home *** CD-ROM | disk | FTP | other *** search
- | mjr: not needed on the TT
-
- #ifndef __M68881__
-
- .text
- .even
- .globl __mulsf3, ___mulsf3
-
- __mulsf3:
- ___mulsf3:
-
- # ifdef sfp004
-
- | single precision floating point stuff for Atari-gcc using the SFP004
- | developed with gas
- |
- | single floating point multiplication routine
- |
- | M. Ritzert (mjr at dfg.dbp.de)
- |
- | 7. Juli 1989
- |
- | revision 1.1: adapted to the gcc lib patchlevel 58
- | 4.10.90
-
- comm = -6
- resp = -16
- zahl = 0
-
- lea 0xfffffa50:w,a0
- movew #0x4400,a0@(comm) | load first argument to fp0
- cmpiw #0x8900,a0@(resp) | check
- movel sp@(4),a0@
- movew #0x4427,a0@(comm)
- .long 0x0c688900, 0xfff067f8
- movel sp@(8),a0@
- movew #0x6400,a0@(comm) | result to d0
- .long 0x0c688900, 0xfff067f8
- movel a0@,d0
- rts
-
- # else sfp004
-
- | single floating point multiplication 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 1.2.4 michal 05-93 (ntomczak@vm.ucs.ualberta.ca)
- | + ensure that Inf * NaN == NaN * Inf == NaN
- | and 0 * Inf = Inf * 0 = NaN
- |
- | Revision 1.2.3 michal 05-93 (ntomczak@vm.ucs.ualberta.ca)
- | + code smoothing
- |
- | patched by Olaf Flebbe (flebbe@tat.physik.uni-tuebingen.de)
- |
- | Revision 1.2.2 olaf 05-93
- | + fixed a bug with -0.
- |
- | Revision 1.2.1 olaf 12-92:
- | + added support for NaN and Infinites
- | + added support for -0
- |
- | Revision 1.2, kub 01-90 :
- | added support for denormalized numbers
- |
- | Revision 1.1, kub 12-89 :
- | Created single float version for 68000. Code could be speed up by having
- | the accumulator in the 68000 register set ...
- |
- | Revision 1.0:
- | original 8088 code from P.S.Housel for double floats
-
- BIAS4 = 0x7F-1
-
- lea sp@(4),a0
- moveml d2-d5,sp@-
- moveml a0@,d4/d5 | d4 = v, d5 = u
-
- movel #0x7fffff,d3
- movel d5,d0 | d0 = u.exp
- andl d3,d5 | remove exponent from u.mantissa
- swap d0
- movew d0,d2 | d2 = u.sign
-
- movel d4,d1 | d1 = v.exp
- andl d3,d4 | remove exponent from v.mantissa
- swap d1
- eorw d1,d2 | d2 = u.sign ^ v.sign (in bit 15)
-
- moveq #15,d3
- bclr d3,d0 | kill sign bit
- bclr d3,d1 | kill sign bit
- tstl d0 | test if one of factors is 0
- beq 1f
- tstl d1
- 1: seq d2 | 'one of factors is 0' flag in the lowest byte
- lsrw #7,d0 | keep here exponents only
- lsrw #7,d1
-
- |
- | Testing for NaN and Infinities
- |
- moveq #-1,d3
- cmpb d3,d0
- beq 0f
- cmpb d3,d1
- bne nospec
- bra 1f
- |
- | first operand is special
- |
- 0: tstl d5 | is it NaN?
- bne retnan
- 1: tstb d2 | 0 times special or special times 0 ?
- bne retnan | yes -> NaN
- cmpb d3,d1 | is the other special ?
- beq 2f | maybe it is NaN
- |
- | Return Infinity with correct sign
- |
- retinf: movel #0xff000000,d0 | we will return #0xff800000 or #0x7f800000
- lslw #1,d2
- roxrl #1,d0 | shift in high bit as given by d2
- return: moveml sp@+,d2-d5
- rts
-
- |
- | v is special
- |
-
- 2: tstl d4 | is this NaN?
- beq retinf | we know that the other is not zero
- retnan: moveql #-1,d0
- lsrl #1,d0 | 0x7fffffff -> d0
- bra return
- |
- | end of NaN and Inf.
- |
- nospec: tstb d2 | not needed - but we can waste two instr.
- bne retzz | return signed 0 if one of factors is 0
- moveq #23,d3
- bset d3,d5 | restore implied leading "1"
- subqw #8,sp | multiplication accumulator
- tstw d0 | check for zero exponent - no leading "1"
- bne 1f
- bclr d3,d5 | remove it
- addqw #1,d0 | "normalize" exponent
- 1: tstl d5
- beq retz | multiplying zero
-
- moveq #23,d3
- bset d3,d4 | restore implied leading "1"
- tstw d1 | check for zero exponent - no leading "1"
- bne 1f
- bclr d3,d4 | remove it
- addqw #1,d1 | "normalize" exponent
- 1: tstl d4
- beq retz | multiply by zero
-
- addw d1,d0 | add exponents,
- subw #BIAS4+16-8,d0 | remove excess bias, acnt for repositioning
-
- clrl sp@ | initialize 64-bit product to zero
- clrl sp@(4)
-
- | see Knuth, Seminumerical Algorithms, section 4.3. algorithm M
-
- movew d4,d3
- mulu d5,d3 | mulitply with bigit from multiplier
- movel d3,sp@(4) | store into result
-
- movel d4,d3
- swap d3
- mulu d5,d3
- addl d3,sp@(2) | add to result
-
- swap d5 | [TOP 8 BITS SHOULD BE ZERO !]
-
- movew d4,d3
- mulu d5,d3 | mulitply with bigit from multiplier
- addl d3,sp@(2) | store into result (no carry can occur here)
-
- movel d4,d3
- swap d3
- mulu d5,d3
- addl d3,sp@ | add to result
- | [TOP 16 BITS SHOULD BE ZERO !]
- moveml sp@(2),d4-d5 | get the 48 valid mantissa bits
- clrw d5 | (pad to 64)
-
- movel #0x0000ffff,d3
- 2:
- cmpl d3,d4 | multiply (shift) until
- bhi 3f | 1 in upper 16 result bits
- cmpw #9,d0 | give up for denormalized numbers
- ble 3f
- swap d4 | (we''re getting here only when multiplying
- swap d5 | with a denormalized number; there''s an
- movew d5,d4 | eventual loss of 4 bits in the rounding
- clrw d5 | byte -- what a pity 8-)
- subqw #8,d0 | decrement exponent
- subqw #8,d0
- bra 2b
- 3:
- movel d5,d1 | get rounding bits
- roll #8,d1
- movel d1,d3 | see if sticky bit should be set
- andl #0xffffff00,d3
- beq 4f
- orb #1,d1 | set "sticky bit" if any low-order set
- 4: addqw #8,sp | remove accumulator from stack
- jmp norm_sf | (result in d4)
-
- retz: addqw #8,sp | release accumulator space
- retzz: moveq #0,d0 | save zero as result
- lslw #1,d2 | and set it sign as for d2
- roxrl #1,d0
- moveml sp@+,d2-d5
- rts | no normalizing neccessary
-
- # endif sfp004
- #endif __M68881__
-