supermario/base/SuperMarioProj.1994-02-09/OS/FPUEmulation/Log.a

727 lines
22 KiB
Plaintext
Raw Normal View History

2019-06-29 15:17:50 +00:00
;
; File: Log.a
;
; Contains: Routines to emulate logarithmic functions
;
; Originally Written by: Motorola Inc.
; Adapted to Apple/MPW: Jon Okada
;
; Copyright: © 1990,1991 by Apple Computer, Inc., all rights reserved.
;
; This file is used in these builds: Mac32
;
; Change History (most recent first):
;
; <3> 4/13/91 BG Modified FLOG2 emulation to not signal inexact on exact cases.
; <2> 3/30/91 BG Rolling in Jon Okada's latest changes.
; <1> 12/14/90 BG First checked into TERROR/BBS.
; log.a
; Based upon Motorola files 'slogn.sa' and 'slog2.sa'.
; CHANGE LOG:
; 04 Jan 91 JPO Changed constant names BOUNDS1 and BOUNDS2 to BND1LOG
; and BND2LOG, respectively. Moved all slogn, slog2, and
; slog10 constants and table LOGTBL to file 'constants.a'.
; Changed variable names X, XDCARE, and XFRAC to XLN,
; XLNDC, and XLNFR, respectively. Deleted unreferenced
; label "HiX_0".
; 28 Mar 91 JPO Modified routines "slog2d" and "slog2" to handle exact
; cases of FLOG2. Streamlined "slognd" and "slogn"
; routines (FLOGN). Created a separate subroutine,
; "lognrm", to normalize subnormal input.
;
; slogn
*
* slogn.sa 3.1 12/10/90
*
* slogn computes the natural logarithm of an
* input value. slognd does the same except the input value is a
* denormalized number. slognp1 computes log(1+X), and slognp1d
* computes log(1+X) for denormalized X.
*
* Input: Double-extended value in memory location pointed to by address
* register a0.
*
* Output: log(X) or log(1+X) returned in floating-point register Fp0.
*
* Accuracy and Monotonicity: The returned result is within 2 ulps in
* 64 significant bit, i.e. within 0.5001 ulp to 53 bits if the
* result is subsequently rounded to double precision. The
* result is provably monotonic in double precision.
*
* Speed: The program slogn takes approximately 190 cycles for input
* argument X such that |X-1| >= 1/16, which is the the usual
* situation. For those arguments, slognp1 takes approximately
* 210 cycles. For the less common arguments, the program will
* run no worse than 10% slower.
*
* Algorithm:
* LOGN:
* Step 1. If |X-1| < 1/16, approximate log(X) by an odd polynomial in
* u, where u = 2(X-1)/(X+1). Otherwise, move on to Step 2.
*
* Step 2. X = 2**k * Y where 1 <= Y < 2. Define F to be the first seven
* significant bits of Y plus 2**(-7), i.e. F = 1.xxxxxx1 in base
* 2 where the six "x" match those of Y. Note that |Y-F| <= 2**(-7).
*
* Step 3. Define u = (Y-F)/F. Approximate log(1+u) by a polynomial in u,
* log(1+u) = poly.
*
* Step 4. Reconstruct log(X) = log( 2**k * Y ) = k*log(2) + log(F) + log(1+u)
* by k*log(2) + (log(F) + poly). The values of log(F) are calculated
* beforehand and stored in the program.
*
* lognp1:
* Step 1: If |X| < 1/16, approximate log(1+X) by an odd polynomial in
* u where u = 2X/(2+X). Otherwise, move on to Step 2.
*
* Step 2: Let 1+X = 2**k * Y, where 1 <= Y < 2. Define F as done in Step 2
* of the algorithm for LOGN and compute log(1+X) as
* k*log(2) + log(F) + poly where poly approximates log(1+u),
* u = (Y-F)/F.
*
* Implementation Notes:
* Note 1. There are 64 different possible values for F, thus 64 log(F)'s
* need to be tabulated. Moreover, the values of 1/F are also
* tabulated so that the division in (Y-F)/F can be performed by a
* multiplication.
*
* Note 2. In Step 2 of lognp1, in order to preserved accuracy, the value
* Y-F has to be calculated carefully when 1/2 <= X < 3/2.
*
* Note 3. To fully exploit the pipeline, polynomials are usually separated
* into two parts evaluated independently before being added up.
*
* Copyright (C) Motorola, Inc. 1990
* All Rights Reserved
*
* THIS IS UNPUBLISHED PROPRIETARY SOURCE CODE OF MOTOROLA
* The copyright notice above does not evidence any
* actual or intended publication of such source code.
* slogn IDNT 2,1 Motorola 040 Floating Point Software Package
ADJK equ L_SCR1
;X equ FP_SCR1 ; deleted <1/4/91, JPO>
;XDCARE equ X+2
;XFRAC equ X+4
XLN equ FP_SCR1 ; <1/4/91, JPO>
XLNDC equ XLN+2
XLNFR equ XLN+4
F equ FP_SCR2
FFRAC equ F+4
KLOG2 equ FP_SCR3
SAVEU equ FP_SCR4
slognd:
*--ENTRY POINT FOR LOG(X) FOR DENORMALIZED INPUT
; MOVE.L #-100,ADJK(a6) ...INPUT = 2^(ADJK) * FP0 - deleted <3/28/91, JPO> <T3>
tst.l (a0) ; invalid if negative operand <3/28/91, JPO> <T3>
bmi t_operr ; <3/28/91, JPO> <T3>
bsr.b lognrm ; normalize input and initialize ADJK <3/28/91, JPO> <T3>
bra.b LOGBGN ; continue below <3/28/91, JPO> <T3>
; NEW SUBROUTINE - <3/28/91, JPO>
; Subroutine lognrm---normalizes the positive extended denormal input at (a0) and
; writes result with zero exponent to XLN(a6). The negative exponent adjustment
; is written to ADJK(a6).
; On input: a0 points to (assumed positive) input argument.
; On output: XLN(a6) contains normal argument with zero exponent.
; ADJK(a6) contains the negative exponent adjustment.
; a0 points to XLN(a6).
*----normalize the input value by left shifting k bits (k to be determined
*----below), adjusting exponent and storing -k to ADJK
*----the value TWOTO100 is no longer needed.
*----Note that this code assumes the denormalized input is NON-ZERO.
lognrm: ; label added <3/28/91, JPO> <T3> thru next <T3>
; MoveM.L D2-D7,-(A7) ...save some registers - DELETED <3/28/91, JPO>
; Move.L #$00000000,D3 ...D3 is exponent of smallest norm. # - DELETED <3/28/91, JPO>
movem.l D3-D7,-(a7) ; save 5 D registers <3/28/91, JPO>
Move.L 4(A0),D4
Move.L 8(A0),D5 ...(D4,D5) is (Hi_X,Lo_X)
; Clr.L D2 ...D2 used for holding K - DELETED <3/28/91, JPO>
clr.l d3 ; d3 used for holding k <3/28/91, JPO>
; Tst.L D4 ; DELETED <3/28/91, JPO>
; BNE.B HiX_not0 ; DELETED <3/28/91, JPO>
bfffo D4{0:32},D6 ; find first set bit in Hi_X <3/28/91, JPO>
bne.b HiX_not0 ; HiX nonzero <3/28/91, JPO>
;HiX_0: ; label DELETED <1/4/91, JPO>
Move.L D5,D4
Clr.L D5
; Move.L #32,D2 ; DELETED <3/28/91, JPO>
moveq.l #32,d3 ; <3/28/91, JPO>
; Clr.L D6 ; DELETED <3/28/91, JPO>
BFFFO D4{0:32},D6
; LSL.L D6,D4 ; DELETED <3/28/91, JPO>
; Add.L D6,D2 ...(D3,D4,D5) is normalized - DELETED <3/28/91, JPO>
; Move.L D3,XLN(a6) ; <1/4/91, JPO> - DELETED <3/28/91, JPO>
; Move.L D4,XLNFR(a6) ; <1/4/91, JPO> - DELETED <3/28/91, JPO>
; Move.L D5,XLNFR+4(a6) ; <1/4/91, JPO> - DELETED <3/28/91, JPO>
; Neg.L D2 ; DELETED <3/28/91, JPO>
; Move.L D2,ADJK(a6) ; DELETED <3/28/91, JPO>
; FMove.X XLN(a6),FP0 ; <1/4/91, JPO> - DELETED <3/28/91, JPO>
; MoveM.L (A7)+,D2-D7 ...restore registers - DELETED <3/28/91, JPO>
; LEA XLN(a6),A0 ; <1/4/91, JPO> - DELETED <3/28/91, JPO>
; Bra.B LOGBGN ...begin regular log(X) - DELETED <3/28/91, JPO>
HiX_not0:
; Clr.L D6 ; DELETED <3/28/91, JPO>
; BFFFO D4{0:32},D6 ...find first 1 - DELETED <3/28/91, JPO>
; Move.L D6,D2 ...get k - DELETED <3/28/91, JPO>
LSL.L D6,D4
add.l d6,d3 ; get k <3/28/91, JPO>
Move.L D5,D7 ...a copy of D5
LSL.L D6,D5
; Neg.L D6 ; DELETED - <3/28/91, JPO>
; AddI.L #32,D6 ; DELETED - <3/28/91, JPO>
neg.b d6 ; do byte operations <3/28/91, JPO>
addi.b #32,d6 ; <3/28/91, JPO>
LSR.L D6,D7
neg.l d3 ; exponent adjust is -k <3/28/91, JPO>
Or.L D7,D4 ...(D3,D4,D5) normalized
; Move.L D3,XLN(a6) ; <1/4/91, JPO> - DELETED <3/28/91, JPO>
clr.l XLN(a6) ; zero exponent <3/28/91, JPO>
Move.L D4,XLNFR(a6) ; <1/4/91, JPO>
Move.L D5,XLNFR+4(a6) ; <1/4/91, JPO>
; Neg.L D2 ; DELETED <3/28/91, JPO>
; Move.L D2,ADJK(a6) ; DELETED <3/28/91, JPO>
move.l D3,ADJK(a6) ; store exponent adjust <3/28/91, JPO>
; FMove.X XLN(a6),FP0 ; <1/4/91, JPO> - DELETED <3/28/91, JPO>
; MoveM.L (A7)+,D2-D7 ...restore registers - DELETED <3/28/91, JPO>
movem.l (a7)+,d3-d7 ; restore 5 registers <3/28/91, JPO>
LEA XLN(a6),A0 ; <1/4/91, JPO>
; Bra.B LOGBGN ...begin regular log(X) - DELETED <3/28/91, JPO>
rts ; return <3/28/91, JPO>
slogn:
*--ENTRY POINT FOR LOG(X) FOR X FINITE, NON-ZERO, NOT NAN'S
tst.l (a0) ; invalid if negative operand
bmi t_operr
; FMOVE.X (A0),FP0 ...LOAD INPUT - moved below <3/28/91, JPO>
; MOVE.L #$00000000,ADJK(a6) ; DELETED <3/28/91, JPO>
clr.l ADJK(a6) ; zero exponent adjustment <3/28/91, JPO>
move.l (a0),XLN(a6) ; transfer normal operand to XLN(a6) <3/28/91, JPO>
move.l 4(a0),XLN+4(a6) ; <3/28/91, JPO>
move.l 8(a0),XLN+8(a6) ; <3/28/91, JPO>
lea.l XLN(a6),a0 ; a0 points to XLN(a6) <3/28/91, JPO> <T3>
LOGBGN:
*--FPCR SAVED AND CLEARED, INPUT IS 2^(ADJK)*FP0, FP0 CONTAINS
*--A FINITE, NON-ZERO, NORMALIZED NUMBER.
move.l (a0),d0
move.w 4(a0),d0
FMOVE.X (a0),FP0 ; FPO <- normal operand <3/28/91, JPO> <T3> thru next <T3>
; move.l (a0),XLN(a6) ; <1/4/91, JPO> - moved to above <3/28/91, JPO>
; move.l 4(a0),XLN+4(a6) ; <1/4/91, JPO> - <3/28/91, JPO>
; move.l 8(a0),XLN+8(a6) ; <1/4/91, JPO> - <3/28/91, JPO>
; CMPI.L #0,D0 ...CHECK IF X IS NEGATIVE - deleted <3/28/91, JPO>
; BLT.W LOGNEG ...LOG OF NEGATIVE ARGUMENT IS INVALID - deleted <3/28/91, JPO>
CMP2.L BND1LOG,D0 ...X IS POSITIVE, CHECK IF X IS NEAR 1
BCC.W LOGNEAR1 ...BOUNDS IS ROUGHLY [15/16, 17/16] <T3>
LOGMAIN:
*--THIS SHOULD BE THE USUAL CASE, X NOT VERY CLOSE TO 1
*--X = 2^(K) * Y, 1 <= Y < 2. THUS, Y = 1.XXXXXXXX....XX IN BINARY.
*--WE DEFINE F = 1.XXXXXX1, I.E. FIRST 7 BITS OF Y AND ATTACH A 1.
*--THE IDEA IS THAT LOG(X) = K*LOG2 + LOG(Y)
*-- = K*LOG2 + LOG(F) + LOG(1 + (Y-F)/F).
*--NOTE THAT U = (Y-F)/F IS VERY SMALL AND THUS APPROXIMATING
*--LOG(1+U) CAN BE VERY EFFICIENT.
*--ALSO NOTE THAT THE VALUE 1/F IS STORED IN A TABLE SO THAT NO
*--DIVISION IS NEEDED TO CALCULATE (Y-F)/F.
*--GET K, Y, F, AND ADDRESS OF 1/F.
ASR.L #8,D0
ASR.L #8,D0 ;...SHIFTED 16 BITS, BIASED EXPO. OF X
SUBI.L #$3FFF,D0 ;...THIS IS K
ADD.L ADJK(a6),D0 ;...ADJUST K, ORIGINAL INPUT MAY BE DENORM.
LEA LOGTBL,A0 ;...BASE ADDRESS OF 1/F AND LOG(F)
FMOVE.L D0,FP1 ;...CONVERT K TO FLOATING-POINT FORMAT
*--WHILE THE CONVERSION IS GOING ON, WE GET F AND ADDRESS OF 1/F
MOVE.L #$3FFF0000,XLN(a6) ;...X IS NOW Y, I.E. 2^(-K)*X <1/4/91, JPO>
MOVE.L XLNFR(a6),FFRAC(a6) ; <1/4/91, JPO>
ANDI.L #$FE000000,FFRAC(a6) ;...FIRST 7 BITS OF Y
ORI.L #$01000000,FFRAC(a6) ;...GET F: ATTACH A 1 AT THE EIGHTH BIT
MOVE.L FFRAC(a6),D0 ;...READY TO GET ADDRESS OF 1/F
ANDI.L #$7E000000,D0
ASR.L #8,D0
ASR.L #8,D0
ASR.L #4,D0 ;...SHIFTED 20, D0 IS THE DISPLACEMENT
ADDA.L D0,A0 ;...A0 IS THE ADDRESS FOR 1/F
FMOVE.X XLN(a6),FP0 ; <1/4/91, JPO>
move.l #$3fff0000,F(a6)
clr.l F+8(a6)
FSUB.X F(a6),FP0 ;...Y-F
FMOVEm.X FP2/fp3,-(sp) ;...SAVE FP2 WHILE FP0 IS NOT READY
*--SUMMARY: FP0 IS Y-F, A0 IS ADDRESS OF 1/F, FP1 IS K
*--REGISTERS SAVED: FPCR, FP1, FP2
LP1CONT1:
*--AN RE-ENTRY POINT FOR LOGNP1
FMUL.X (A0),FP0 ...FP0 IS U = (Y-F)/F
FMUL.X LOGOF2,FP1 ...GET K*LOG2 WHILE FP0 IS NOT READY
FMOVE.X FP0,FP2
FMUL.X FP2,FP2 ;...FP2 IS V=U*U
FMOVE.X FP1,KLOG2(a6) ;...PUT K*LOG2 IN MEMEORY, FREE FP1
*--LOG(1+U) IS APPROXIMATED BY
*--U + V*(A1+U*(A2+U*(A3+U*(A4+U*(A5+U*A6))))) WHICH IS
*--[U + V*(A1+V*(A3+V*A5))] + [U*V*(A2+V*(A4+V*A6))]
FMOVE.X FP2,FP3
FMOVE.X FP2,FP1
FMUL.D LOGA6,FP1 ;...V*A6
FMUL.D LOGA5,FP2 ;...V*A5
FADD.D LOGA4,FP1 ;...A4+V*A6
FADD.D LOGA3,FP2 ;...A3+V*A5
FMUL.X FP3,FP1 ;...V*(A4+V*A6)
FMUL.X FP3,FP2 ;...V*(A3+V*A5)
FADD.D LOGA2,FP1 ;...A2+V*(A4+V*A6)
FADD.D LOGA1,FP2 ;...A1+V*(A3+V*A5)
FMUL.X FP3,FP1 ;...V*(A2+V*(A4+V*A6))
ADDA.L #16,A0 ;...ADDRESS OF LOG(F)
FMUL.X FP3,FP2 ;...V*(A1+V*(A3+V*A5)), FP3 RELEASED
FMUL.X FP0,FP1 ;...U*V*(A2+V*(A4+V*A6))
FADD.X FP2,FP0 ;...U+V*(A1+V*(A3+V*A5)), FP2 RELEASED
FADD.X (A0),FP1 ;...LOG(F)+U*V*(A2+V*(A4+V*A6))
FMOVEm.X (sp)+,FP2/fp3 ;...RESTORE FP2
FADD.X FP1,FP0 ;...FP0 IS LOG(F) + LOG(1+U)
fmove.l d1,fpcr
FADD.X KLOG2(a6),FP0 ;...FINAL ADD
bra t_frcinx
LOGNEAR1:
*--REGISTERS SAVED: FPCR, FP1. FP0 CONTAINS THE INPUT.
FMOVE.X FP0,FP1
FSUB.S one,FP1 ;...FP1 IS X-1
FADD.S one,FP0 ;...FP0 IS X+1
FADD.X FP1,FP1 ;...FP1 IS 2(X-1)
*--LOG(X) = LOG(1+U/2)-LOG(1-U/2) WHICH IS AN ODD POLYNOMIAL
*--IN U, U = 2(X-1)/(X+1) = FP1/FP0
LP1CONT2:
*--THIS IS AN RE-ENTRY POINT FOR LOGNP1
FDIV.X FP0,FP1 ;...FP1 IS U
FMOVEm.X FP2/fp3,-(sp) ;...SAVE FP2
*--REGISTERS SAVED ARE NOW FPCR,FP1,FP2,FP3
*--LET V=U*U, W=V*V, CALCULATE
*--U + U*V*(B1 + V*(B2 + V*(B3 + V*(B4 + V*B5)))) BY
*--U + U*V*( [B1 + W*(B3 + W*B5)] + [V*(B2 + W*B4)] )
FMOVE.X FP1,FP0
FMUL.X FP0,FP0 ;...FP0 IS V
FMOVE.X FP1,SAVEU(a6) ;...STORE U IN MEMORY, FREE FP1
FMOVE.X FP0,FP1
FMUL.X FP1,FP1 ;...FP1 IS W
FMOVE.D LOGB5,FP3
FMOVE.D LOGB4,FP2
FMUL.X FP1,FP3 ;...W*B5
FMUL.X FP1,FP2 ;...W*B4
FADD.D LOGB3,FP3 ;...B3+W*B5
FADD.D LOGB2,FP2 ;...B2+W*B4
FMUL.X FP3,FP1 ;...W*(B3+W*B5), FP3 RELEASED
FMUL.X FP0,FP2 ;...V*(B2+W*B4)
FADD.D LOGB1,FP1 ;...B1+W*(B3+W*B5)
FMUL.X SAVEU(a6),FP0 ;...FP0 IS U*V
FADD.X FP2,FP1 ;...B1+W*(B3+W*B5) + V*(B2+W*B4), FP2 RELEASED
FMOVEm.X (sp)+,FP2/fp3 ;...FP2 RESTORED
FMUL.X FP1,FP0 ;...U*V*( [B1+W*(B3+W*B5)] + [V*(B2+W*B4)] )
fmove.l d1,fpcr
FADD.X SAVEU(a6),FP0
bra t_frcinx
rts
;LOGNEG: ; label DELETED <3/28/91, JPO>
;*--REGISTERS SAVED FPCR. LOG(-VE) IS INVALID
; bra t_operr ; DELETED <3/28/91, JPO>
slognp1d:
*--ENTRY POINT FOR LOG(1+Z) FOR DENORMALIZED INPUT
* Simply return the denorm
bra t_extdnrm
slognp1:
*--ENTRY POINT FOR LOG(1+X) FOR X FINITE, NON-ZERO, NOT NAN'S
FMOVE.X (A0),FP0 ...LOAD INPUT
fabs.x fp0 ;test magnitude
fcmp.x LTHOLD,fp0 ;compare with min threshold
fbgt.w LP1REAL ;if greater, continue
fmove.l #0,fpsr ;clr N flag from compare
fmove.l d1,fpcr
fmove.x (a0),fp0 ;return signed argument
bra t_frcinx
LP1REAL:
FMOVE.X (A0),FP0 ...LOAD INPUT
MOVE.L #$00000000,ADJK(a6)
FMOVE.X FP0,FP1 ...FP1 IS INPUT Z
FADD.S one,FP0 ...X := ROUND(1+Z)
FMOVE.X FP0,XLN(a6) ; <1/4/91, JPO>
MOVE.W XLNFR(a6),XLNDC(a6) ; <1/4/91, JPO>
MOVE.L XLN(a6),D0 ; <1/4/91, JPO>
; CMPI.L #0,D0 ; DELETED <3/28/91, JPO> <T3>
tst.l d0 ; <3/28/91, JPO> <T3>
BLE.W LP1NEG0 ...LOG OF ZERO OR -VE <T3>
CMP2.L BND2LOG,D0 ; <T3>
BCS.W LOGMAIN ...BND2LOG IS [1/2,3/2] <T3>
*--IF 1+Z > 3/2 OR 1+Z < 1/2, THEN X, WHICH IS ROUNDING 1+Z,
*--CONTAINS AT LEAST 63 BITS OF INFORMATION OF Z. IN THAT CASE,
*--SIMPLY INVOKE LOG(X) FOR LOG(1+Z).
LP1NEAR1:
*--NEXT SEE IF EXP(-1/16) < X < EXP(1/16)
CMP2.L BND1LOG,D0
BCS.B LP1CARE
LP1ONE16:
*--EXP(-1/16) < X < EXP(1/16). LOG(1+Z) = LOG(1+U/2) - LOG(1-U/2)
*--WHERE U = 2Z/(2+Z) = 2Z/(1+X).
FADD.X FP1,FP1 ...FP1 IS 2Z
FADD.S one,FP0 ...FP0 IS 1+X
*--U = FP1/FP0
BRA.W LP1CONT2
LP1CARE:
*--HERE WE USE THE USUAL TABLE DRIVEN APPROACH. CARE HAS TO BE
*--TAKEN BECAUSE 1+Z CAN HAVE 67 BITS OF INFORMATION AND WE MUST
*--PRESERVE ALL THE INFORMATION. BECAUSE 1+Z IS IN [1/2,3/2],
*--THERE ARE ONLY TWO CASES.
*--CASE 1: 1+Z < 1, THEN K = -1 AND Y-F = (2-F) + 2Z
*--CASE 2: 1+Z > 1, THEN K = 0 AND Y-F = (1-F) + Z
*--ON RETURNING TO LP1CONT1, WE MUST HAVE K IN FP1, ADDRESS OF
*--(1/F) IN A0, Y-F IN FP0, AND FP2 SAVED.
MOVE.L XLNFR(a6),FFRAC(a6) ; <1/4/91, JPO>
ANDI.L #$FE000000,FFRAC(a6)
ORI.L #$01000000,FFRAC(a6) ;...F OBTAINED
CMPI.L #$3FFF8000,D0 ;...SEE IF 1+Z > 1
BGE.B KISZERO
KISNEG1:
FMOVE.S TWO,FP0
move.l #$3FFF0000,F(a6)
clr.l F+8(a6)
FSUB.X F(a6),FP0 ;...2-F
MOVE.L FFRAC(a6),D0
ANDI.L #$7E000000,D0
ASR.L #8,D0
ASR.L #8,D0
ASR.L #4,D0 ;...D0 CONTAINS DISPLACEMENT FOR 1/F
FADD.X FP1,FP1 ;...GET 2Z
FMOVEm.X FP2/fp3,-(sp) ;...SAVE FP2
FADD.X FP1,FP0 ;...FP0 IS Y-F = (2-F)+2Z
LEA LOGTBL,A0 ;...A0 IS ADDRESS OF 1/F
ADDA.L D0,A0
FMOVE.S negone,FP1 ;...FP1 IS K = -1
BRA.W LP1CONT1
KISZERO:
FMOVE.S one,FP0
move.l #$3fff0000,F(a6)
clr.l F+8(a6)
FSUB.X F(a6),FP0 ;...1-F
MOVE.L FFRAC(a6),D0
ANDI.L #$7E000000,D0
ASR.L #8,D0
ASR.L #8,D0
ASR.L #4,D0
FADD.X FP1,FP0 ;...FP0 IS Y-F
FMOVEm.X FP2/fp3,-(sp) ;...FP2 SAVED
LEA LOGTBL,A0
ADDA.L D0,A0 ;...A0 IS ADDRESS OF 1/F
FMOVE.S zero,FP1 ;...FP1 IS K = 0
BRA.W LP1CONT1
LP1NEG0:
*--FPCR SAVED. D0 IS X IN COMPACT FORM.
; CMPI.L #0,D0 ; DELETED <3/28/91, JPO> <T3>
; BLT.B LP1NEG ; DELETED <3/28/91, JPO> <T3>
tst.l d0 ; <3/28/91, JPO> <T3>
bmi.b LP1NEG ; <3/28/91, JPO> <T3>
LP1ZERO:
FMOVE.S negone,FP0
fmove.l d1,fpcr
bra t_dz
LP1NEG:
FMOVE.S zero,FP0
fmove.l d1,fpcr
bra t_operr
; slog2
*
* slog2.sa 3.1 12/10/90
*
* The entry point slog10 computes the base-10
* logarithm of an input argument X.
* slog10d does the same except the input value is a
* denormalized number.
* sLog2 and sLog2d are the base-2 analogues.
*
* INPUT: Double-extended value in memory location pointed to
* by address register a0.
*
* OUTPUT: log_10(X) or log_2(X) returned in floating-point
* register fp0.
*
* ACCURACY and MONOTONICITY: The returned result is within 1.7
* ulps in 64 significant bit, i.e. within 0.5003 ulp
* to 53 bits if the result is subsequently rounded
* to double precision. The result is provably monotonic
* in double precision.
*
* SPEED: Two timings are measured, both in the copy-back mode.
* The first one is measured when the function is invoked
* the first time (so the instructions and data are not
* in cache), and the second one is measured when the
* function is reinvoked at the same input argument.
*
* ALGORITHM and IMPLEMENTATION NOTES:
*
* slog10d:
*
* Step 0. If X < 0, create a NaN and raise the invalid operation
* flag. Otherwise, save FPCR in D1; set FpCR to default.
* Notes: Default means round-to-nearest mode, no floating-point
* traps, and precision control = double extended.
*
* Step 1. Call slognd to obtain Y = log(X), the natural log of X.
* Notes: Even if X is denormalized, log(X) is always normalized.
*
* Step 2. Compute log_10(X) = log(X) * (1/log(10)).
* 2.1 Restore the user FPCR
* 2.2 Return ans := Y * INV_L10.
*
*
* slog10:
*
* Step 0. If X < 0, create a NaN and raise the invalid operation
* flag. Otherwise, save FPCR in D1; set FpCR to default.
* Notes: Default means round-to-nearest mode, no floating-point
* traps, and precision control = double extended.
*
* Step 1. Call sLogN to obtain Y = log(X), the natural log of X.
*
* Step 2. Compute log_10(X) = log(X) * (1/log(10)).
* 2.1 Restore the user FPCR
* 2.2 Return ans := Y * INV_L10.
*
*
* sLog2d:
*
* Step 0. If X < 0, create a NaN and raise the invalid operation
* flag. Otherwise, save FPCR in D1; set FpCR to default.
* Notes: Default means round-to-nearest mode, no floating-point
* traps, and precision control = double extended.
*
* Step 1. Call slognd to obtain Y = log(X), the natural log of X.
* Notes: Even if X is denormalized, log(X) is always normalized.
*
* Step 2. Compute log_2(X) = log(X) * (1/log(2)).
* 2.1 Restore the user FPCR
* 2.2 Return ans := Y * INV_L2.
*
*
* sLog2:
*
* Step 0. If X < 0, create a NaN and raise the invalid operation
* flag. Otherwise, save FPCR in D1; set FpCR to default.
* Notes: Default means round-to-nearest mode, no floating-point
* traps, and precision control = double extended.
*
* Step 1. If X is not an integer power of two, i.e., X != 2^k,
* go to Step 3.
*
* Step 2. Return k.
* 2.1 Get integer k, X = 2^k.
* 2.2 Restore the user FPCR.
* 2.3 Return ans := convert-to-double-extended(k).
*
* Step 3. Call sLogN to obtain Y = log(X), the natural log of X.
*
* Step 4. Compute log_2(X) = log(X) * (1/log(2)).
* 4.1 Restore the user FPCR
* 4.2 Return ans := Y * INV_L2.
*
* Copyright (C) Motorola, Inc. 1990
* All Rights Reserved
*
* THIS IS UNPUBLISHED PROPRIETARY SOURCE CODE OF MOTOROLA
* The copyright notice above does not evidence any
* actual or intended publication of such source code.
* SLOG2 IDNT 2,1 Motorola 040 Floating Point Software Package
slog10d:
*--entry point for Log10(X), X is denormalized
; move.l (a0),d0 ; DELETED <3/28/91, JPO> <T3> thru next <T3>
; blt.w invalid ; DELETED <3/28/91, JPO>
tst.l (a0) ; <3/28/91, JPO>
bmi t_operr ; <3/28/91, JPO>
move.l d1,-(sp)
clr.l d1
bsr slognd ...log(X), X denorm. <T3>
fmove.l (sp)+,fpcr
fmul.x INV_L10,fp0
bra t_frcinx
slog10:
*--entry point for Log10(X), X is normalized
; move.l (a0),d0 ; DELETED <3/28/91, JPO> <T3> thru next <T3>
; blt.b invalid ; DELETED <3/28/91, JPO>
tst.l (a0) ; <3/28/91, JPO>
bmi t_operr ; <3/28/91, JPO>
move.l d1,-(sp)
clr.l d1
bsr slogn ...log(X), X normal. <T3>
fmove.l (sp)+,fpcr
fmul.x INV_L10,fp0
bra t_frcinx
slog2d:
*--entry point for Log2(X), X is denormalized
; move.l (a0),d0 ; DELETED <3/28/91, JPO> <T3> thru next <T3>
; blt.b invalid ; DELETED <3/28/91, JPO>
tst.l (a0) ; <3/28/91, JPO>
bmi t_operr ; <3/28/91, JPO>
bsr lognrm ; normalize with exponent adjust in ADJK(A6) <3/28/91, JPO>
tst.l 8(a0) ; check for exact integral power of 2 <3/28/91, JPO>
bne.b @1 ; inexact <3/28/91, JPO>
move.l 4(a0),d0 ; <3/28/91, JPO>
lsl.l #1,d0 ; <3/28/91, JPO>
bne.b @1 ; inexact <3/28/91, JPO>
move.l ADJK(a6),d0 ; exact. get negative exponent into d0 <3/28/91, JPO>
bra.b slog2ex ; continue below <3/28/91, JPO>
@1: ; label added <3/28/91, JPO>
move.l d1,-(sp)
clr.l d1
; bsr slognd ...log(X), X denorm. - DELETED <3/28/91, JPO>
bsr LOGBGN ; log(X), X norm with negative ADJK <T3>
fmove.l (sp)+,fpcr
fmul.x INV_L2,fp0
bra t_frcinx
slog2:
*--entry point for Log2(X), X is normalized
; move.l (a0),d0 ; DELETED <3/28/91, JPO> <T3> thru next <T3>
; blt.b invalid ; DELETED <3/28/91, JPO>
tst.l (a0) ; <3/28/91, JPO>
bmi t_operr ; <3/28/91, JPO>
; move.l 8(a0),d0 ; DELETED <3/28/91, JPO>
tst.l 8(a0)
bne.b continue ...X is not 2^k
move.l 4(a0),d0
and.l #$7FFFFFFF,d0
; tst.l d0 ; DELETED <3/28/91, JPO> <T3>
bne.b continue
*--X = 2^k.
move.w (a0),d0
slog2ex: ; label ADDED <3/28/91, JPO> <T3> thru next <T3>
; and.l #$00007FFF,d0 ; DELETED <3/28/91, JPO>
; sub.l #$3FFF,d0 ; DELETED <3/28/91, JPO>
sub.w #$3FFF,d0 ; word operation sufficient <3/28/91, JPO>
fmove.l d1,fpcr
; fmove.l d0,fp0 ; DELETED <3/28/91, JPO>
fmove.w d0,fp0 ; exact result <3/28/91, JPO>
; bra t_frcinx ; DELETED
rts ; return exact result <3/28/91, JPO> <T3>
continue:
move.l d1,-(sp)
clr.l d1
bsr slogn ;...log(X), X normal.
fmove.l (sp)+,fpcr
fmul.x INV_L2,fp0
bra t_frcinx
;invalid: ; label DELETED <3/28/91, JPO> <T3>
; bra t_operr ; DELETED <3/28/91, JPO> <T3>