Aren/NetBSD
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
47226343a9
NetBSD/sys/arch/m68k/fpsp/bindec.sa
briggs c869df6afb The new toolchain wants to generate relocatable branches to globally-defined 
symbols--even when the branches are local and would otherwise work with
bxx.b.  Compensate for this by shadowing the relevant labels with local
labels.
2001-12-09 01:43:13 +00:00

948 lines
29 KiB
Plaintext
Raw Blame History

* $NetBSD: bindec.sa,v 1.5 2001/12/09 01:43:13 briggs Exp $
* MOTOROLA MICROPROCESSOR & MEMORY TECHNOLOGY GROUP
* M68000 Hi-Performance Microprocessor Division
* M68040 Software Package
*
* M68040 Software Package Copyright (c) 1993, 1994 Motorola Inc.
* All rights reserved.
*
* THE SOFTWARE is provided on an "AS IS" basis and without warranty.
* To the maximum extent permitted by applicable law,
* MOTOROLA DISCLAIMS ALL WARRANTIES WHETHER EXPRESS OR IMPLIED,
* INCLUDING IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A
* PARTICULAR PURPOSE and any warranty against infringement with
* regard to the SOFTWARE (INCLUDING ANY MODIFIED VERSIONS THEREOF)
* and any accompanying written materials.
*
* To the maximum extent permitted by applicable law,
* IN NO EVENT SHALL MOTOROLA BE LIABLE FOR ANY DAMAGES WHATSOEVER
* (INCLUDING WITHOUT LIMITATION, DAMAGES FOR LOSS OF BUSINESS
* PROFITS, BUSINESS INTERRUPTION, LOSS OF BUSINESS INFORMATION, OR
* OTHER PECUNIARY LOSS) ARISING OF THE USE OR INABILITY TO USE THE
* SOFTWARE. Motorola assumes no responsibility for the maintenance
* and support of the SOFTWARE.
*
* You are hereby granted a copyright license to use, modify, and
* distribute the SOFTWARE so long as this entire notice is retained
* without alteration in any modified and/or redistributed versions,
* and that such modified versions are clearly identified as such.
* No licenses are granted by implication, estoppel or otherwise
* under any patents or trademarks of Motorola, Inc.
*
* bindec.sa 3.4 1/3/91
*
* bindec
*
* Description:
* Converts an input in extended precision format
* to bcd format.
*
* Input:
* a0 points to the input extended precision value
* value in memory; d0 contains the k-factor sign-extended
* to 32-bits. The input may be either normalized,
* unnormalized, or denormalized.
*
* Output: result in the FP_SCR1 space on the stack.
*
* Saves and Modifies: D2-D7,A2,FP2
*
* Algorithm:
*
* A1. Set RM and size ext; Set SIGMA = sign of input.
* The k-factor is saved for use in d7. Clear the
* BINDEC_FLG for separating normalized/denormalized
* input. If input is unnormalized or denormalized,
* normalize it.
*
* A2. Set X = abs(input).
*
* A3. Compute ILOG.
* ILOG is the log base 10 of the input value. It is
* approximated by adding e + 0.f when the original
* value is viewed as 2^^e * 1.f in extended precision.
* This value is stored in d6.
*
* A4. Clr INEX bit.
* The operation in A3 above may have set INEX2.
*
* A5. Set ICTR = 0;
* ICTR is a flag used in A13. It must be set before the
* loop entry A6.
*
* A6. Calculate LEN.
* LEN is the number of digits to be displayed. The
* k-factor can dictate either the total number of digits,
* if it is a positive number, or the number of digits
* after the decimal point which are to be included as
* significant. See the 68882 manual for examples.
* If LEN is computed to be greater than 17, set OPERR in
* USER_FPSR. LEN is stored in d4.
*
* A7. Calculate SCALE.
* SCALE is equal to 10^ISCALE, where ISCALE is the number
* of decimal places needed to insure LEN integer digits
* in the output before conversion to bcd. LAMBDA is the
* sign of ISCALE, used in A9. Fp1 contains
* 10^^(abs(ISCALE)) using a rounding mode which is a
* function of the original rounding mode and the signs
* of ISCALE and X. A table is given in the code.
*
* A8. Clr INEX; Force RZ.
* The operation in A3 above may have set INEX2.
* RZ mode is forced for the scaling operation to insure
* only one rounding error. The grs bits are collected in
* the INEX flag for use in A10.
*
* A9. Scale X -> Y.
* The mantissa is scaled to the desired number of
* significant digits. The excess digits are collected
* in INEX2.
*
* A10. Or in INEX.
* If INEX is set, round error occurred. This is
* compensated for by 'or-ing' in the INEX2 flag to
* the lsb of Y.
*
* A11. Restore original FPCR; set size ext.
* Perform FINT operation in the user's rounding mode.
* Keep the size to extended.
*
* A12. Calculate YINT = FINT(Y) according to user's rounding
* mode. The FPSP routine sintd0 is used. The output
* is in fp0.
*
* A13. Check for LEN digits.
* If the int operation results in more than LEN digits,
* or less than LEN -1 digits, adjust ILOG and repeat from
* A6. This test occurs only on the first pass. If the
* result is exactly 10^LEN, decrement ILOG and divide
* the mantissa by 10.
*
* A14. Convert the mantissa to bcd.
* The binstr routine is used to convert the LEN digit
* mantissa to bcd in memory. The input to binstr is
* to be a fraction; i.e. (mantissa)/10^LEN and adjusted
* such that the decimal point is to the left of bit 63.
* The bcd digits are stored in the correct position in
* the final string area in memory.
*
* A15. Convert the exponent to bcd.
* As in A14 above, the exp is converted to bcd and the
* digits are stored in the final string.
* Test the length of the final exponent string. If the
* length is 4, set operr.
*
* A16. Write sign bits to final string.
*
* Implementation Notes:
*
* The registers are used as follows:
*
* d0: scratch; LEN input to binstr
* d1: scratch
* d2: upper 32-bits of mantissa for binstr
* d3: scratch;lower 32-bits of mantissa for binstr
* d4: LEN
* d5: LAMBDA/ICTR
* d6: ILOG
* d7: k-factor
* a0: ptr for original operand/final result
* a1: scratch pointer
* a2: pointer to FP_X; abs(original value) in ext
* fp0: scratch
* fp1: scratch
* fp2: scratch
* F_SCR1:
* F_SCR2:
* L_SCR1:
* L_SCR2:
*
BINDEC IDNT 2,1 Motorola 040 Floating Point Software Package
include fpsp.h
section 8
* Constants in extended precision
LOG2 dc.l $3FFD0000,$9A209A84,$FBCFF798,$00000000
LOG2UP1 dc.l $3FFD0000,$9A209A84,$FBCFF799,$00000000
* Constants in single precision
FONE dc.l $3F800000,$00000000,$00000000,$00000000
FTWO dc.l $40000000,$00000000,$00000000,$00000000
FTEN dc.l $41200000,$00000000,$00000000,$00000000
F4933 dc.l $459A2800,$00000000,$00000000,$00000000
RBDTBL dc.b 0,0,0,0
dc.b 3,3,2,2
dc.b 3,2,2,3
dc.b 2,3,3,2
xref binstr
xref sintdo
xref ptenrn,ptenrm,ptenrp
xdef bindec
xdef sc_mul
bindec:
movem.l d2-d7/a2,-(a7)
fmovem.x fp0-fp2,-(a7)
* A1. Set RM and size ext. Set SIGMA = sign input;
* The k-factor is saved for use in d7. Clear BINDEC_FLG for
* separating normalized/denormalized input. If the input
* is a denormalized number, set the BINDEC_FLG memory word
* to signal denorm. If the input is unnormalized, normalize
* the input and test for denormalized result.
*
fmove.l #rm_mode,FPCR ;set RM and ext
move.l (a0),L_SCR2(a6) ;save exponent for sign check
move.l d0,d7 ;move k-factor to d7
clr.b BINDEC_FLG(a6) ;clr norm/denorm flag
move.w STAG(a6),d0 ;get stag
andi.w #$e000,d0 ;isolate stag bits
beq A2_str ;if zero, input is norm
*
* Normalize the denorm
*
un_de_norm:
move.w (a0),d0
andi.w #$7fff,d0 ;strip sign of normalized exp
move.l 4(a0),d1
move.l 8(a0),d2
norm_loop:
sub.w #1,d0
add.l d2,d2
addx.l d1,d1
tst.l d1
bge.b norm_loop
*
* Test if the normalized input is denormalized
*
tst.w d0
bgt.b pos_exp ;if greater than zero, it is a norm
st BINDEC_FLG(a6) ;set flag for denorm
pos_exp:
andi.w #$7fff,d0 ;strip sign of normalized exp
move.w d0,(a0)
move.l d1,4(a0)
move.l d2,8(a0)
* A2. Set X = abs(input).
*
A2_str:
move.l (a0),FP_SCR2(a6) ; move input to work space
move.l 4(a0),FP_SCR2+4(a6) ; move input to work space
move.l 8(a0),FP_SCR2+8(a6) ; move input to work space
andi.l #$7fffffff,FP_SCR2(a6) ;create abs(X)
* A3. Compute ILOG.
* ILOG is the log base 10 of the input value. It is approx-
* imated by adding e + 0.f when the original value is viewed
* as 2^^e * 1.f in extended precision. This value is stored
* in d6.
*
* Register usage:
* Input/Output
* d0: k-factor/exponent
* d2: x/x
* d3: x/x
* d4: x/x
* d5: x/x
* d6: x/ILOG
* d7: k-factor/Unchanged
* a0: ptr for original operand/final result
* a1: x/x
* a2: x/x
* fp0: x/float(ILOG)
* fp1: x/x
* fp2: x/x
* F_SCR1:x/x
* F_SCR2:Abs(X)/Abs(X) with $3fff exponent
* L_SCR1:x/x
* L_SCR2:first word of X packed/Unchanged
tst.b BINDEC_FLG(a6) ;check for denorm
beq.b A3_cont ;if clr, continue with norm
move.l #-4933,d6 ;force ILOG = -4933
bra.b A4_str
A3_cont:
move.w FP_SCR2(a6),d0 ;move exp to d0
move.w #$3fff,FP_SCR2(a6) ;replace exponent with 0x3fff
fmove.x FP_SCR2(a6),fp0 ;now fp0 has 1.f
sub.w #$3fff,d0 ;strip off bias
fadd.w d0,fp0 ;add in exp
fsub.s FONE,fp0 ;subtract off 1.0
fbge.w pos_res ;if pos, branch
fmul.x LOG2UP1,fp0 ;if neg, mul by LOG2UP1
fmove.l fp0,d6 ;put ILOG in d6 as a lword
bra.b A4_str ;go move out ILOG
pos_res:
fmul.x LOG2,fp0 ;if pos, mul by LOG2
fmove.l fp0,d6 ;put ILOG in d6 as a lword
* A4. Clr INEX bit.
* The operation in A3 above may have set INEX2.
A4_str:
fmove.l #0,FPSR ;zero all of fpsr - nothing needed
* A5. Set ICTR = 0;
* ICTR is a flag used in A13. It must be set before the
* loop entry A6. The lower word of d5 is used for ICTR.
clr.w d5 ;clear ICTR
* A6. Calculate LEN.
* LEN is the number of digits to be displayed. The k-factor
* can dictate either the total number of digits, if it is
* a positive number, or the number of digits after the
* original decimal point which are to be included as
* significant. See the 68882 manual for examples.
* If LEN is computed to be greater than 17, set OPERR in
* USER_FPSR. LEN is stored in d4.
*
* Register usage:
* Input/Output
* d0: exponent/Unchanged
* d2: x/x/scratch
* d3: x/x
* d4: exc picture/LEN
* d5: ICTR/Unchanged
* d6: ILOG/Unchanged
* d7: k-factor/Unchanged
* a0: ptr for original operand/final result
* a1: x/x
* a2: x/x
* fp0: float(ILOG)/Unchanged
* fp1: x/x
* fp2: x/x
* F_SCR1:x/x
* F_SCR2:Abs(X) with $3fff exponent/Unchanged
* L_SCR1:x/x
* L_SCR2:first word of X packed/Unchanged
A6_str:
tst.l d7 ;branch on sign of k
ble.b k_neg ;if k <= 0, LEN = ILOG + 1 - k
move.l d7,d4 ;if k > 0, LEN = k
bra.b len_ck ;skip to LEN check
k_neg:
move.l d6,d4 ;first load ILOG to d4
sub.l d7,d4 ;subtract off k
addq.l #1,d4 ;add in the 1
len_ck:
tst.l d4 ;LEN check: branch on sign of LEN
ble.b LEN_ng ;if neg, set LEN = 1
cmp.l #17,d4 ;test if LEN > 17
ble.b A7_str ;if not, forget it
move.l #17,d4 ;set max LEN = 17
tst.l d7 ;if negative, never set OPERR
ble.b A7_str ;if positive, continue
or.l #opaop_mask,USER_FPSR(a6) ;set OPERR & AIOP in USER_FPSR
bra.b A7_str ;finished here
LEN_ng:
moveq.l #1,d4 ;min LEN is 1
* A7. Calculate SCALE.
* SCALE is equal to 10^ISCALE, where ISCALE is the number
* of decimal places needed to insure LEN integer digits
* in the output before conversion to bcd. LAMBDA is the sign
* of ISCALE, used in A9. Fp1 contains 10^^(abs(ISCALE)) using
* the rounding mode as given in the following table (see
* Coonen, p. 7.23 as ref.; however, the SCALE variable is
* of opposite sign in bindec.sa from Coonen).
*
* Initial USE
* FPCR[6:5] LAMBDA SIGN(X) FPCR[6:5]
* ----------------------------------------------
* RN 00 0 0 00/0 RN
* RN 00 0 1 00/0 RN
* RN 00 1 0 00/0 RN
* RN 00 1 1 00/0 RN
* RZ 01 0 0 11/3 RP
* RZ 01 0 1 11/3 RP
* RZ 01 1 0 10/2 RM
* RZ 01 1 1 10/2 RM
* RM 10 0 0 11/3 RP
* RM 10 0 1 10/2 RM
* RM 10 1 0 10/2 RM
* RM 10 1 1 11/3 RP
* RP 11 0 0 10/2 RM
* RP 11 0 1 11/3 RP
* RP 11 1 0 11/3 RP
* RP 11 1 1 10/2 RM
*
* Register usage:
* Input/Output
* d0: exponent/scratch - final is 0
* d2: x/0 or 24 for A9
* d3: x/scratch - offset ptr into PTENRM array
* d4: LEN/Unchanged
* d5: 0/ICTR:LAMBDA
* d6: ILOG/ILOG or k if ((k<=0)&(ILOG<k))
* d7: k-factor/Unchanged
* a0: ptr for original operand/final result
* a1: x/ptr to PTENRM array
* a2: x/x
* fp0: float(ILOG)/Unchanged
* fp1: x/10^ISCALE
* fp2: x/x
* F_SCR1:x/x
* F_SCR2:Abs(X) with $3fff exponent/Unchanged
* L_SCR1:x/x
* L_SCR2:first word of X packed/Unchanged
A7_str:
tst.l d7 ;test sign of k
bgt.b k_pos ;if pos and > 0, skip this
cmp.l d6,d7 ;test k - ILOG
blt.b k_pos ;if ILOG >= k, skip this
move.l d7,d6 ;if ((k<0) & (ILOG < k)) ILOG = k
k_pos:
move.l d6,d0 ;calc ILOG + 1 - LEN in d0
addq.l #1,d0 ;add the 1
sub.l d4,d0 ;sub off LEN
swap d5 ;use upper word of d5 for LAMBDA
clr.w d5 ;set it zero initially
clr.w d2 ;set up d2 for very small case
tst.l d0 ;test sign of ISCALE
bge.b iscale ;if pos, skip next inst
addq.w #1,d5 ;if neg, set LAMBDA true
cmp.l #$ffffecd4,d0 ;test iscale <= -4908
bgt.b no_inf ;if false, skip rest
addi.l #24,d0 ;add in 24 to iscale
move.l #24,d2 ;put 24 in d2 for A9
no_inf:
neg.l d0 ;and take abs of ISCALE
iscale:
fmove.s FONE,fp1 ;init fp1 to 1
bfextu USER_FPCR(a6){26:2},d1 ;get initial rmode bits
add.w d1,d1 ;put them in bits 2:1
add.w d5,d1 ;add in LAMBDA
add.w d1,d1 ;put them in bits 3:1
tst.l L_SCR2(a6) ;test sign of original x
bge.b x_pos ;if pos, don't set bit 0
addq.l #1,d1 ;if neg, set bit 0
x_pos:
lea.l RBDTBL,a2 ;load rbdtbl base
move.b (a2,d1),d3 ;load d3 with new rmode
lsl.l #4,d3 ;put bits in proper position
fmove.l d3,fpcr ;load bits into fpu
lsr.l #4,d3 ;put bits in proper position
tst.b d3 ;decode new rmode for pten table
bne.b not_rn ;if zero, it is RN
lea.l PTENRN,a1 ;load a1 with RN table base
bra.b rmode ;exit decode
not_rn:
lsr.b #1,d3 ;get lsb in carry
bcc.b not_rp ;if carry clear, it is RM
lea.l PTENRP,a1 ;load a1 with RP table base
bra.b rmode ;exit decode
not_rp:
lea.l PTENRM,a1 ;load a1 with RM table base
rmode:
clr.l d3 ;clr table index
e_loop:
lsr.l #1,d0 ;shift next bit into carry
bcc.b e_next ;if zero, skip the mul
fmul.x (a1,d3),fp1 ;mul by 10**(d3_bit_no)
e_next:
add.l #12,d3 ;inc d3 to next pwrten table entry
tst.l d0 ;test if ISCALE is zero
bne.b e_loop ;if not, loop
* A8. Clr INEX; Force RZ.
* The operation in A3 above may have set INEX2.
* RZ mode is forced for the scaling operation to insure
* only one rounding error. The grs bits are collected in
* the INEX flag for use in A10.
*
* Register usage:
* Input/Output
fmove.l #0,FPSR ;clr INEX
fmove.l #rz_mode,FPCR ;set RZ rounding mode
* A9. Scale X -> Y.
* The mantissa is scaled to the desired number of significant
* digits. The excess digits are collected in INEX2. If mul,
* Check d2 for excess 10 exponential value. If not zero,
* the iscale value would have caused the pwrten calculation
* to overflow. Only a negative iscale can cause this, so
* multiply by 10^(d2), which is now only allowed to be 24,
* with a multiply by 10^8 and 10^16, which is exact since
* 10^24 is exact. If the input was denormalized, we must
* create a busy stack frame with the mul command and the
* two operands, and allow the fpu to complete the multiply.
*
* Register usage:
* Input/Output
* d0: FPCR with RZ mode/Unchanged
* d2: 0 or 24/unchanged
* d3: x/x
* d4: LEN/Unchanged
* d5: ICTR:LAMBDA
* d6: ILOG/Unchanged
* d7: k-factor/Unchanged
* a0: ptr for original operand/final result
* a1: ptr to PTENRM array/Unchanged
* a2: x/x
* fp0: float(ILOG)/X adjusted for SCALE (Y)
* fp1: 10^ISCALE/Unchanged
* fp2: x/x
* F_SCR1:x/x
* F_SCR2:Abs(X) with $3fff exponent/Unchanged
* L_SCR1:x/x
* L_SCR2:first word of X packed/Unchanged
A9_str:
fmove.x (a0),fp0 ;load X from memory
fabs.x fp0 ;use abs(X)
tst.w d5 ;LAMBDA is in lower word of d5
bne.b short_sc_mul ;if neg (LAMBDA = 1), scale by mul
fdiv.x fp1,fp0 ;calculate X / SCALE -> Y to fp0
bra.b A10_st ;branch to A10
sc_mul:
short_sc_mul:
tst.b BINDEC_FLG(a6) ;check for denorm
beq.b A9_norm ;if norm, continue with mul
fmovem.x fp1,-(a7) ;load ETEMP with 10^ISCALE
move.l 8(a0),-(a7) ;load FPTEMP with input arg
move.l 4(a0),-(a7)
move.l (a0),-(a7)
move.l #18,d3 ;load count for busy stack
A9_loop:
clr.l -(a7) ;clear lword on stack
dbf.w d3,A9_loop
move.b VER_TMP(a6),(a7) ;write current version number
move.b #BUSY_SIZE-4,1(a7) ;write current busy size
move.b #$10,$44(a7) ;set fcefpte[15] bit
move.w #$0023,$40(a7) ;load cmdreg1b with mul command
move.b #$fe,$8(a7) ;load all 1s to cu savepc
frestore (a7)+ ;restore frame to fpu for completion
fmul.x 36(a1),fp0 ;multiply fp0 by 10^8
fmul.x 48(a1),fp0 ;multiply fp0 by 10^16
bra.b A10_st
A9_norm:
tst.w d2 ;test for small exp case
beq.b A9_con ;if zero, continue as normal
fmul.x 36(a1),fp0 ;multiply fp0 by 10^8
fmul.x 48(a1),fp0 ;multiply fp0 by 10^16
A9_con:
fmul.x fp1,fp0 ;calculate X * SCALE -> Y to fp0
* A10. Or in INEX.
* If INEX is set, round error occurred. This is compensated
* for by 'or-ing' in the INEX2 flag to the lsb of Y.
*
* Register usage:
* Input/Output
* d0: FPCR with RZ mode/FPSR with INEX2 isolated
* d2: x/x
* d3: x/x
* d4: LEN/Unchanged
* d5: ICTR:LAMBDA
* d6: ILOG/Unchanged
* d7: k-factor/Unchanged
* a0: ptr for original operand/final result
* a1: ptr to PTENxx array/Unchanged
* a2: x/ptr to FP_SCR2(a6)
* fp0: Y/Y with lsb adjusted
* fp1: 10^ISCALE/Unchanged
* fp2: x/x
A10_st:
fmove.l FPSR,d0 ;get FPSR
fmove.x fp0,FP_SCR2(a6) ;move Y to memory
lea.l FP_SCR2(a6),a2 ;load a2 with ptr to FP_SCR2
btst.l #9,d0 ;check if INEX2 set
beq.b A11_st ;if clear, skip rest
ori.l #1,8(a2) ;or in 1 to lsb of mantissa
fmove.x FP_SCR2(a6),fp0 ;write adjusted Y back to fpu
* A11. Restore original FPCR; set size ext.
* Perform FINT operation in the user's rounding mode. Keep
* the size to extended. The sintdo entry point in the sint
* routine expects the FPCR value to be in USER_FPCR for
* mode and precision. The original FPCR is saved in L_SCR1.
A11_st:
move.l USER_FPCR(a6),L_SCR1(a6) ;save it for later
andi.l #$00000030,USER_FPCR(a6) ;set size to ext,
* ;block exceptions
* A12. Calculate YINT = FINT(Y) according to user's rounding mode.
* The FPSP routine sintd0 is used. The output is in fp0.
*
* Register usage:
* Input/Output
* d0: FPSR with AINEX cleared/FPCR with size set to ext
* d2: x/x/scratch
* d3: x/x
* d4: LEN/Unchanged
* d5: ICTR:LAMBDA/Unchanged
* d6: ILOG/Unchanged
* d7: k-factor/Unchanged
* a0: ptr for original operand/src ptr for sintdo
* a1: ptr to PTENxx array/Unchanged
* a2: ptr to FP_SCR2(a6)/Unchanged
* a6: temp pointer to FP_SCR2(a6) - orig value saved and restored
* fp0: Y/YINT
* fp1: 10^ISCALE/Unchanged
* fp2: x/x
* F_SCR1:x/x
* F_SCR2:Y adjusted for inex/Y with original exponent
* L_SCR1:x/original USER_FPCR
* L_SCR2:first word of X packed/Unchanged
A12_st:
movem.l d0-d1/a0-a1,-(a7) ;save regs used by sintd0
move.l L_SCR1(a6),-(a7)
move.l L_SCR2(a6),-(a7)
lea.l FP_SCR2(a6),a0 ;a0 is ptr to F_SCR2(a6)
fmove.x fp0,(a0) ;move Y to memory at FP_SCR2(a6)
tst.l L_SCR2(a6) ;test sign of original operand
bge.b do_fint ;if pos, use Y
or.l #$80000000,(a0) ;if neg, use -Y
do_fint:
move.l USER_FPSR(a6),-(a7)
bsr sintdo ;sint routine returns int in fp0
move.b (a7),USER_FPSR(a6)
add.l #4,a7
move.l (a7)+,L_SCR2(a6)
move.l (a7)+,L_SCR1(a6)
movem.l (a7)+,d0-d1/a0-a1 ;restore regs used by sint
move.l L_SCR2(a6),FP_SCR2(a6) ;restore original exponent
move.l L_SCR1(a6),USER_FPCR(a6) ;restore user's FPCR
* A13. Check for LEN digits.
* If the int operation results in more than LEN digits,
* or less than LEN -1 digits, adjust ILOG and repeat from
* A6. This test occurs only on the first pass. If the
* result is exactly 10^LEN, decrement ILOG and divide
* the mantissa by 10. The calculation of 10^LEN cannot
* be inexact, since all powers of ten upto 10^27 are exact
* in extended precision, so the use of a previous power-of-ten
* table will introduce no error.
*
*
* Register usage:
* Input/Output
* d0: FPCR with size set to ext/scratch final = 0
* d2: x/x
* d3: x/scratch final = x
* d4: LEN/LEN adjusted
* d5: ICTR:LAMBDA/LAMBDA:ICTR
* d6: ILOG/ILOG adjusted
* d7: k-factor/Unchanged
* a0: pointer into memory for packed bcd string formation
* a1: ptr to PTENxx array/Unchanged
* a2: ptr to FP_SCR2(a6)/Unchanged
* fp0: int portion of Y/abs(YINT) adjusted
* fp1: 10^ISCALE/Unchanged
* fp2: x/10^LEN
* F_SCR1:x/x
* F_SCR2:Y with original exponent/Unchanged
* L_SCR1:original USER_FPCR/Unchanged
* L_SCR2:first word of X packed/Unchanged
A13_st:
swap d5 ;put ICTR in lower word of d5
tst.w d5 ;check if ICTR = 0
bne not_zr ;if non-zero, go to second test
*
* Compute 10^(LEN-1)
*
fmove.s FONE,fp2 ;init fp2 to 1.0
move.l d4,d0 ;put LEN in d0
subq.l #1,d0 ;d0 = LEN -1
clr.l d3 ;clr table index
l_loop:
lsr.l #1,d0 ;shift next bit into carry
bcc.b l_next ;if zero, skip the mul
fmul.x (a1,d3),fp2 ;mul by 10**(d3_bit_no)
l_next:
add.l #12,d3 ;inc d3 to next pwrten table entry
tst.l d0 ;test if LEN is zero
bne.b l_loop ;if not, loop
*
* 10^LEN-1 is computed for this test and A14. If the input was
* denormalized, check only the case in which YINT > 10^LEN.
*
tst.b BINDEC_FLG(a6) ;check if input was norm
beq.b A13_con ;if norm, continue with checking
fabs.x fp0 ;take abs of YINT
bra test_2
*
* Compare abs(YINT) to 10^(LEN-1) and 10^LEN
*
A13_con:
fabs.x fp0 ;take abs of YINT
fcmp.x fp2,fp0 ;compare abs(YINT) with 10^(LEN-1)
fbge.w test_2 ;if greater, do next test
subq.l #1,d6 ;subtract 1 from ILOG
move.w #1,d5 ;set ICTR
fmove.l #rm_mode,FPCR ;set rmode to RM
fmul.s FTEN,fp2 ;compute 10^LEN
bra.w A6_str ;return to A6 and recompute YINT
test_2:
fmul.s FTEN,fp2 ;compute 10^LEN
fcmp.x fp2,fp0 ;compare abs(YINT) with 10^LEN
fblt.w A14_st ;if less, all is ok, go to A14
fbgt.w fix_ex ;if greater, fix and redo
fdiv.s FTEN,fp0 ;if equal, divide by 10
addq.l #1,d6 ; and inc ILOG
bra.b A14_st ; and continue elsewhere
fix_ex:
addq.l #1,d6 ;increment ILOG by 1
move.w #1,d5 ;set ICTR
fmove.l #rm_mode,FPCR ;set rmode to RM
bra.w A6_str ;return to A6 and recompute YINT
*
* Since ICTR <> 0, we have already been through one adjustment,
* and shouldn't have another; this is to check if abs(YINT) = 10^LEN
* 10^LEN is again computed using whatever table is in a1 since the
* value calculated cannot be inexact.
*
not_zr:
fmove.s FONE,fp2 ;init fp2 to 1.0
move.l d4,d0 ;put LEN in d0
clr.l d3 ;clr table index
z_loop:
lsr.l #1,d0 ;shift next bit into carry
bcc.b z_next ;if zero, skip the mul
fmul.x (a1,d3),fp2 ;mul by 10**(d3_bit_no)
z_next:
add.l #12,d3 ;inc d3 to next pwrten table entry
tst.l d0 ;test if LEN is zero
bne.b z_loop ;if not, loop
fabs.x fp0 ;get abs(YINT)
fcmp.x fp2,fp0 ;check if abs(YINT) = 10^LEN
fbne.w A14_st ;if not, skip this
fdiv.s FTEN,fp0 ;divide abs(YINT) by 10
addq.l #1,d6 ;and inc ILOG by 1
addq.l #1,d4 ; and inc LEN
fmul.s FTEN,fp2 ; if LEN++, the get 10^^LEN
* A14. Convert the mantissa to bcd.
* The binstr routine is used to convert the LEN digit
* mantissa to bcd in memory. The input to binstr is
* to be a fraction; i.e. (mantissa)/10^LEN and adjusted
* such that the decimal point is to the left of bit 63.
* The bcd digits are stored in the correct position in
* the final string area in memory.
*
*
* Register usage:
* Input/Output
* d0: x/LEN call to binstr - final is 0
* d1: x/0
* d2: x/ms 32-bits of mant of abs(YINT)
* d3: x/ls 32-bits of mant of abs(YINT)
* d4: LEN/Unchanged
* d5: ICTR:LAMBDA/LAMBDA:ICTR
* d6: ILOG
* d7: k-factor/Unchanged
* a0: pointer into memory for packed bcd string formation
* /ptr to first mantissa byte in result string
* a1: ptr to PTENxx array/Unchanged
* a2: ptr to FP_SCR2(a6)/Unchanged
* fp0: int portion of Y/abs(YINT) adjusted
* fp1: 10^ISCALE/Unchanged
* fp2: 10^LEN/Unchanged
* F_SCR1:x/Work area for final result
* F_SCR2:Y with original exponent/Unchanged
* L_SCR1:original USER_FPCR/Unchanged
* L_SCR2:first word of X packed/Unchanged
A14_st:
fmove.l #rz_mode,FPCR ;force rz for conversion
fdiv.x fp2,fp0 ;divide abs(YINT) by 10^LEN
lea.l FP_SCR1(a6),a0
fmove.x fp0,(a0) ;move abs(YINT)/10^LEN to memory
move.l 4(a0),d2 ;move 2nd word of FP_RES to d2
move.l 8(a0),d3 ;move 3rd word of FP_RES to d3
clr.l 4(a0) ;zero word 2 of FP_RES
clr.l 8(a0) ;zero word 3 of FP_RES
move.l (a0),d0 ;move exponent to d0
swap d0 ;put exponent in lower word
beq.b no_sft ;if zero, don't shift
subi.l #$3ffd,d0 ;sub bias less 2 to make fract
tst.l d0 ;check if > 1
bgt.b no_sft ;if so, don't shift
neg.l d0 ;make exp positive
m_loop:
lsr.l #1,d2 ;shift d2:d3 right, add 0s
roxr.l #1,d3 ;the number of places
dbf.w d0,m_loop ;given in d0
no_sft:
tst.l d2 ;check for mantissa of zero
bne.b no_zr ;if not, go on
tst.l d3 ;continue zero check
beq.b zer_m ;if zero, go directly to binstr
no_zr:
clr.l d1 ;put zero in d1 for addx
addi.l #$00000080,d3 ;inc at bit 7
addx.l d1,d2 ;continue inc
andi.l #$ffffff80,d3 ;strip off lsb not used by 882
zer_m:
move.l d4,d0 ;put LEN in d0 for binstr call
addq.l #3,a0 ;a0 points to M16 byte in result
bsr binstr ;call binstr to convert mant
* A15. Convert the exponent to bcd.
* As in A14 above, the exp is converted to bcd and the
* digits are stored in the final string.
*
* Digits are stored in L_SCR1(a6) on return from BINDEC as:
*
* 32 16 15 0
* -----------------------------------------
* | 0 | e3 | e2 | e1 | e4 | X | X | X |
* -----------------------------------------
*
* And are moved into their proper places in FP_SCR1. If digit e4
* is non-zero, OPERR is signaled. In all cases, all 4 digits are
* written as specified in the 881/882 manual for packed decimal.
*
* Register usage:
* Input/Output
* d0: x/LEN call to binstr - final is 0
* d1: x/scratch (0);shift count for final exponent packing
* d2: x/ms 32-bits of exp fraction/scratch
* d3: x/ls 32-bits of exp fraction
* d4: LEN/Unchanged
* d5: ICTR:LAMBDA/LAMBDA:ICTR
* d6: ILOG
* d7: k-factor/Unchanged
* a0: ptr to result string/ptr to L_SCR1(a6)
* a1: ptr to PTENxx array/Unchanged
* a2: ptr to FP_SCR2(a6)/Unchanged
* fp0: abs(YINT) adjusted/float(ILOG)
* fp1: 10^ISCALE/Unchanged
* fp2: 10^LEN/Unchanged
* F_SCR1:Work area for final result/BCD result
* F_SCR2:Y with original exponent/ILOG/10^4
* L_SCR1:original USER_FPCR/Exponent digits on return from binstr
* L_SCR2:first word of X packed/Unchanged
A15_st:
tst.b BINDEC_FLG(a6) ;check for denorm
beq.b not_denorm
ftst.x fp0 ;test for zero
fbeq.w den_zero ;if zero, use k-factor or 4933
fmove.l d6,fp0 ;float ILOG
fabs.x fp0 ;get abs of ILOG
bra.b convrt
den_zero:
tst.l d7 ;check sign of the k-factor
blt.b use_ilog ;if negative, use ILOG
fmove.s F4933,fp0 ;force exponent to 4933
bra.b convrt ;do it
use_ilog:
fmove.l d6,fp0 ;float ILOG
fabs.x fp0 ;get abs of ILOG
bra.b convrt
not_denorm:
ftst.x fp0 ;test for zero
fbne.w not_zero ;if zero, force exponent
fmove.s FONE,fp0 ;force exponent to 1
bra.b convrt ;do it
not_zero:
fmove.l d6,fp0 ;float ILOG
fabs.x fp0 ;get abs of ILOG
convrt:
fdiv.x 24(a1),fp0 ;compute ILOG/10^4
fmove.x fp0,FP_SCR2(a6) ;store fp0 in memory
move.l 4(a2),d2 ;move word 2 to d2
move.l 8(a2),d3 ;move word 3 to d3
move.w (a2),d0 ;move exp to d0
beq.b x_loop_fin ;if zero, skip the shift
subi.w #$3ffd,d0 ;subtract off bias
neg.w d0 ;make exp positive
x_loop:
lsr.l #1,d2 ;shift d2:d3 right
roxr.l #1,d3 ;the number of places
dbf.w d0,x_loop ;given in d0
x_loop_fin:
clr.l d1 ;put zero in d1 for addx
addi.l #$00000080,d3 ;inc at bit 6
addx.l d1,d2 ;continue inc
andi.l #$ffffff80,d3 ;strip off lsb not used by 882
move.l #4,d0 ;put 4 in d0 for binstr call
lea.l L_SCR1(a6),a0 ;a0 is ptr to L_SCR1 for exp digits
bsr binstr ;call binstr to convert exp
move.l L_SCR1(a6),d0 ;load L_SCR1 lword to d0
move.l #12,d1 ;use d1 for shift count
lsr.l d1,d0 ;shift d0 right by 12
bfins d0,FP_SCR1(a6){4:12} ;put e3:e2:e1 in FP_SCR1
lsr.l d1,d0 ;shift d0 right by 12
bfins d0,FP_SCR1(a6){16:4} ;put e4 in FP_SCR1
tst.b d0 ;check if e4 is zero
beq.b A16_st ;if zero, skip rest
or.l #opaop_mask,USER_FPSR(a6) ;set OPERR & AIOP in USER_FPSR
* A16. Write sign bits to final string.
* Sigma is bit 31 of initial value; RHO is bit 31 of d6 (ILOG).
*
* Register usage:
* Input/Output
* d0: x/scratch - final is x
* d2: x/x
* d3: x/x
* d4: LEN/Unchanged
* d5: ICTR:LAMBDA/LAMBDA:ICTR
* d6: ILOG/ILOG adjusted
* d7: k-factor/Unchanged
* a0: ptr to L_SCR1(a6)/Unchanged
* a1: ptr to PTENxx array/Unchanged
* a2: ptr to FP_SCR2(a6)/Unchanged
* fp0: float(ILOG)/Unchanged
* fp1: 10^ISCALE/Unchanged
* fp2: 10^LEN/Unchanged
* F_SCR1:BCD result with correct signs
* F_SCR2:ILOG/10^4
* L_SCR1:Exponent digits on return from binstr
* L_SCR2:first word of X packed/Unchanged
A16_st:
clr.l d0 ;clr d0 for collection of signs
andi.b #$0f,FP_SCR1(a6) ;clear first nibble of FP_SCR1
tst.l L_SCR2(a6) ;check sign of original mantissa
bge.b mant_p ;if pos, don't set SM
moveq.l #2,d0 ;move 2 in to d0 for SM
mant_p:
tst.l d6 ;check sign of ILOG
bge.b wr_sgn ;if pos, don't set SE
addq.l #1,d0 ;set bit 0 in d0 for SE
wr_sgn:
bfins d0,FP_SCR1(a6){0:2} ;insert SM and SE into FP_SCR1
* Clean up and restore all registers used.
fmove.l #0,FPSR ;clear possible inex2/ainex bits
fmovem.x (a7)+,fp0-fp2
movem.l (a7)+,d2-d7/a2
rts
end
Reference in New Issue View Git Blame Copy Permalink