From 1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 Mon Sep 17 00:00:00 2001 From: Linus Torvalds Date: Sat, 16 Apr 2005 15:20:36 -0700 Subject: Linux-2.6.12-rc2 Initial git repository build. I'm not bothering with the full history, even though we have it. We can create a separate "historical" git archive of that later if we want to, and in the meantime it's about 3.2GB when imported into git - space that would just make the early git days unnecessarily complicated, when we don't have a lot of good infrastructure for it. Let it rip! --- arch/m68k/fpsp040/round.S | 649 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 649 insertions(+) create mode 100644 arch/m68k/fpsp040/round.S (limited to 'arch/m68k/fpsp040/round.S') diff --git a/arch/m68k/fpsp040/round.S b/arch/m68k/fpsp040/round.S new file mode 100644 index 000000000000..00f98068783f --- /dev/null +++ b/arch/m68k/fpsp040/round.S @@ -0,0 +1,649 @@ +| +| round.sa 3.4 7/29/91 +| +| handle rounding and normalization tasks +| +| +| +| Copyright (C) Motorola, Inc. 1990 +| All Rights Reserved +| +| THIS IS UNPUBLISHED PROPRIETARY SOURCE CODE OF MOTOROLA +| The copyright notice above does not evidence any +| actual or intended publication of such source code. + +|ROUND idnt 2,1 | Motorola 040 Floating Point Software Package + + |section 8 + +#include "fpsp.h" + +| +| round --- round result according to precision/mode +| +| a0 points to the input operand in the internal extended format +| d1(high word) contains rounding precision: +| ext = $0000xxxx +| sgl = $0001xxxx +| dbl = $0002xxxx +| d1(low word) contains rounding mode: +| RN = $xxxx0000 +| RZ = $xxxx0001 +| RM = $xxxx0010 +| RP = $xxxx0011 +| d0{31:29} contains the g,r,s bits (extended) +| +| On return the value pointed to by a0 is correctly rounded, +| a0 is preserved and the g-r-s bits in d0 are cleared. +| The result is not typed - the tag field is invalid. The +| result is still in the internal extended format. +| +| The INEX bit of USER_FPSR will be set if the rounded result was +| inexact (i.e. if any of the g-r-s bits were set). +| + + .global round +round: +| If g=r=s=0 then result is exact and round is done, else set +| the inex flag in status reg and continue. +| + bsrs ext_grs |this subroutine looks at the +| :rounding precision and sets +| ;the appropriate g-r-s bits. + tstl %d0 |if grs are zero, go force + bne rnd_cont |lower bits to zero for size + + swap %d1 |set up d1.w for round prec. + bra truncate + +rnd_cont: +| +| Use rounding mode as an index into a jump table for these modes. +| + orl #inx2a_mask,USER_FPSR(%a6) |set inex2/ainex + lea mode_tab,%a1 + movel (%a1,%d1.w*4),%a1 + jmp (%a1) +| +| Jump table indexed by rounding mode in d1.w. All following assumes +| grs != 0. +| +mode_tab: + .long rnd_near + .long rnd_zero + .long rnd_mnus + .long rnd_plus +| +| ROUND PLUS INFINITY +| +| If sign of fp number = 0 (positive), then add 1 to l. +| +rnd_plus: + swap %d1 |set up d1 for round prec. + tstb LOCAL_SGN(%a0) |check for sign + bmi truncate |if positive then truncate + movel #0xffffffff,%d0 |force g,r,s to be all f's + lea add_to_l,%a1 + movel (%a1,%d1.w*4),%a1 + jmp (%a1) +| +| ROUND MINUS INFINITY +| +| If sign of fp number = 1 (negative), then add 1 to l. +| +rnd_mnus: + swap %d1 |set up d1 for round prec. + tstb LOCAL_SGN(%a0) |check for sign + bpl truncate |if negative then truncate + movel #0xffffffff,%d0 |force g,r,s to be all f's + lea add_to_l,%a1 + movel (%a1,%d1.w*4),%a1 + jmp (%a1) +| +| ROUND ZERO +| +| Always truncate. +rnd_zero: + swap %d1 |set up d1 for round prec. + bra truncate +| +| +| ROUND NEAREST +| +| If (g=1), then add 1 to l and if (r=s=0), then clear l +| Note that this will round to even in case of a tie. +| +rnd_near: + swap %d1 |set up d1 for round prec. + asll #1,%d0 |shift g-bit to c-bit + bcc truncate |if (g=1) then + lea add_to_l,%a1 + movel (%a1,%d1.w*4),%a1 + jmp (%a1) + +| +| ext_grs --- extract guard, round and sticky bits +| +| Input: d1 = PREC:ROUND +| Output: d0{31:29}= guard, round, sticky +| +| The ext_grs extract the guard/round/sticky bits according to the +| selected rounding precision. It is called by the round subroutine +| only. All registers except d0 are kept intact. d0 becomes an +| updated guard,round,sticky in d0{31:29} +| +| Notes: the ext_grs uses the round PREC, and therefore has to swap d1 +| prior to usage, and needs to restore d1 to original. +| +ext_grs: + swap %d1 |have d1.w point to round precision + cmpiw #0,%d1 + bnes sgl_or_dbl + bras end_ext_grs + +sgl_or_dbl: + moveml %d2/%d3,-(%a7) |make some temp registers + cmpiw #1,%d1 + bnes grs_dbl +grs_sgl: + bfextu LOCAL_HI(%a0){#24:#2},%d3 |sgl prec. g-r are 2 bits right + movel #30,%d2 |of the sgl prec. limits + lsll %d2,%d3 |shift g-r bits to MSB of d3 + movel LOCAL_HI(%a0),%d2 |get word 2 for s-bit test + andil #0x0000003f,%d2 |s bit is the or of all other + bnes st_stky |bits to the right of g-r + tstl LOCAL_LO(%a0) |test lower mantissa + bnes st_stky |if any are set, set sticky + tstl %d0 |test original g,r,s + bnes st_stky |if any are set, set sticky + bras end_sd |if words 3 and 4 are clr, exit +grs_dbl: + bfextu LOCAL_LO(%a0){#21:#2},%d3 |dbl-prec. g-r are 2 bits right + movel #30,%d2 |of the dbl prec. limits + lsll %d2,%d3 |shift g-r bits to the MSB of d3 + movel LOCAL_LO(%a0),%d2 |get lower mantissa for s-bit test + andil #0x000001ff,%d2 |s bit is the or-ing of all + bnes st_stky |other bits to the right of g-r + tstl %d0 |test word original g,r,s + bnes st_stky |if any are set, set sticky + bras end_sd |if clear, exit +st_stky: + bset #rnd_stky_bit,%d3 +end_sd: + movel %d3,%d0 |return grs to d0 + moveml (%a7)+,%d2/%d3 |restore scratch registers +end_ext_grs: + swap %d1 |restore d1 to original + rts + +|******************* Local Equates + .set ad_1_sgl,0x00000100 | constant to add 1 to l-bit in sgl prec + .set ad_1_dbl,0x00000800 | constant to add 1 to l-bit in dbl prec + + +|Jump table for adding 1 to the l-bit indexed by rnd prec + +add_to_l: + .long add_ext + .long add_sgl + .long add_dbl + .long add_dbl +| +| ADD SINGLE +| +add_sgl: + addl #ad_1_sgl,LOCAL_HI(%a0) + bccs scc_clr |no mantissa overflow + roxrw LOCAL_HI(%a0) |shift v-bit back in + roxrw LOCAL_HI+2(%a0) |shift v-bit back in + addw #0x1,LOCAL_EX(%a0) |and incr exponent +scc_clr: + tstl %d0 |test for rs = 0 + bnes sgl_done + andiw #0xfe00,LOCAL_HI+2(%a0) |clear the l-bit +sgl_done: + andil #0xffffff00,LOCAL_HI(%a0) |truncate bits beyond sgl limit + clrl LOCAL_LO(%a0) |clear d2 + rts + +| +| ADD EXTENDED +| +add_ext: + addql #1,LOCAL_LO(%a0) |add 1 to l-bit + bccs xcc_clr |test for carry out + addql #1,LOCAL_HI(%a0) |propagate carry + bccs xcc_clr + roxrw LOCAL_HI(%a0) |mant is 0 so restore v-bit + roxrw LOCAL_HI+2(%a0) |mant is 0 so restore v-bit + roxrw LOCAL_LO(%a0) + roxrw LOCAL_LO+2(%a0) + addw #0x1,LOCAL_EX(%a0) |and inc exp +xcc_clr: + tstl %d0 |test rs = 0 + bnes add_ext_done + andib #0xfe,LOCAL_LO+3(%a0) |clear the l bit +add_ext_done: + rts +| +| ADD DOUBLE +| +add_dbl: + addl #ad_1_dbl,LOCAL_LO(%a0) + bccs dcc_clr + addql #1,LOCAL_HI(%a0) |propagate carry + bccs dcc_clr + roxrw LOCAL_HI(%a0) |mant is 0 so restore v-bit + roxrw LOCAL_HI+2(%a0) |mant is 0 so restore v-bit + roxrw LOCAL_LO(%a0) + roxrw LOCAL_LO+2(%a0) + addw #0x1,LOCAL_EX(%a0) |incr exponent +dcc_clr: + tstl %d0 |test for rs = 0 + bnes dbl_done + andiw #0xf000,LOCAL_LO+2(%a0) |clear the l-bit + +dbl_done: + andil #0xfffff800,LOCAL_LO(%a0) |truncate bits beyond dbl limit + rts + +error: + rts +| +| Truncate all other bits +| +trunct: + .long end_rnd + .long sgl_done + .long dbl_done + .long dbl_done + +truncate: + lea trunct,%a1 + movel (%a1,%d1.w*4),%a1 + jmp (%a1) + +end_rnd: + rts + +| +| NORMALIZE +| +| These routines (nrm_zero & nrm_set) normalize the unnorm. This +| is done by shifting the mantissa left while decrementing the +| exponent. +| +| NRM_SET shifts and decrements until there is a 1 set in the integer +| bit of the mantissa (msb in d1). +| +| NRM_ZERO shifts and decrements until there is a 1 set in the integer +| bit of the mantissa (msb in d1) unless this would mean the exponent +| would go less than 0. In that case the number becomes a denorm - the +| exponent (d0) is set to 0 and the mantissa (d1 & d2) is not +| normalized. +| +| Note that both routines have been optimized (for the worst case) and +| therefore do not have the easy to follow decrement/shift loop. +| +| NRM_ZERO +| +| Distance to first 1 bit in mantissa = X +| Distance to 0 from exponent = Y +| If X < Y +| Then +| nrm_set +| Else +| shift mantissa by Y +| set exponent = 0 +| +|input: +| FP_SCR1 = exponent, ms mantissa part, ls mantissa part +|output: +| L_SCR1{4} = fpte15 or ete15 bit +| + .global nrm_zero +nrm_zero: + movew LOCAL_EX(%a0),%d0 + cmpw #64,%d0 |see if exp > 64 + bmis d0_less + bsr nrm_set |exp > 64 so exp won't exceed 0 + rts +d0_less: + moveml %d2/%d3/%d5/%d6,-(%a7) + movel LOCAL_HI(%a0),%d1 + movel LOCAL_LO(%a0),%d2 + + bfffo %d1{#0:#32},%d3 |get the distance to the first 1 +| ;in ms mant + beqs ms_clr |branch if no bits were set + cmpw %d3,%d0 |of X>Y + bmis greater |then exp will go past 0 (neg) if +| ;it is just shifted + bsr nrm_set |else exp won't go past 0 + moveml (%a7)+,%d2/%d3/%d5/%d6 + rts +greater: + movel %d2,%d6 |save ls mant in d6 + lsll %d0,%d2 |shift ls mant by count + lsll %d0,%d1 |shift ms mant by count + movel #32,%d5 + subl %d0,%d5 |make op a denorm by shifting bits + lsrl %d5,%d6 |by the number in the exp, then +| ;set exp = 0. + orl %d6,%d1 |shift the ls mant bits into the ms mant + movel #0,%d0 |same as if decremented exp to 0 +| ;while shifting + movew %d0,LOCAL_EX(%a0) + movel %d1,LOCAL_HI(%a0) + movel %d2,LOCAL_LO(%a0) + moveml (%a7)+,%d2/%d3/%d5/%d6 + rts +ms_clr: + bfffo %d2{#0:#32},%d3 |check if any bits set in ls mant + beqs all_clr |branch if none set + addw #32,%d3 + cmpw %d3,%d0 |if X>Y + bmis greater |then branch + bsr nrm_set |else exp won't go past 0 + moveml (%a7)+,%d2/%d3/%d5/%d6 + rts +all_clr: + movew #0,LOCAL_EX(%a0) |no mantissa bits set. Set exp = 0. + moveml (%a7)+,%d2/%d3/%d5/%d6 + rts +| +| NRM_SET +| + .global nrm_set +nrm_set: + movel %d7,-(%a7) + bfffo LOCAL_HI(%a0){#0:#32},%d7 |find first 1 in ms mant to d7) + beqs lower |branch if ms mant is all 0's + + movel %d6,-(%a7) + + subw %d7,LOCAL_EX(%a0) |sub exponent by count + movel LOCAL_HI(%a0),%d0 |d0 has ms mant + movel LOCAL_LO(%a0),%d1 |d1 has ls mant + + lsll %d7,%d0 |shift first 1 to j bit position + movel %d1,%d6 |copy ls mant into d6 + lsll %d7,%d6 |shift ls mant by count + movel %d6,LOCAL_LO(%a0) |store ls mant into memory + moveql #32,%d6 + subl %d7,%d6 |continue shift + lsrl %d6,%d1 |shift off all bits but those that will +| ;be shifted into ms mant + orl %d1,%d0 |shift the ls mant bits into the ms mant + movel %d0,LOCAL_HI(%a0) |store ms mant into memory + moveml (%a7)+,%d7/%d6 |restore registers + rts + +| +| We get here if ms mant was = 0, and we assume ls mant has bits +| set (otherwise this would have been tagged a zero not a denorm). +| +lower: + movew LOCAL_EX(%a0),%d0 |d0 has exponent + movel LOCAL_LO(%a0),%d1 |d1 has ls mant + subw #32,%d0 |account for ms mant being all zeros + bfffo %d1{#0:#32},%d7 |find first 1 in ls mant to d7) + subw %d7,%d0 |subtract shift count from exp + lsll %d7,%d1 |shift first 1 to integer bit in ms mant + movew %d0,LOCAL_EX(%a0) |store ms mant + movel %d1,LOCAL_HI(%a0) |store exp + clrl LOCAL_LO(%a0) |clear ls mant + movel (%a7)+,%d7 + rts +| +| denorm --- denormalize an intermediate result +| +| Used by underflow. +| +| Input: +| a0 points to the operand to be denormalized +| (in the internal extended format) +| +| d0: rounding precision +| Output: +| a0 points to the denormalized result +| (in the internal extended format) +| +| d0 is guard,round,sticky +| +| d0 comes into this routine with the rounding precision. It +| is then loaded with the denormalized exponent threshold for the +| rounding precision. +| + + .global denorm +denorm: + btstb #6,LOCAL_EX(%a0) |check for exponents between $7fff-$4000 + beqs no_sgn_ext + bsetb #7,LOCAL_EX(%a0) |sign extend if it is so +no_sgn_ext: + + cmpib #0,%d0 |if 0 then extended precision + bnes not_ext |else branch + + clrl %d1 |load d1 with ext threshold + clrl %d0 |clear the sticky flag + bsr dnrm_lp |denormalize the number + tstb %d1 |check for inex + beq no_inex |if clr, no inex + bras dnrm_inex |if set, set inex + +not_ext: + cmpil #1,%d0 |if 1 then single precision + beqs load_sgl |else must be 2, double prec + +load_dbl: + movew #dbl_thresh,%d1 |put copy of threshold in d1 + movel %d1,%d0 |copy d1 into d0 + subw LOCAL_EX(%a0),%d0 |diff = threshold - exp + cmpw #67,%d0 |if diff > 67 (mant + grs bits) + bpls chk_stky |then branch (all bits would be +| ; shifted off in denorm routine) + clrl %d0 |else clear the sticky flag + bsr dnrm_lp |denormalize the number + tstb %d1 |check flag + beqs no_inex |if clr, no inex + bras dnrm_inex |if set, set inex + +load_sgl: + movew #sgl_thresh,%d1 |put copy of threshold in d1 + movel %d1,%d0 |copy d1 into d0 + subw LOCAL_EX(%a0),%d0 |diff = threshold - exp + cmpw #67,%d0 |if diff > 67 (mant + grs bits) + bpls chk_stky |then branch (all bits would be +| ; shifted off in denorm routine) + clrl %d0 |else clear the sticky flag + bsr dnrm_lp |denormalize the number + tstb %d1 |check flag + beqs no_inex |if clr, no inex + bras dnrm_inex |if set, set inex + +chk_stky: + tstl LOCAL_HI(%a0) |check for any bits set + bnes set_stky + tstl LOCAL_LO(%a0) |check for any bits set + bnes set_stky + bras clr_mant +set_stky: + orl #inx2a_mask,USER_FPSR(%a6) |set inex2/ainex + movel #0x20000000,%d0 |set sticky bit in return value +clr_mant: + movew %d1,LOCAL_EX(%a0) |load exp with threshold + movel #0,LOCAL_HI(%a0) |set d1 = 0 (ms mantissa) + movel #0,LOCAL_LO(%a0) |set d2 = 0 (ms mantissa) + rts +dnrm_inex: + orl #inx2a_mask,USER_FPSR(%a6) |set inex2/ainex +no_inex: + rts + +| +| dnrm_lp --- normalize exponent/mantissa to specified threshold +| +| Input: +| a0 points to the operand to be denormalized +| d0{31:29} initial guard,round,sticky +| d1{15:0} denormalization threshold +| Output: +| a0 points to the denormalized operand +| d0{31:29} final guard,round,sticky +| d1.b inexact flag: all ones means inexact result +| +| The LOCAL_LO and LOCAL_GRS parts of the value are copied to FP_SCR2 +| so that bfext can be used to extract the new low part of the mantissa. +| Dnrm_lp can be called with a0 pointing to ETEMP or WBTEMP and there +| is no LOCAL_GRS scratch word following it on the fsave frame. +| + .global dnrm_lp +dnrm_lp: + movel %d2,-(%sp) |save d2 for temp use + btstb #E3,E_BYTE(%a6) |test for type E3 exception + beqs not_E3 |not type E3 exception + bfextu WBTEMP_GRS(%a6){#6:#3},%d2 |extract guard,round, sticky bit + movel #29,%d0 + lsll %d0,%d2 |shift g,r,s to their positions + movel %d2,%d0 +not_E3: + movel (%sp)+,%d2 |restore d2 + movel LOCAL_LO(%a0),FP_SCR2+LOCAL_LO(%a6) + movel %d0,FP_SCR2+LOCAL_GRS(%a6) + movel %d1,%d0 |copy the denorm threshold + subw LOCAL_EX(%a0),%d1 |d1 = threshold - uns exponent + bles no_lp |d1 <= 0 + cmpw #32,%d1 + blts case_1 |0 = d1 < 32 + cmpw #64,%d1 + blts case_2 |32 <= d1 < 64 + bra case_3 |d1 >= 64 +| +| No normalization necessary +| +no_lp: + clrb %d1 |set no inex2 reported + movel FP_SCR2+LOCAL_GRS(%a6),%d0 |restore original g,r,s + rts +| +| case (0= 64 Force the exponent to be the denorm threshold with the +| correct sign. +| +case_3: + movew %d0,LOCAL_EX(%a0) + tstw LOCAL_SGN(%a0) + bges c3con +c3neg: + orl #0x80000000,LOCAL_EX(%a0) +c3con: + cmpw #64,%d1 + beqs sixty_four + cmpw #65,%d1 + beqs sixty_five +| +| Shift value is out of range. Set d1 for inex2 flag and +| return a zero with the given threshold. +| + clrl LOCAL_HI(%a0) + clrl LOCAL_LO(%a0) + movel #0x20000000,%d0 + st %d1 + rts + +sixty_four: + movel LOCAL_HI(%a0),%d0 + bfextu %d0{#2:#30},%d1 + andil #0xc0000000,%d0 + bras c3com + +sixty_five: + movel LOCAL_HI(%a0),%d0 + bfextu %d0{#1:#31},%d1 + andil #0x80000000,%d0 + lsrl #1,%d0 |shift high bit into R bit + +c3com: + tstl %d1 + bnes c3ssticky + tstl LOCAL_LO(%a0) + bnes c3ssticky + tstb FP_SCR2+LOCAL_GRS(%a6) + bnes c3ssticky + clrb %d1 + bras c3end + +c3ssticky: + bsetl #rnd_stky_bit,%d0 + st %d1 +c3end: + clrl LOCAL_HI(%a0) + clrl LOCAL_LO(%a0) + rts + + |end -- cgit v1.2.3