![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to mc68020.s CVS log ![[TXT]](/cvsweb/icons/text.gif){kind=icon}
![[DIR]](/cvsweb/icons/dir.gif){kind=icon}
Up to [PGP] / pgp / src|
Return to mc68020.s CVS log | Up to [PGP] / pgp / src |
PGP 2.6.3i
| Fast assembly routines for MC68020 (Sun-3) | Assumptions: | Arguments start at sp@(0x4) | Return value is in d0 | d0/d1/a0/a1 are scratch | P_SMUL needs MULTUNIT set to "unsigned long" in mpilib.c | P_DMUL replaces mp_smul and mp_dmul in mpilib.c | | 92.9.21 - Tsutomu Shimomura, [email protected] | 93.5.14 - Bug in P_DMUL fixed -- now works with small bignums .text | P_SETP(p) sets the current precision to be p longwords. No-op. .proc .globl _P_SETP _P_SETP: | movl #L2000, a0 |% | jsr mcount |% | .bss |% | .even |% |L2000: .skip 4 |% | .text |% rts | P_ADDC(*a, *b, c) performs a += b + c (carry). Carry is returned. .proc .globl _P_ADDC _P_ADDC: | movl #L2001, a0 |% | jsr mcount |% | .bss |% | .even |% |L2001: .skip 4 |% | .text |% movl sp@(0x4), a0 | claim arguments movl sp@(0x8), a1 movl sp@(0xc), d0 movl d2, sp@- | preserve d2 movw _global_precision, d1 | longword count movw d1, d2 | save a copy lslw #2, d1 addw d1, a0 | adjust array pointers addw d1, a1 lsrw #1, d1 | compute initial branch offset andw #0xe, d1 negw d1 | branch offset in d1 lsrw #3, d2 | 8 longwords/loop; count in d2 asrl #1, d0 | set X if necessary jmp pc@(0x12,d1:w) 1: addxl a1@-, a0@- addxl a1@-, a0@- addxl a1@-, a0@- addxl a1@-, a0@- addxl a1@-, a0@- addxl a1@-, a0@- addxl a1@-, a0@- addxl a1@-, a0@- dbf d2, 1b roxll #1, d0 movl sp@+, d2 rts | P_SUBB(*a, *b, c) performs a -= b + c (borrow). Borrow is returned. .proc .globl _P_SUBB _P_SUBB: | movl #L2002, a0 |% | jsr mcount |% | .bss |% | .even |% |L2002: .skip 4 |% | .text |% movl sp@(0x4), a0 | claim arguments movl sp@(0x8), a1 movl sp@(0xc), d0 movl d2, sp@- | preserve d2 movw _global_precision, d1 | longword count movw d1, d2 | save a copy lslw #2, d1 addw d1, a0 | adjust array pointers addw d1, a1 lsrw #1, d1 | compute initial branch offset andw #0xe, d1 negw d1 | branch offset in d1 lsrw #3, d2 | 8 longwords/loop; count in d2 asrl #1, d0 | set X if necessary jmp pc@(0x12,d1:w) 1: subxl a1@-, a0@- subxl a1@-, a0@- subxl a1@-, a0@- subxl a1@-, a0@- subxl a1@-, a0@- subxl a1@-, a0@- subxl a1@-, a0@- subxl a1@-, a0@- dbf d2, 1b roxll #1, d0 movl sp@+, d2 rts | P_ROTL(*a, c) performs a = (a<<1) | c (lo-bit). Hi-bit is returned. .proc .globl _P_ROTL _P_ROTL: | movl #L2003, a0 |% | jsr mcount |% | .bss |% | .even |% |L2003: .skip 4 |% | .text |% movl sp@(0x4), a0 | claim arguments movl sp@(0x8), d0 movl d2, a1 | preserve d2 movw _global_precision, d1 | longword count movw d1, d2 | save a copy lslw #2, d1 addw d1, a0 | adjust array pointer andw #0x1c, d1 negw d1 | branch offset in d1 lsrw #3, d2 | 8 longwords/loop; count in d2 asrl #1, d0 | set X if necessary jmp pc@(0x22,d1:w) 1: roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- roxlw a0@- dbf d2, 1b roxll #1, d0 movl a1, d2 rts | P_SMUL(*a, *b, x) performs a += b * x. Pointers are to the LSB. .proc .globl _P_SMUL _P_SMUL: | movl #L2004, a0 |% | jsr mcount |% | .bss |% | .even |% |L2004: .skip 4 |% | .text |% movl sp@(0x4), a0 | claim arguments movl sp@(0x8), a1 movl sp@(0xc), d1 tstl d1 | horrible kludge to speed multiply by 0 beq 3f moveml #0x3c00, sp@- | d2/d3/d4/d5; Sun's optimizer is really *DUMB* movw _global_precision, d5 | longword count; 0 will fail subqw #2, d5 | first longword not handled in loop clrl d4 movl a1@, d2 mulul d1, d3:d2 | d3 is carry addl d2, a0@ | accumulate | tstw d5 | This code needed if global_precision < 2 | blt 2f | only one longword? 1: movl a0@-, d0 addxl d3, d0 | accumulate carry and X-bit movl a1@-, d2 mulul d1, d3:d2 | d3 is carry addxl d4, d3 | add X-bit to carry addl d2, d0 | accumulate movl d0, a0@ dbf d5, 1b 2: addxl d4, d3 | add X-bit to carry movl d3, a0@- moveml sp@+, #0x3c | d2/d3/d4/d5; Sun's optimizer is really *DUMB* 3: rts | P_DMUL(*a, *b, *c) performs a = b * c. .proc .globl _P_DMUL _P_DMUL: | movl #L2005, a0 |% | jsr mcount |% | .bss |% | .even |% |L2005: .skip 4 |% | .text |% moveml #0x3f38, sp@- | d2-d7/a2-a4; Sun's optimizer is really *DUMB* movl sp@(0x28), a0 | claim arguments movl sp@(0x2c), a1 movl sp@(0x30), a2 movw _global_precision, d0 subqw #2, d0 | global_precision - 2 movl a0, a4 | product movw d0, d7 | count for multiplicand movl a1, a3 | multiplicand tstl a3@+ bne 2f subqw #1, d7 1: clrl a4@+ tstl a3@+ dbne d7, 1b addqw #1, d7 | d7 contains effective size of the multiplicand-2 2: movl d0, d6 | count for multiplier movl a2, a3 | multiplier tstl a3@+ bne 2f subqw #1, d6 1: clrl a4@+ tstl a3@+ dbne d6, 1b addqw #1, d6 | d6 contains effective size of the multiplier-2 2: addqw #1, d0 | global_precision - 1 lslw #2, d0 addw d0, a1 | pointer to LSB of the multiplicand addw d0, a2 | pointer to LSB of the multiplier addw d0, a0 addw d0, a0 addql #4, a0 | pointer to LSB of product - KLUDGE! | First partial product not handled in loop | Assumes that the X-bit is clear from the above contortions. clrl d4 movl a0, a3 | product movl a1, a4 | multiplicand movl a2@, d1 | one longword of the multiplier movw d7, d5 | loop count movl a4@, d2 mulul d1, d3:d2 | d3 is carry movl d2, a3@ | store product 1: movl a4@-, d2 mulul d1, d0:d2 addxl d3, d2 movl d0, d3 movl d2, a3@- dbf d5, 1b addxl d4, d3 movl d3, a3@- | The other partial products 2: movl a1, a4 | multiplicand movl a2@-, d1 | another longword of the multiplier movw d7, d5 | loop count movl a4@, d2 mulul d1, d3:d2 | d3 is carry addl d2, a0@- | accumulate movl a0, a3 | product 1: movl a3@-, d0 addxl d3, d0 | accumulate carry and X-bit movl a4@-, d2 mulul d1, d3:d2 | d3 is carry addxl d4, d3 | add X-bit to carry addl d2, d0 | accumulate movl d0, a3@ dbf d5, 1b addxl d4, d3 | add X-bit to carry movl d3, a3@- dbf d6, 2b moveml sp@+, #0x1cfc | d2-d7/a2-a4; Sun's optimizer is really *DUMB* rts
This archive runs on limited infrastructure. Preserving old code on modern bandwidth. Automated agents are requested to crawl responsibly.