| 1 | # |
| 2 | # Copyright (c) 1987 Regents of the University of California. |
| 3 | # |
| 4 | # Use and reproduction of this software are granted in accordance with |
| 5 | # the terms and conditions specified in the Berkeley Software License |
| 6 | # Agreement (in particular, this entails acknowledgement of the programs' |
| 7 | # source, and inclusion of this notice) with the additional understanding |
| 8 | # that all recipients should regard themselves as participants in an |
| 9 | # ongoing research project and hence should feel obligated to report |
| 10 | # their experiences (good or bad) with these elementary function codes, |
| 11 | # using "sendbug 4bsd-bugs@BERKELEY", to the authors. |
| 12 | # |
| 13 | .data |
| 14 | .align 2 |
| 15 | _sccsid: |
| 16 | .asciz "@(#)cbrt.s 1.1 (ucb.elefunt) %G%" |
| 17 | |
| 18 | # double cbrt(double arg) |
| 19 | # W. Kahan, 10/13/80. revised 1/13/84 for keeping sign symmetry |
| 20 | # Revised and tested by Z. Alex Liu (7/11/87) |
| 21 | # Max error less than 0.667 ULPs _if_ +,-,*,/ were all correctly rounded... |
| 22 | .globl _cbrt |
| 23 | .globl _d_cbrt |
| 24 | .globl _dcbrt_ |
| 25 | .text |
| 26 | .align 1 |
| 27 | _cbrt: |
| 28 | _d_cbrt: |
| 29 | .word 0x01fc # save r2 to r8 |
| 30 | movl 4(fp),r0 # r0:r1 = x |
| 31 | movl 8(fp),r1 |
| 32 | brb 1f |
| 33 | _dcbrt_: |
| 34 | .word 0x01fc # save r2 to r8 |
| 35 | movl 4(fp),r8 |
| 36 | movl (r8),r0 |
| 37 | movl 4(r8),r1 # r0:r1 = x |
| 38 | |
| 39 | 1: andl3 $0x7f800000,r0,r2 # biased exponent of x |
| 40 | beql return # dcbrt(0)=0 dcbrt(res)=res. operand |
| 41 | andl3 $0x80000000,r0,r8 # r8 has sign(x) |
| 42 | xorl2 r8,r0 # r0 is abs(x) |
| 43 | movl r0,r2 # r2 has abs(x) |
| 44 | divl2 $3,r2 # rough dcbrt with bias/3 |
| 45 | addl2 B,r2 # restore bias, diminish fraction |
| 46 | ldf r2 # acc = |q|=|dcbrt| to 5 bits |
| 47 | mulf r2 # acc = qq |
| 48 | divf r0 # acc = qq/|x| |
| 49 | mulf r2 # acc = qqq/|x| |
| 50 | addf C # acc = C+qqq/|x| |
| 51 | stf r3 # r3 = s = C+qqq/|x| |
| 52 | ldf D # acc = D |
| 53 | divf r3 # acc = D/s |
| 54 | addf E # acc = E+D/s |
| 55 | addf r3 # acc = s+E+D/s |
| 56 | stf r3 # r3 = s+E+D/s |
| 57 | ldf F # acc = F |
| 58 | divf r3 # acc = F/(s+E+D/s) |
| 59 | addf G # acc = G+F/(s+E+D/s) |
| 60 | mulf r2 # acc = q*(G+F/(s+E+D/s)) = new q to 23 bits |
| 61 | stf r2 # r2 = q*(G+F/(s+E+D/s)) = new q to 23 bits |
| 62 | clrl r3 # r2:r3 = q as double float |
| 63 | ldd r2 # acc = q as double float |
| 64 | muld r2 # acc = qq exactly |
| 65 | std r4 # r4:r5 = qq exactly |
| 66 | ldd r0 # acc = |x| |
| 67 | divd r4 # acc = |x|/(q*q) rounded |
| 68 | std r0 # r0:r1 = |x|/(q*q) rounded |
| 69 | subd r2 # acc = |x|/(q*q)-q exactly |
| 70 | std r6 # r6:r7 = |x|/(q*q)-q exactly |
| 71 | movl r2,r4 |
| 72 | clrl r5 # r4:r5 = q as double float |
| 73 | addl2 $0x800000,r4 # r4:r5 = 2*q |
| 74 | ldd r4 # acc = 2*q |
| 75 | addd r0 # acc = 2*q+|x|/(q*q) |
| 76 | std r4 # r4:r5 = 2*q+|x|/(q*q) |
| 77 | ldd r6 # acc = |x|/(q*q)-q |
| 78 | divd r4 # acc = (|x|/(q*q)-q)/(2*q+|x|/(q*q)) |
| 79 | muld r2 # acc = q*(|x|/(q*q)-q)/(2*q+|x|/(q*q)) |
| 80 | addd r2 # acc = q+q*(|x|/(q*q)-q)/(2*q+|x|/(q*q)) |
| 81 | std r0 # r0:r1 = |result| |
| 82 | orl2 r8,r0 # restore the sign bit |
| 83 | return: ret # error less than 0.667ULPs? |
| 84 | |
| 85 | .data |
| 86 | .align 2 |
| 87 | B : .long 721142941 #(86-0.03306235651)*(2^23) |
| 88 | .align 2 |
| 89 | C: .long 0x400af8b0 #.float 0f0.5428571429 # 19/35 |
| 90 | .align 2 |
| 91 | D: .long 0xc0348ef1 #.float 0f-0.7053061224 # -864/1225 |
| 92 | .align 2 |
| 93 | E: .long 0x40b50750 #.float 0f1.414285714 # 99/70 |
| 94 | .align 2 |
| 95 | F: .long 0x40cdb6db #.float 0f1.607142857 # 45/28 |
| 96 | .align 2 |
| 97 | G: .long 0x3fb6db6e #.float 0f0.3571428571 # 5/14 |