 # s0 = 0 cos_coef: .double 0d-1.13006966202629430300e-11 # s7 = 2^-25 -1.8D9BA04D1374BE.. .double 0d+2.08746646574796004700e-09 # s6 = 2^-1D 1.1EE632650350BA.. .double 0d-2.75573073031284417300e-07 # s5 = 2^-16 -1.27E4F31411719E.. .double 0d+2.48015872682668025200e-05 # s4 = 2^-10 1.A01A0196B902E8.. .double 0d-1.38888888888464709200e-03 # s3 = 2^-0A -1.6C16C16C11FACE.. .double 0d+4.16666666666664761400e-02 # s2 = 2^-05 1.5555555555539E .double 0d+0.00000000000000000000e+00 # s1 = 0 .double 0d+0.00000000000000000000e+00 # s0 = 0 # # Multiples of pi/2 expressed as the sum of three doubles, # # trailing: n * pi/2 , n = 0, 1, 2, ..., 29 # trailing[n] , # # middle: n * pi/2 , n = 0, 1, 2, ..., 29 # middle[n] , # # leading: n * pi/2 , n = 0, 1, 2, ..., 29 # leading[n] , # # where # leading[n] := (n * pi/2) rounded, # middle[n] := (n * pi/2 - leading[n]) rounded, # trailing[n] := (( n * pi/2 - leading[n]) - middle[n]) rounded . trailing: .double 0d+0.00000000000000000000e+00 # 0 * pi/2 trailing .double 0d+4.33590506506189049611e-35 # 1 * pi/2 trailing .double 0d+8.67181013012378099223e-35 # 2 * pi/2 trailing .double 0d+1.30077151951856714215e-34 # 3 * pi/2 trailing .double 0d+1.73436202602475619845e-34 # 4 * pi/2 trailing .double 0d-1.68390735624352669192e-34 # 5 * pi/2 trailing .double 0d+2.60154303903713428430e-34 # 6 * pi/2 trailing .double 0d-8.16726343231148352150e-35 # 7 * pi/2 trailing .double 0d+3.46872405204951239689e-34 # 8 * pi/2 trailing .double 0d+3.90231455855570147991e-34 # 9 * pi/2 trailing .double 0d-3.36781471248705338384e-34 # 10 * pi/2 trailing .double 0d-1.06379439835298071785e-33 # 11 * pi/2 trailing .double 0d+5.20308607807426856861e-34 # 12 * pi/2 trailing .double 0d+5.63667658458045770509e-34 # 13 * pi/2 trailing .double 0d-1.63345268646229670430e-34 # 14 * pi/2 trailing .double 0d-1.19986217995610764801e-34 # 15 * pi/2 trailing .double 0d+6.93744810409902479378e-34 # 16 * pi/2 trailing .double 0d-8.03640094449267300110e-34 # 17 * pi/2 trailing .double 0d+7.80462911711140295982e-34 # 18 * pi/2 trailing .double 0d-7.16921993148029483506e-34 # 19 * pi/2 trailing .double 0d-6.73562942497410676769e-34 # 20 * pi/2 trailing .double 0d-6.30203891846791677593e-34 # 21 * pi/2 trailing .double 0d-2.12758879670596143570e-33 # 22 * pi/2 trailing .double 0d+2.53800212047402350390e-33 # 23 * pi/2 trailing .double 0d+1.04061721561485371372e-33 # 24 * pi/2 trailing .double 0d+6.11729905311472319056e-32 # 25 * pi/2 trailing .double 0d+1.12733531691609154102e-33 # 26 * pi/2 trailing .double 0d-3.70049587943078297272e-34 # 27 * pi/2 trailing .double 0d-3.26690537292459340860e-34 # 28 * pi/2 trailing .double 0d-1.14812616507957271361e-34 # 29 * pi/2 trailing middle: .double 0d+0.00000000000000000000e+00 # 0 * pi/2 middle .double 0d+5.72118872610983179676e-18 # 1 * pi/2 middle .double 0d+1.14423774522196635935e-17 # 2 * pi/2 middle .double 0d-3.83475850529283316309e-17 # 3 * pi/2 middle .double 0d+2.28847549044393271871e-17 # 4 * pi/2 middle .double 0d-2.69052076007086676522e-17 # 5 * pi/2 middle .double 0d-7.66951701058566632618e-17 # 6 * pi/2 middle .double 0d-1.54628301484890040587e-17 # 7 * pi/2 middle .double 0d+4.57695098088786543741e-17 # 8 * pi/2 middle .double 0d+1.07001849766246313192e-16 # 9 * pi/2 middle .double 0d-5.38104152014173353044e-17 # 10 * pi/2 middle .double 0d-2.14622680169080983801e-16 # 11 * pi/2 middle .double 0d-1.53390340211713326524e-16 # 12 * pi/2 middle .double 0d-9.21580002543456677056e-17 # 13 * pi/2 middle .double 0d-3.09256602969780081173e-17 # 14 * pi/2 middle .double 0d+3.03066796603896507006e-17 # 15 * pi/2 middle .double 0d+9.15390196177573087482e-17 # 16 * pi/2 middle .double 0d+1.52771359575124969107e-16 # 17 * pi/2 middle .double 0d+2.14003699532492626384e-16 # 18 * pi/2 middle .double 0d-1.68853170360202329427e-16 # 19 * pi/2 middle .double 0d-1.07620830402834670609e-16 # 20 * pi/2 middle .double 0d+3.97700719404595604379e-16 # 21 * pi/2 middle .double 0d-4.29245360338161967602e-16 # 22 * pi/2 middle .double 0d-3.68013020380794313406e-16 # 23 * pi/2 middle .double 0d-3.06780680423426653047e-16 # 24 * pi/2 middle .double 0d-2.45548340466059054318e-16 # 25 * pi/2 middle .double 0d-1.84316000508691335411e-16 # 26 * pi/2 middle .double 0d-1.23083660551323675053e-16 # 27 * pi/2 middle .double 0d-6.18513205939560162346e-17 # 28 * pi/2 middle .double 0d-6.18980636588357585202e-19 # 29 * pi/2 middle leading: .double 0d+0.00000000000000000000e+00 # 0 * pi/2 leading .double 0d+1.57079632679489661351e+00 # 1 * pi/2 leading .double 0d+3.14159265358979322702e+00 # 2 * pi/2 leading .double 0d+4.71238898038468989604e+00 # 3 * pi/2 leading .double 0d+6.28318530717958645404e+00 # 4 * pi/2 leading .double 0d+7.85398163397448312306e+00 # 5 * pi/2 leading .double 0d+9.42477796076937979208e+00 # 6 * pi/2 leading .double 0d+1.09955742875642763501e+01 # 7 * pi/2 leading .double 0d+1.25663706143591729081e+01 # 8 * pi/2 leading .double 0d+1.41371669411540694661e+01 # 9 * pi/2 leading .double 0d+1.57079632679489662461e+01 # 10 * pi/2 leading .double 0d+1.72787595947438630262e+01 # 11 * pi/2 leading .double 0d+1.88495559215387595842e+01 # 12 * pi/2 leading .double 0d+2.04203522483336561422e+01 # 13 * pi/2 leading .double 0d+2.19911485751285527002e+01 # 14 * pi/2 leading .double 0d+2.35619449019234492582e+01 # 15 * pi/2 leading .double 0d+2.51327412287183458162e+01 # 16 * pi/2 leading .double 0d+2.67035375555132423742e+01 # 17 * pi/2 leading .double 0d+2.82743338823081389322e+01 # 18 * pi/2 leading .double 0d+2.98451302091030359342e+01 # 19 * pi/2 leading .double 0d+3.14159265358979324922e+01 # 20 * pi/2 leading .double 0d+3.29867228626928286062e+01 # 21 * pi/2 leading .double 0d+3.45575191894877260523e+01 # 22 * pi/2 leading .double 0d+3.61283155162826226103e+01 # 23 * pi/2 leading .double 0d+3.76991118430775191683e+01 # 24 * pi/2 leading .double 0d+3.92699081698724157263e+01 # 25 * pi/2 leading .double 0d+4.08407044966673122843e+01 # 26 * pi/2 leading .double 0d+4.24115008234622088423e+01 # 27 * pi/2 leading .double 0d+4.39822971502571054003e+01 # 28 * pi/2 leading .double 0d+4.55530934770520019583e+01 # 29 * pi/2 leading twoOverPi: .double 0d+6.36619772367581343076e-01 .text .align 1 table_lookup: muld3 r3,twoOverPi,r0 cvtrdl r0,r0 # n = nearest int to ((2/pi)*|x|) rnded mull3 $8,r0,r5 subd2 leading(r5),r3 # p = (|x| - leading n*pi/2) exactly subd3 middle(r5),r3,r1 # q = (p - middle n*pi/2) rounded subd2 r1,r3 # r = (p - q) subd2 middle(r5),r3 # r = r - middle n*pi/2 subd2 trailing(r5),r3 # r = r - trailing n*pi/2 rounded # # If the original argument was negative, # negate the reduce argument and # adjust the octant/quadrant number. # tstw 4(ap) bgeq abs2 mnegf r1,r1 mnegf r3,r3 # subb3 r0,$8,r0 ...used for pi/4 reduction -S.McD subb3 r0,$4,r0 abs2: # # Clear all unneeded octant/quadrant bits. # # bicb2 $0xf8,r0 ...used for pi/4 reduction -S.McD bicb2 $0xfc,r0 rsb # # p.0 .text .align 2 # # Only 256 (actually 225) bits of 2/pi are needed for VAX double # precision; this was determined by enumerating all the nearest # machine integer multiples of pi/2 using continued fractions. # (8a8d3673775b7ff7 required the most bits.) -S.McD # .long 0 .long 0 .long 0xaef1586d .long 0x9458eaf7 .long 0x10e4107f .long 0xd8a5664f .long 0x4d377036 .long 0x09d5f47d .long 0x91054a7f .long 0xbe60db93 bits2opi: .long 0x00000028 .long 0 # # Note: wherever you see the word `octant', read `quadrant'. # Currently this code is set up for pi/2 argument reduction. # By uncommenting/commenting the appropriate lines, it will # also serve as a pi/4 argument reduction code. # # p.1 # Trigred preforms argument reduction # for the trigonometric functions. It # takes one input argument, a D-format # number in r1/r0 . The magnitude of # the input argument must be greater # than or equal to 1/2 . Trigred produces # three results: the number of the octant # occupied by the argument, the reduced # argument, and an extension of the # reduced argument. The octant number is # returned in r0 . The reduced argument # is returned as a D-format number in # r2/r1 . An 8 bit extension of the # reduced argument is returned as an # F-format number in r3. # p.2 trigred: # # Save the sign of the input argument. # movw r0,-(sp) # # Extract the exponent field. # extzv $7,$7,r0,r2 # # Convert the fraction part of the input # argument into a quadword integer. # bicw2 $0xff80,r0 bisb2 $0x80,r0 # -S.McD rotl $16,r0,r0 rotl $16,r1,r1 # # If r1 is negative, add 1 to r0 . This # adjustment is made so that the two's # complement multiplications done later # will produce unsigned results. # bgeq posmid incl r0 posmid: # p.3 # # Set r3 to the address of the first quadword # used to obtain the needed portion of 2/pi . # The address is longword aligned to ensure # efficient access. # ashl $-3,r2,r3 bicb2 $3,r3 subl3 r3,$bits2opi,r3 # # Set r2 to the size of the shift needed to # obtain the correct portion of 2/pi . # bicb2 $0xe0,r2 # p.4 # # Move the needed 128 bits of 2/pi into # r11 - r8 . Adjust the numbers to allow # for unsigned multiplication. # ashq r2,(r3),r10 subl2 $4,r3 ashq r2,(r3),r9 bgeq signoff1 incl r11 signoff1: subl2 $4,r3 ashq r2,(r3),r8 bgeq signoff2 incl r10 signoff2: subl2 $4,r3 ashq r2,(r3),r7 bgeq signoff3 incl r9 signoff3: # p.5 # # Multiply the contents of r0/r1 by the # slice of 2/pi in r11 - r8 . # emul r0,r8,$0,r4 emul r0,r9,r5,r5 emul r0,r10,r6,r6 emul r1,r8,$0,r7 emul r1,r9,r8,r8 emul r1,r10,r9,r9 emul r1,r11,r10,r10 addl2 r4,r8 adwc r5,r9 adwc r6,r10 # p.6 # # If there are more than five leading zeros # after the first two quotient bits or if there # are more than five leading ones after the first # two quotient bits, generate more fraction bits. # Otherwise, branch to code to produce the result. # bicl3 $0xc1ffffff,r10,r4 beql more1 cmpl $0x3e000000,r4 bneq result more1: # p.7 # # generate another 32 result bits. # subl2 $4,r3 ashq r2,(r3),r5 bgeq signoff4 emul r1,r6,$0,r4 addl2 r1,r5 emul r0,r6,r5,r5 addl2 r0,r6 brb addbits1 signoff4: emul r1,r6,$0,r4 emul r0,r6,r5,r5 addbits1: addl2 r5,r7 adwc r6,r8 adwc $0,r9 adwc $0,r10 # p.8 # # Check for massive cancellation. # bicl3 $0xc0000000,r10,r6 # bneq more2 -S.McD Test was backwards beql more2 cmpl $0x3fffffff,r6 bneq result more2: # p.9 # # If massive cancellation has occurred, # generate another 24 result bits. # Testing has shown there will always be # enough bits after this point. # subl2 $4,r3 ashq r2,(r3),r5 bgeq signoff5 emul r0,r6,r4,r5 addl2 r0,r6 brb addbits2 signoff5: emul r0,r6,r4,r5 addbits2: addl2 r6,r7 adwc $0,r8 adwc $0,r9 adwc $0,r10 # p.10 # # The following code produces the reduced # argument from the product bits contained # in r10 - r7 . # result: # # Extract the octant number from r10 . # # extzv $29,$3,r10,r0 ...used for pi/4 reduction -S.McD extzv $30,$2,r10,r0 # # Clear the octant bits in r10 . # # bicl2 $0xe0000000,r10 ...used for pi/4 reduction -S.McD bicl2 $0xc0000000,r10 # # Zero the sign flag. # clrl r5 # p.11 # # Check to see if the fraction is greater than # or equal to one-half. If it is, add one # to the octant number, set the sign flag # on, and replace the fraction with 1 minus # the fraction. # # bitl $0x10000000,r10 ...used for pi/4 reduction -S.McD bitl $0x20000000,r10 beql small incl r0 incl r5 # subl3 r10,$0x1fffffff,r10 ...used for pi/4 reduction -S.McD subl3 r10,$0x3fffffff,r10 mcoml r9,r9 mcoml r8,r8 mcoml r7,r7 small: # p.12 # ## Test whether the first 29 bits of the ...used for pi/4 reduction -S.McD # Test whether the first 30 bits of the # fraction are zero. # tstl r10 beql tiny # # Find the position of the first one bit in r10 . # cvtld r10,r1 extzv $7,$7,r1,r1 # # Compute the size of the shift needed. # subl3 r1,$32,r6 # # Shift up the high order 64 bits of the # product. # ashq r6,r9,r10 ashq r6,r8,r9 brb mult # p.13 # # Test to see if the sign bit of r9 is on. # tiny: tstl r9 bgeq tinier # # If it is, shift the product bits up 32 bits. # movl $32,r6 movq r8,r10 tstl r10 brb mult # p.14 # # Test whether r9 is zero. It is probably # impossible for both r10 and r9 to be # zero, but until proven to be so, the test # must be made. # tinier: beql zero # # Find the position of the first one bit in r9 . # cvtld r9,r1 extzv $7,$7,r1,r1 # # Compute the size of the shift needed. # subl3 r1,$32,r1 addl3 $32,r1,r6 # # Shift up the high order 64 bits of the # product. # ashq r1,r8,r10 ashq r1,r7,r9 brb mult # p.15 # # The following code sets the reduced # argument to zero. # zero: clrl r1 clrl r2 clrl r3 brw return # p.16 # # At this point, r0 contains the octant number, # r6 indicates the number of bits the fraction # has been shifted, r5 indicates the sign of # the fraction, r11/r10 contain the high order # 64 bits of the fraction, and the condition # codes indicate where the sign bit of r10 # is on. The following code multiplies the # fraction by pi/2 . # mult: # # Save r11/r10 in r4/r1 . -S.McD movl r11,r4 movl r10,r1 # # If the sign bit of r10 is on, add 1 to r11 . # bgeq signoff6 incl r11 signoff6: # p.17 # # Move pi/2 into r3/r2 . # movq $0xc90fdaa22168c235,r2 # # Multiply the fraction by the portion of pi/2 # in r2 . # emul r2,r10,$0,r7 emul r2,r11,r8,r7 # # Multiply the fraction by the portion of pi/2 # in r3 . emul r3,r10,$0,r9 emul r3,r11,r10,r10 # # Add the product bits together. # addl2 r7,r9 adwc r8,r10 adwc $0,r11 # # Compensate for not sign extending r8 above.-S.McD # tstl r8 bgeq signoff6a decl r11 signoff6a: # # Compensate for r11/r10 being unsigned. -S.McD # addl2 r2,r10 adwc r3,r11 # # Compensate for r3/r2 being unsigned. -S.McD # addl2 r1,r10 adwc r4,r11 # p.18 # # If the sign bit of r11 is zero, shift the # product bits up one bit and increment r6 . # blss signon incl r6 ashq $1,r10,r10 tstl r9 bgeq signoff7 incl r10 signoff7: signon: # p.19 # # Shift the 56 most significant product # bits into r9/r8 . The sign extension # will be handled later. # ashq $-8,r10,r8 # # Convert the low order 8 bits of r10 # into an F-format number. # cvtbf r10,r3 # # If the result of the conversion was # negative, add 1 to r9/r8 . # bgeq chop incl r8 adwc $0,r9 # # If r9 is now zero, branch to special # code to handle that possibility. # beql carryout chop: # p.20 # # Convert the number in r9/r8 into # D-format number in r2/r1 . # rotl $16,r8,r2 rotl $16,r9,r1 # # Set the exponent field to the appropriate # value. Note that the extra bits created by # sign extension are now eliminated. # subw3 r6,$131,r6 insv r6,$7,$9,r1 # # Set the exponent field of the F-format # number in r3 to the appropriate value. # tstf r3 beql return # extzv $7,$8,r3,r4 -S.McD extzv $7,$7,r3,r4 addw2 r4,r6 # subw2 $217,r6 -S.McD subw2 $64,r6 insv r6,$7,$8,r3 brb return # p.21 # # The following code generates the appropriate # result for the unlikely possibility that # rounding the number in r9/r8 resulted in # a carry out. # carryout: clrl r1 clrl r2 subw3 r6,$132,r6 insv r6,$7,$9,r1 tstf r3 beql return extzv $7,$8,r3,r4 addw2 r4,r6 subw2 $218,r6 insv r6,$7,$8,r3 # p.22 # # The following code makes an needed # adjustments to the signs of the # results or to the octant number, and # then returns. # return: # # Test if the fraction was greater than or # equal to 1/2 . If so, negate the reduced # argument. # blbc r5,signoff8 mnegf r1,r1 mnegf r3,r3 signoff8: # p.23 # # If the original argument was negative, # negate the reduce argument and # adjust the octant number. # tstw (sp)+ bgeq signoff9 mnegf r1,r1 mnegf r3,r3 # subb3 r0,$8,r0 ...used for pi/4 reduction -S.McD subb3 r0,$4,r0 signoff9: # # Clear all unneeded octant bits. # # bicb2 $0xf8,r0 ...used for pi/4 reduction -S.McD bicb2 $0xfc,r0 # # Return. # rsb