home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
GEMini Atari
/
GEMini_Atari_CD-ROM_Walnut_Creek_December_1993.iso
/
files
/
mint
/
mntlib18
/
_addsubd.cpp
next >
Wrap
Text File
|
1993-08-03
|
5KB
|
202 lines
#ifndef __M68881__
#ifndef sfp004
| double floating point add/subtract routine
|
| written by Kai-Uwe Bloem (I5110401@dbstu1.bitnet).
| Based on a 80x86 floating point packet from comp.os.minix, written by P.Housel
|
|
| Revision 1.3, kub 01-90 :
| added support for denormalized numbers
|
| Revision 1.2, kub 01-90 :
| replace far shifts by swaps to gain speed
|
| Revision 1.1, kub 12-89 :
| Ported over to 68k assembler
|
| Revision 1.0:
| original 8088 code from P.S.Housel
.text
.even
.globl __subdf3, ___subdf3
.globl __adddf3, ___adddf3
__subdf3:
___subdf3:
eorb #0x80,sp@(12) | reverse sign of v
__adddf3:
___adddf3:
lea sp@(4),a0 | pointer to u and v parameter
moveml d2-d7,sp@- | save registers
moveml a0@,d4-d5/d6-d7 | d4-d5 = v, d6-d7 = u
movel d6,d0 | d0 = u.exp
swap d0
movel d6,d2 | d2.h = u.sign
movew d0,d2
lsrw #4,d0
andw #0x07ff,d0 | kill sign bit
movel d4,d1 | d1 = v.exp
swap d1
eorw d1,d2 | d2.l = u.sign ^ v.sign
lsrw #4,d1
andw #0x07ff,d1 | kill sign bit
andl #0x0fffff,d6 | remove exponent from u.mantissa
tstw d0 | check for zero exponent - no leading "1"
beq 0f
orl #0x100000,d6 | restore implied leading "1"
bra 1f
0: addw #1,d0 | "normalize" exponent
1:
andl #0x0fffff,d4 | remove exponent from v.mantissa
tstw d1 | check for zero exponent - no leading "1"
beq 0f
orl #0x100000,d4 | restore implied leading "1"
bra 1f
0: addw #1,d1 | "normalize" exponent
1:
clrw d3 | (put initial zero rounding bits in d3)
negw d1 | d1 = u.exp - v.exp
addw d0,d1
beq 5f | exponents are equal - no shifting neccessary
bgt 1f | not equal but no exchange neccessary
exg d4,d6 | exchange u and v
exg d5,d7
subw d1,d0 | d0 = u.exp - (u.exp - v.exp) = v.exp
negw d1
tstw d2 | d2.h = u.sign ^ (u.sign ^ v.sign) = v.sign
bpl 1f
bchg #31,d2
1:
cmpw #53,d1 | is u so much bigger that v is not
bge 7f | significant ?
movew #10-1,d3 | shift u left up to 10 bits to minimize loss
2:
addl d7,d7
addxl d6,d6
subw #1,d0 | decrement exponent
subw #1,d1 | done shifting altogether ?
dbeq d3,2b | loop if still can shift u.mant more
clrw d3
3:
cmpw #16,d1 | see if fast rotate possible
blt 4f
orb d5,d3 | set rounding bits
orb d2,d3
sne d2 | "sticky byte"
movew d5,d3
lsrw #8,d3
movew d4,d5 | rotate by swapping register halfs
swap d5
clrw d4
swap d4
subw #16,d1
bra 3b
0:
lsrl #1,d4 | shift v.mant right the rest of the way
roxrl #1,d5 | to line it up with u.mant
orb d3,d2 | set "sticky byte" if necessary
roxrw #1,d3 | shift into rounding bits
4: dbra d1,0b | loop
andb #1,d2 | see if "sticky bit" should be set
orb d2,d3
5:
tstw d2 | are the signs equal ?
bpl 6f | yes, no negate necessary
negb d3 | negate rounding bits and v.mant
negl d5
negxl d4
6:
addl d5,d7 | u.mant = u.mant + v.mant
addxl d4,d6
bcs 7f | need not negate
tstw d2 | opposite signs ?
bpl 7f | do not need to negate result
negb d3 | negate rounding bits and u.mant
negl d7
negxl d6
notl d2 | switch sign
7:
movel d6,d4 | move result for normalization
movel d7,d5
moveb d3,d1 | put rounding bits in d1 for norm_df
swap d2 | put sign into d2 (exponent is in d0)
jmp norm_df | leave registers on stack for norm_df
#else sfp004
| double precision floating point stuff for Atari-gcc using the SFP004
| developed with gas
|
| double floating point add/subtract routine
|
| M. Ritzert (mjr at dmzrzu71)
|
| 4.10.1990
|
| no NAN checking implemented since the 68881 treats this situation "correct",
| i.e. according to IEEE
| addresses of the 68881 data port. This choice is fastest when much data is
| transferred between the two processors.
comm = -6
resp = -16
zahl = 0
| waiting loop ...
|
| wait:
| ww: cmpiw #0x8900,a0@(resp)
| beq ww
| is coded directly by
| .long 0x0c688900, 0xfff067f8
.text
.even
.globl __subdf3, ___subdf3
.globl __adddf3, ___adddf3
__subdf3:
___subdf3:
| double precision subtraction
| sub second arg from fp0
lea 0xfffa50,a0
movew #0x5400,a0@(comm) | load first argument to fp0
cmpiw #0x8900,a0@(resp) | check
movel a7@(4),a0@
movel a7@(8),a0@
movew #0x5428,a0@(comm)
.long 0x0c688900, 0xfff067f8
movel a7@(12),a0@
movel a7@(16),a0@
movew #0x7400,a0@(comm) | result to d0/d1
.long 0x0c688900, 0xfff067f8
movel a0@,d0
movel a0@,d1
rts
__adddf3:
___adddf3:
lea 0xfffa50,a0
movew #0x5400,a0@(comm) | load fp0
cmpiw #0x8900,a0@(resp) | got it?
movel a7@(4),a0@ | take a hi from stack to FPU
movel a7@(8),a0@ | take a lo from stack to FPU
movew #0x5422,a0@(comm) | add second arg to fp0
.long 0x0c688900, 0xfff067f8
movel a7@(12),a0@ | move b hi from stack to FPU
movel a7@(16),a0@ | move b lo from stack to FPU
movew #0x7400,a0@(comm) | result to d0/d1
.long 0x0c688900, 0xfff067f8
movel a0@,d0 | download result
movel a0@,d1 | download result
rts
#endif sfp004
#endif __M68881__