|
|
subttl emftran.asm - Transcendental instructions page ;******************************************************************************* ; Copyright (c) Microsoft Corporation 1991 ; All Rights Reserved ; ;emftran.asm - Transcendental instructions ; by Tim Paterson ; ;Purpose: ; F2XM1, FPATAN, FYL2X, FYL2XP1 instructions ;Inputs: ; edi = [CURstk] ; ;Revision History: ; ; [] 09/05/91 TP Initial 32-bit version. ; ;*******************************************************************************
;********************* Polynomial Coefficients *********************
;These polynomial coefficients were all taken from "Computer Approximations" ;by J.F. Hart (reprinted 1978 w/corrections). All calculations and ;conversions to hexadecimal were done with a character-string calculator ;written in Visual Basic with precision set to 30 digits. Once the constants ;were typed into this file, all transfers were done with cut-and-paste ;operations to and from the calculator to help eliminate any typographical ;errors.
tAtanPoly label word
;These constants are from Hart #5056: atan(x) = x * P(x^2) / Q(x^2), ;accurate to 20.78 digits over interval [0, tan(pi/12)].
dd 4 ;P() is degree four
; Hart constant ; ;+.16241 70218 72227 96595 08 E0 ;Hex value: 0.A650A5D5050DE43A2C25A8C00 HFFFE dq 0A650A5D5050DE43AH dw bTAG_VALID,0FFFEH-1
;+.65293 76545 29069 63960 675 E1 ;Hex value: 0.D0F0A714A9604993AC4AC49A0 H3 dq 0D0F0A714A9604994H dw bTAG_VALID,03H-1
;+.39072 57269 45281 71734 92684 E2 ;Hex value: 0.9C4A507F16530AC3CDDEFA3DE H6 dq 09C4A507F16530AC4H dw bTAG_VALID,06H-1
;+.72468 55912 17450 17145 90416 9 E2 ;Hex value: 0.90EFE6FB30465042CF089D1310 H7 dq 090EFE6FB30465043H dw bTAG_VALID,07H-1
;+.41066 29181 34876 24224 77349 62 E2 ;Hex value: 0.A443E2004BB000B84A5154D44 H6 dq 0A443E2004BB000B8H dw bTAG_VALID,06H-1
dd 4 ;Q() is degree four
; Hart constant ; ;+.15023 99905 56978 85827 4928 E2 ;Hex value: 0.F0624CD575B782643AFB912D0 H4 dq 0F0624CD575B78264H dw bTAG_VALID,04H-1
;+.59578 42201 83554 49303 22456 E2 ;Hex value: 0.EE504DDC907DEAEB7D7473B82 H6 dq 0EE504DDC907DEAEBH dw bTAG_VALID,06H-1
;+.86157 32305 95742 25062 42472 E2 ;Hex value: 0.AC508CA5E78E504AB2032E864 H7 dq 0AC508CA5E78E504BH dw bTAG_VALID,07H-1
;+.41066 29181 34876 24224 84140 84 E2 ;Hex value: 0.A443E2004BB000B84F542813C H6 dq 0A443E2004BB000B8H dw bTAG_VALID,06H-1
;tan(pi/12) = tan(15 deg.) = 2 - sqrt(3) ;= 0.26794 91924 31122 70647 25536 58494 12763 ;From Hart appendix ;Hex value: 0.8930A2F4F66AB189B517A51F2 HFFFF Tan15Hi equ 08930A2F4H Tan15Lo equ 0F66AB18AH Tan15exp equ 0FFFFH-1
;1/tan(pi/6) = sqrt(3) = 1.73205 08075 68877 29352 74463 41505 87236 ;From Hart appendix ;Hex value: 0.DDB3D742C265539D92BA16B8 H1 Sqrt3Hi equ 0DDB3D742H Sqrt3Lo equ 0C265539EH Sqrt3exp equ 01H-1
;pi = +3.14159265358979323846264338328 ;Hex value: 0.C90FDAA22168C234C4C6628B8 H2 PiHi equ 0C90FDAA2H PiLo equ 02168C235H PiExp equ 02H-1
;3*pi = +9.42477796076937971538793014984 ;Hex value: 0.96CBE3F9990E91A79394C9E890 H4 XThreePiHi equ 096CBE3F9H XThreePiMid equ 0990E91A7H XThreePiLo equ 090000000H ThreePiExp equ 04H-1
;This is a table of multiples of pi/6. It is used to adjust the ;final result angle after atan(). Derived from Hart appendix ;pi/180 = 0.01745 32925 19943 29576 92369 07684 88612 ; ;When the reduced argument for atan() is very small, these correction ;constants simply become the result. These constants have all been ;rounded to nearest, but the user may have selected a different rounding ;mode. The tag byte is not needed for these constants, so its space ;is used to indicate if it was rounded. To determine if a constant ;was rounded, 7FH is subtracted from this flag; CY set means it was ;rounded up.
RoundedUp equ 040H RoundedDown equ 0C0H
tAtanPiFrac label dword ;pi/2 = +1.57079632679489661923132169163 ;Hex value: 0.C90FDAA22168C234C4C6628B0 H1 dq 0C90FDAA22168C235H dw RoundedUp,01H-1
;2*pi/3 = +2.09439510239319549230842892218 ;Hex value: 0.860A91C16B9B2C232DD997078 H2 dq 0860A91C16B9B2C23H dw RoundedDown,02H-1
;none dd 0,0,0
;pi/6 = +0.523598775598298873077107230544E0 ;Hex value: 0.860A91C16B9B2C232DD99707A H0 dq 0860A91C16B9B2C23H dw RoundedDown,00H-1
;pi/2 = +1.57079632679489661923132169163 ;Hex value: 0.C90FDAA22168C234C4C6628B0 H1 dq 0C90FDAA22168C235H dw RoundedUp,01H-1
;pi/3 = +1.04719755119659774615421446109 ;Hex value: 0.860A91C16B9B2C232DD997078 H1 dq 0860A91C16B9B2C23H dw RoundedDown,01H-1
;pi = +3.14159265358979323846264338328 ;Hex value: 0.C90FDAA22168C234C4C6628B8 H2 dq 0C90FDAA22168C235H dw RoundedUp,02H-1
;5*pi/6 = +2.61799387799149436538553615272 ;Hex value: 0.A78D3631C681F72BF94FFCC96 H2 dq 0A78D3631C681F72CH dw RoundedUp,02H-1
;*********************
tExpPoly label word
;These constants are from Hart #1324: 2^x - 1 = ; 2 * x * P(x^2) / ( Q(x^2) - x * P(x^2) ) ;accurate to 21.54 digits over interval [0, 0.5].
dd 2 ;P() is degree two
; Hart constant ; ;+.60613 30790 74800 42574 84896 07 E2 ;Hex value: 0.F27406FCF405189818F68BB78 H6 dq 0F27406FCF4051898H dw bTAG_VALID,06H-1
;+.30285 61978 21164 59206 24269 927 E5 ;Hex value: 0.EC9B3D5414E1AD0852E432A18 HF dq 0EC9B3D5414E1AD08H dw bTAG_VALID,0FH-1
;+.20802 83036 50596 27128 55955 242 E7 ;Hex value: 0.FDF0D84AC3A35FAF89A690CC4 H15 dq 0FDF0D84AC3A35FB0H dw bTAG_VALID,015H-1
dd 3 ;Q() is degree three. First ;coefficient is 1.0 and is not listed. ; Hart constant ; ;+.17492 20769 51057 14558 99141 717 E4 ;Hex value: 0.DAA7108B387B776F212ECFBEC HB dq 0DAA7108B387B776FH dw bTAG_VALID,0BH-1
;+.32770 95471 93281 18053 40200 719 E6 ;Hex value: 0.A003B1829B7BE85CC81BD5309 H13 dq 0A003B1829B7BE85DH dw bTAG_VALID,013H-1
;+.60024 28040 82517 36653 36946 908 E7 ;Hex value: 0.B72DF814E709837E066855BDD H17 dq 0B72DF814E709837EH dw bTAG_VALID,017H-1
;sqrt(2) = 1.41421 35623 73095 04880 16887 24209 69808 ;From Hart appendix ;Hex value: 0.B504F333F9DE6484597D89B30 H1 Sqrt2Hi equ 0B504F333H Sqrt2Lo equ 0F9DE6484H Sqrt2Exp equ 01H-1
;sqrt(2) - 1 = +0.4142135623730950488016887242E0 ;Hex value: 0.D413CCCFE779921165F626CC4 HFFFF Sqrt2m1Hi equ 0D413CCCFH Sqrt2m1Lo equ 0E7799211H XSqrt2m1Lo equ 060000000H Sqrt2m1Exp equ 0FFFFH-1
;2 - sqrt(2) = +0.5857864376269049511983112758E0 ;Hex value: 0.95F619980C4336F74D04EC9A0 H0 TwoMinusSqrt2Hi equ 095F61998H TwoMinusSqrt2Lo equ 00C4336F7H TwoMinusSqrt2Exp equ 00H-1
;*********************
tLogPoly label dword
;These constants are derived from Hart #2355: log2(x) = z * P(z^2) / Q(z^2), ; z = (x+1) / (x-1) accurate to 19.74 digits over interval ;[1/sqrt(2), sqrt(2)]. The original Hart coefficients were for log10(); ;the P() coefficients have been scaled by log2(10) to compute log2(). ; ;log2(10) = 3.32192 80948 87362 34787 03194 29489 39017 ;From Hart appendix
dd 3 ;P() is degree three
; Original Hart constant Scaled value ; ;+.18287 59212 09199 9337 E0 +0.607500660543248917834110566373E0 ;Hex value: 0.9B8529CD54E72022A12BAEC53 H0 dq 09B8529CD54E72023H dw bTAG_VALID,00H-1
;-.41855 96001 31266 20633 E1 -13.9042489506087332809657007634 ;Hex value: 0.DE77CDBF64E8C53F0DCD458D0 H4 dq 0DE77CDBF64E8C53FH dw bSign shl 8 + bTAG_VALID,04H-1
;+.13444 58152 27503 62236 E2 +44.6619330844279438866067340334 ;Hex value: 0.B2A5D1C95708A0C9FE50F6F97 H6 dq 0B2A5D1C95708A0CAH dw bTAG_VALID,06H-1
;-.10429 11213 72526 69497 44122 E2 -34.6447606134704282123622236943 ;Hex value: 0.8A943C20526AE439A98B30F6A H6 dq 08A943C20526AE43AH dw bSign shl 8 + bTAG_VALID,06H-1
dd 3 ;Q() is degree three. First ;coefficient is 1.0 and is not listed. ; Hart constant ; ;-.89111 09060 90270 85654 E1 ;Hex value: 0.8E93E7183AA998D74F45CDFF0 H4 dq 08E93E7183AA998D7H dw bSign shl 8 + bTAG_VALID,04H-1
;+.19480 96618 79809 36524 155 E2 ;Hex value: 0.9BD904CCFEE118D4BEF319716 H5 dq 09BD904CCFEE118D5H dw bTAG_VALID,05H-1
;-.12006 95907 02006 34243 4218 E2 ;Hex value: 0.C01C811D2EC1B5806304B1858 H4 dq 0C01C811D2EC1B580H dw bSign shl 8 + bTAG_VALID,04H-1
;Log2(e) = 1.44269 50408 88963 40735 99246 81001 89213 ;From Hart appendix ;Hex value: 0.B8AA3B295C17F0BBBE87FED04 H1 Log2OfEHi equ 0B8AA3B29H Log2OfELo equ 05C17F0BCH Log2OfEexp equ 01H-1
;********************* Generic polynomial evaluation ********************* ; ;EvalPoly, EvalPolyAdd, EvalPolySetup, Eval2Poly ; ;Inputs: ; ebx:esi,ecx = floating point number, internal format ; edi = pointer to polynomial degree and coefficients ;Outputs: ; result in ebx:esi,ecx ; edi incremented to start of last coefficient in list ; ;EvalPoly is the basic polynomial evaluator, using Horner's rule. The ;polynomial pointer in edi points to a list: the first dword in the list ;is the degree of the polynomial (n); it is followed by the n+1 ;coefficients in internal (12-byte) format. The argment for EvalPoly ;must be stored in the static FloatTemp in addition to being in ;registers. ; ;EvalPolyAdd is an alternate entry point into the middle of EvalPoly. ;It is used when the first coefficient is 1.0, so it skips the first ;multiplication. It requires that the degree of the polynomial be ;already loaded into ebp. ; ;EvalPolySetup store a copy of the argument in the static ArgTemp, ;and stores the square of the argument in the static FloatTemp. ;Then it falls into EvalPoly to evaluate the polynomial on the square. ; ;Eval2Poly evaluate two polynomials on its argument. The first ;polynomial is x * P(x^2), and its result is left at [[CURstk]]. ;The second polynomial is Q(x^2), and its result is left in registers. ;The most significant coefficient of Q() is 1. ; ;Polynomial evaluation uses a slight variation on the standard add ;and multiply routines. PolyAddDouble and PolyMulDouble both check ;to see if the argument in registers (the current accumulation) is ;zero. The argument pointed to by edi is a coefficient and is never ;zero. ; ;In addition, the [RoundMode] and [ZeroVector] vectors are "trapped", ;i.e., redirected to special handlers for polynomial evaluation. ;[RoundMode] ordinarily points to the routine that handles the ;the current rounding mode and precision control; however, during ;polynomial evaluation, we always want full precision and round ;nearest. The normal rounding routines also store their result ;at [[Result]], but we want the result left in registers. ;[ZeroVector] exists solely so polynomial evaluation can trap ;when AddDouble results of zero. The normal response is to store ;a zero at [[Result]], but we need the zero left in registers. ;PolyRound and PolyZero handle these traps.
EvalPolySetup: ;Save x in ArgTemp mov EMSEG:[ArgTemp].ExpSgn,ecx mov EMSEG:[ArgTemp].lManHi,ebx mov EMSEG:[ArgTemp].lManLo,esi mov EMSEG:[RoundMode],offset PolyRound mov EMSEG:[ZeroVector],offset PolyZero push edi ;Save pointer to polynomials ;op1 mantissa in ebx:esi, exponent in high ecx, sign in ch bit 7 mov edx,ebx mov edi,esi mov eax,ecx ;op2 mantissa in edx:edi, exponent in high eax, sign in ah bit 7 call MulDoubleReg ;Compute x^2 ;Save x^2 in FloatTemp mov EMSEG:[FloatTemp].ExpSgn,ecx mov EMSEG:[FloatTemp].lManHi,ebx mov EMSEG:[FloatTemp].lManLo,esi pop edi EvalPoly: ;ebx:esi,ecx = arg to evaluate, also in FloatTemp ;edi = pointer to degree and list of coefficients. push edi mov eax,cs:[edi+4].ExpSgn mov edx,cs:[edi+4].lManHi mov edi,cs:[edi+4].lManLo call MulDoubleReg ;Multiply arg by first coef. pop edi mov ebp,cs:[edi] ;Get polynomial degree add edi,4+Reg87Len ;Point to second coefficient jmp EvalPolyAdd
PolyLoop: push ebp ;Save loop count ifdef NT386 mov edi,YFloatTemp else mov edi,offset edata:FloatTemp endif call PolyMulDouble pop ebp pop edi add di,Reg87Len EvalPolyAdd: push edi mov eax,cs:[edi].ExpSgn mov edx,cs:[edi].lManHi mov edi,cs:[edi].lManLo cmp cl,bTAG_ZERO ;Adding to zero? jz AddToZero call AddDoubleReg ;ebp preserved ContPolyLoop: dec ebp jnz PolyLoop pop edi ret
AddToZero: ;Number in registers is zero, so just return value from memory. mov ecx,eax mov ebx,edx mov esi,edi jmp ContPolyLoop
Eval2Poly: call EvalPolySetup push edi ifdef NT386 mov edi,YArgTemp else mov edi,offset edata:ArgTemp endif call PolyMulDouble ;Multiply first result by argument pop edi ;Save result of first polynomial at [[CURstk]] mov edx,EMSEG:[CURstk] mov EMSEG:[edx].ExpSgn,ecx mov EMSEG:[edx].lManHi,ebx mov EMSEG:[edx].lManLo,esi ;Load x^2 back into registers mov ecx,EMSEG:[FloatTemp].ExpSgn mov ebx,EMSEG:[FloatTemp].lManHi mov esi,EMSEG:[FloatTemp].lManLo ;Start second polynomial evaluation add edi,4+Reg87Len ;Point to coefficient mov ebp,cs:[edi-4] ;Get polynomial degree jmp EvalPolyAdd
PolyRound: ;This routine handles all rounding during polynomial evaluation. ;It performs 64-but round nearest, with result left in registers. ; ;Inputs: ; mantissa in ebx:esi:eax, exponent in high ecx, sign in ch bit 7 ;Outputs: ; same, plus tag in cl. ; ;To perform "round even" when the round bit is set and the sticky bits ;are zero, we treat the LSB as if it were a sticky bit. Thus if the LSB ;is set, that will always force a round up (to even) if the round bit is ;set. If the LSB is zero, then the sticky bits remain zero and we always ;round down. This rounding rule is implemented by adding RoundBit-1 ;(7F..FFH), setting CY if round up. ; ;This routine needs to be reversible in case we're at the last step ;in the polynomial and final rounding uses a different rounding mode. ;We do this by copying the LSB of esi into al. While the rounding is ;reversible, you can't tell if the answer was exact.
mov edx,esi and dl,1 ;Look at LSB or al,dl ;Set LSB as sticky bit add eax,(1 shl 31)-1 ;Sum LSB & sticky bits--CY if round up adc esi,0 adc ebx,0 jc PolyBumpExponent ;Overflowed, increment exponent or esi,esi ;Any bits in low half? .erre bTAG_VALID eq 1 .erre bTAG_SNGL eq 0 setnz cl ;if low half==0 then cl=0 else cl=1 ret
PolyBumpExponent: add ecx,1 shl 16 ;Mantissa overflowed, bump exponent or ebx,1 shl 31 ;Set MSB mov cl,bTAG_SNGL PolyZero: ;Enter here when result is zero ret
;*******************************************************************************
;FPATAN instruction
;Actual instruction entry point is in emarith.asm
tFpatanDisp label dword ;Source (ST(0)) Dest (*[di] = ST(1)) dd AtanDouble ;single single dd AtanDouble ;single double dd AtanZeroDest ;single zero dd AtanSpclDest ;single special dd AtanDouble ;double single dd AtanDouble ;double double dd AtanZeroDest ;double zero dd AtanSpclDest ;double special dd AtanZeroSource ;zero single dd AtanZeroSource ;zero double dd AtanZeroDest ;zero zero dd AtanSpclDest ;zero special dd AtanSpclSource ;special single dd AtanSpclSource ;special double dd AtanSpclSource ;special zero dd TwoOpBothSpcl ;special special dd AtanTwoInf ;Two infinites
;Compute atan( st(1)/st(0) ). Neither st(0) or st(1) are zero or ;infinity at this point. ; ;Argument reduction starts by dividing the smaller by the larger, ;ensuring that the result x is <= 1. The absolute value of the quotient ;is used and the quadrant is fixed up later. If x = st(0)/st(1), then ;the final atan result is subtracted from pi/2 (and normalized for the ;correct range of -pi to +pi). ; ;The range of x is further reduced using the formulas: ; t = (x - k) / (1 + kx) ; atan(x) = atan(k) + atan(t) ; ;Given that x <= 1, if we choose k = tan(pi/6) = 1/sqrt(3), then we ;are assured that t <= tan(pi/12) = 2 - sqrt(3), and ;for x >= tan(pi/12) = 2 - sqrt(3), t >= -tan(pi/12). ;Thus we can always reduce the argument to abs(t) <= tan(pi/12). ; ;Since k = 1/sqrt(3), it is convenient to multiply the numerator ;and denominator of t by 1/k, which gives ;t = (x/k - 1) / (1/k + x) = ( x*sqrt(3) - 1 ) / ( sqrt(3) + x ). ;This is the form found in Cody and Waite and in previous versions ;of the emulator. It requires one each add, subtract, multiply, and ;divide. ; ;Hart has derived a simpler version of this formula: ;t = 1/k - (1/k^2 + 1) / (1/k + x) = sqrt(3) - 4 / ( sqrt(3) + x ). ;Note that this computation requires one each add, subtract, and ;divide, but no multiply.
;st(0) mantissa in ebx:esi, exponent in high ecx, sign in ch bit 7 ;[edi] points to st(1), where result is returned
AtanDouble: mov EMSEG:[Result],edi mov EMSEG:[RoundMode],offset PolyRound mov EMSEG:[ZeroVector],offset PolyZero mov ah,EMSEG:[edi].bSgn ;Sign of result mov al,ch ;Affects quadrant of result and al,bSign ;Zero other bits, used as flags push eax ;Save flag ;First figure out which is larger push offset AtanQuo ;Return address for DivDouble shld edx,ecx,16 ;Get exponent to ax cmp dx,EMSEG:[edi].wExp ;Compare exponents jl DivrDoubleSetFlag ;ST(0) is smaller, make it dividend jg DivDouble ; ...is bigger, make it divisor ;Exponents are equal, compare mantissas cmp ebx,EMSEG:[edi].lManHi jb DivrDoubleSetFlag ;ST(0) is smaller, make it dividend ja DivDouble ; ...is bigger, make it divisor cmp esi,EMSEG:[edi].lManLo jbe DivrDoubleSetFlag ;ST(0) is smaller, make it dividend jmp DivDouble
TinyAtan: ;Come here if the angle was reduced to zero, or the divide resulted in ;unmasked underflow so that the quotient exponent was biased. ;Note that an angle of zero means reduction was performed, and the ;result will be corrected to a non-zero value. mov dl,[esp] ;Get flag byte or dl,dl ;No correction needed? jz AtanSetSign ;Just return result of divide and EMSEG:[CURerr],not Underflow ;Angle in registers is too small to affect correction amount. Just ;load up correction angle instead of adding it in. add dl,40H ;Change flags for correction lookup shr dl,5-2 ;Now in bits 2,3,4 and edx,7 shl 2 mov ebx,[edx+2*edx+tAtanPiFrac].lManHi mov esi,[edx+2*edx+tAtanPiFrac].lManLo mov ecx,[edx+2*edx+tAtanPiFrac].ExpSgn shrd eax,ecx,8 ;Copy rounding flag to high eax jmp AtanSetSign
AtanQuo: ;Return here after divide. Underflow flag is set only for "big underflow", ;meaning the (15-bit) exponent couldn't even be kept in 16 bits. This can ;only happen dividing a denormal by one of the largest numbers. ; ;Rounded mantissa in ebx:esi:eax, exp/sign in high ecx test EMSEG:[CURerr],Underflow;Did we underflow? jnz TinyAtan ;Now compare quotient in ebx:esi,ecx with tan(pi/12) = 2 - sqrt(3) xor cx,cx ;Use absolute value cmp ecx,Tan15exp shl 16 jg AtnNeedReduce jl AtnReduced cmp ebx,Tan15Hi ja AtnNeedReduce jb AtnReduced cmp esi,Tan15Lo jbe AtnReduced AtnNeedReduce: or byte ptr [esp],20H ;Note reduction in flags on stack ;Compute t = sqrt(3) - 4 / ( sqrt(3) + x ). mov eax,Sqrt3exp shl 16 mov edx,Sqrt3Hi mov edi,Sqrt3Lo call AddDoubleReg ;x + sqrt(3) mov edi,esi mov esi,ebx ;Mantissa in esi:edi mov ebx,ecx ;ExpSgn to ebx mov ecx,(2+TexpBias) shl 16 mov edx,1 shl 31 xor eax,eax ;edx:edi,eax = 4.0 ;dividend mantissa in edx:eax, exponent in high ecx, sign in ch bit 7 ;divisor mantissa in esi:edi, exponent in high ebx, sign in bh bit 7 call DivDoubleReg ;4 / ( x + sqrt(3) ) not ch ;Flip sign mov eax,Sqrt3exp shl 16 mov edx,Sqrt3Hi mov edi,Sqrt3Lo call AddDoubleReg ;sqrt(3) - 4 / ( x + sqrt(3) ) ;Result in ebx:esi,ecx could be very small (or zero) if arg was near tan(pi/6). cmp cl,bTAG_ZERO jz TinyAtan AtnReduced: ;If angle is small, skip the polynomial. atan(x) = x when x - x^3/3 = x ;[or 1 - x^2/3 = 1], which happens when x < 2^-32. This prevents underflow ;in computing x^2. TinyAtanArg equ -32 cmp ecx,TinyAtanArg shl 16 jl AtanCorrection mov edi,offset tAtanPoly call Eval2Poly mov edi,EMSEG:[CURstk] ;Point to first result call DivDouble ;x * P(x^2) / Q(x^2) AtanCorrection: ;Rounded mantissa in ebx:esi:eax, exp/sign in high ecx ; ;Correct sign and add fraction of pi to account for various angle reductions: ; ; flag bit indicates correction ;---------------------------------------------------- ; 5 arg > tan(pi/12) add pi/6 ; 6 st(1) > st(0) sub from pi/2 ; 7 st(0) < 0 sub from pi ; ;This results in the following correction for the result R: ; ;bit 7 6 5 correction ;--------------------------- ; 0 0 0 none ; 0 0 1 pi/6 + R ; 0 1 0 pi/2 - R ; 0 1 1 pi/3 - R ; 1 0 0 pi - R ; 1 0 1 5*pi/6 - R ; 1 1 0 pi/2 + R ; 1 1 1 2*pi/3 + R
mov dl,[esp] ;Get flag byte or dl,dl ;No correction needed? jz AtanSetSign add dl,40H ;Set bit 7 for all -R cases
;This changes the meaning of the flag bits to the following: ; ;bit 7 6 5 correction ;--------------------------- ; 0 0 0 pi/2 + R ; 0 0 1 2*pi/3 + R ; 0 1 0 none ; 0 1 1 pi/6 + R ; 1 0 0 pi/2 - R ; 1 0 1 pi/3 - R ; 1 1 0 pi - R ; 1 1 1 5*pi/6 - R
xor ch,dl ;Flip sign bit in cases 4 - 7 shr dl,5-2 ;Now in bits 2,3,4 and edx,7 shl 2 mov eax,[edx+2*edx+tAtanPiFrac].ExpSgn mov edi,[edx+2*edx+tAtanPiFrac].lManLo mov edx,[edx+2*edx+tAtanPiFrac].lManHi call AddDoubleReg ;Add in correction angle AtanSetSign: pop edx ;Get flags again mov ch,dh ;Set sign to original ST(1) ;Rounded mantissa in ebx:esi:eax, exp/sign in ecx jmp TransUnround
;*** AtanSpclDest: mov al,EMSEG:[edi].bTag ;Pick up tag ; cmp cl,bTAG_INF ;Is argument infinity? cmp al,bTAG_INF ;Is argument infinity? jnz SpclDest ;In emarith.asm AtanZeroSource: ;Dividend is infinity or divisor is zero. Return pi/2 with ;same sign as dividend. mov ecx,(PiExp-1) shl 16 + bTAG_VALID ;Exponent for pi/2 PiMant: ;For storing multiples of pi. Exponent/tag is in ecx. mov ch,EMSEG:[edi].bSgn ;Get dividend's sign mov ebx,XPiHi mov esi,XPiMid mov eax,XPiLo ;A jump through [TransRound] is only valid if the number is known not to ;underflow. Unmasked underflow requires [RoundMode] be set. jmp EMSEG:[TransRound]
;*** AtanSpclSource: cmp cl,bTAG_INF ;Scaling by infinity? jnz SpclSource ;in emarith.asm AtanZeroDest: ;Divisor is infinity or dividend is zero. Return zero for +divisor, ;pi for -divisor. Result sign is same is dividend. or ch,ch ;Check divisor's sign mov ecx,PiExp shl 16 + bTAG_VALID ;Exponent for pi js PiMant ;Store pi ;Result is zero mov EMSEG:[edi].lManHi,0 mov EMSEG:[edi].lManLo,0 mov EMSEG:[edi].wExp,0 mov EMSEG:[edi].bTAG,bTAG_ZERO ret
;*** AtanTwoInf: ;Return pi/4 for +infinity divisor, 3*pi/4 for -infinity divisor. ;Result sign is same is dividend infinity. or ch,ch ;Check divisor's sign mov ecx,(PiExp-2) shl 16 + bTAG_VALID ;Exponent for pi/4 jns PiMant ;Store pi/4 mov ecx,(ThreePiExp-2) shl 16 + bTAG_VALID ;Exponent for 3*pi/4 mov ch,EMSEG:[edi].bSgn ;Get dividend's sign mov ebx,XThreePiHi mov esi,XThreePiMid mov eax,XThreePiLo ;A jump through [TransRound] is only valid if the number is known not to ;underflow. Unmasked underflow requires [RoundMode] be set. jmp EMSEG:[TransRound]
;*******************************************************************************
ExpSpcl: ;Tagged special cmp cl,bTAG_DEN jz ExpDenorm cmp cl,bTAG_INF mov al, cl jnz SpclDestNotDen ;Check for Empty or NAN ;Have infinity, check its sign. ;Return -1 for -infinity, no change if +infinity or ch,ch ;Check sign jns ExpRet ;Just return the +inifinity mov EMSEG:[edi].lManLo,0 mov EMSEG:[edi].lManHi,1 shl 31 mov EMSEG:[edi].ExpSgn,bSign shl 8 + bTAG_SNGL ;-1.0 (exponent is zero) ret
ExpDenorm: mov EMSEG:[CURerr],Denormal test EMSEG:[CWmask],Denormal ;Is denormal exception masked? jnz ExpCont ;Yes, continue ExpRet: ret
EM_ENTRY eF2XM1 eF2XM1: ;edi = [CURstk] mov ecx,EMSEG:[edi].ExpSgn cmp cl,bTAG_ZERO jz ExpRet ;Return same zero ja ExpSpcl ExpCont:
;The input range specified for the function is (-1, +1). The polynomial ;used for this function is valid only over the range [0, +0.5], so range ;reduction is needed. Range reduction is based on the identity: ; ; 2^(a+b) = 2^a * 2^b ; ;1.0 or 0.5 can be added/subtracted from the argument to bring it into ;range. We calculate 2^x - 1 with a polynomial, and then adjust the ;result according to the amount added or subtracted, as shown in the table: ; ;Arg range Adj Polynomial result Required result, 2^x - 1 ; ; (-1, -0.5] +1 P = 2^(x+1) - 1 (P - 1)/2 ; ; (-0.5, 0) +0.5 P = 2^(x+0.5) - 1 P * sqrt(2)/2 + (sqrt(2)/2 - 1) ; ; (0, 0.5) 0 P = 2^x - 1 P ; ; [0.5, 1) -0.5 P = 2^(x-0.5) - 1 P * sqrt(2) + (sqrt(2)-1) ; ;Since the valid input range does not include +1.0 or -1.0, and zero is ;handled separately, the precision exception will always be set.
mov EMSEG:[Result],edi mov EMSEG:[RoundMode],offset PolyRound mov EMSEG:[ZeroVector],offset PolyZero push offset TransUnround ;Always exit through here mov ebx,EMSEG:[edi].lManHi mov esi,EMSEG:[edi].lManLo ;Check for small argument, so that x^2 does not underflow. Note that ;e^x = 1+x for small x, where small x means x + x^2/2 = x [or 1 + x/2 = 1], ;which happens when x < 2^-64, so 2^x - 1 = x * ln(2) for small x. TinyExpArg equ -64 cmp ecx,TinyExpArg shl 16 jl TinyExp cmp ecx,-1 shl 16 + bSign shl 8 ;See if positive, < 0.5 jl ExpReduced ;Argument was not in range (0, 0.5), so we need some kind of reduction or ecx,ecx ;Exp >= 0 means arg >= 1.0 --> too big ;CONSIDER: this returns through TransUnround which restores the rounding ;vectors, but it also randomly rounds the result becase eax is not set. jge ExpRet ;Give up if arg out of range ;We're going to need to add/subtract 1.0 or 0.5, so load up the constant mov edx,1 shl 31 xor edi,edi mov eax,-1 shl 16 + bSign shl 8 ;edx:edi,eax = -0.5 mov ebp,offset ExpReducedMinusHalf or ch,ch ;If it's positive, must be [0.5, 1) jns ExpReduction xor ah,ah ;edx:edi,eax = +0.5 mov ebp,offset ExpReducedPlusHalf cmp ecx,eax ;See if abs(arg) >= 0.5 jl ExpReduction ;No, adjust by .5 xor eax,eax ;edx:edi,eax = 1.0 mov ebp,offset ExpReducedPlusOne ExpReduction: call AddDoubleReg ;Argument now in range [0, 0.5] cmp cl,bTAG_ZERO ;Did reduction result in zero? jz ExpHalf ;If so, must have been exactly 0.5 push ebp ;Address of reduction cleanup ExpReduced: mov edi,offset tExpPoly call Eval2Poly ;2^x - 1 is approximated with 2 * x*P(x^2) / ( Q(x^2) - x*P(x^2) ) ;Q(x^2) is in registers, P(x^2) is at [[CURstk]] mov edi,EMSEG:[CURstk] mov dx,bSign shl 8 ;Subtract memory operand ;Note that Q() and P() have no roots over the input range ;(they will never be zero). call AddDouble ;Q(x^2) - x*P(x^2) sub ecx,1 shl 16 ;Divide by two mov edi,EMSEG:[CURstk] jmp DivDouble ;2 * x*P(x^2) / ( Q(x^2) - x*P(x^2) ) ;Returns to correct argument reduction correction routine or TransUnround
TinyExp: ;Exponent is very small (and was not reduced) mov edx,cFLDLN2hi mov edi,cFLDLN2lo mov eax,cFLDLN2exp shl 16 ;This could underflow (but not big time) jmp MulDoubleReg ;Returns to TransUnround
ExpHalf: ;Argument of exactly 0.5 was reduced to zero. Just return result. mov ebx,Sqrt2m1Hi mov esi,Sqrt2m1Lo mov eax,XSqrt2m1Lo + 1 shl 31 - 1 mov ecx,Sqrt2m1Exp shl 16 ret ;Exit through TransUnround
ExpReducedPlusOne: ;Correct result is (P - 1)/2 sub ecx,1 shl 16 ;Divide by two mov edx,1 shl 31 xor edi,edi mov eax,-1 shl 16 + bSign shl 8 ;edx:edi,eax = -0.5 jmp AddDoubleReg
ExpReducedPlusHalf: ;Correct result is P * sqrt(2)/2 - (1 - sqrt(2)/2) mov edx,Sqrt2Hi mov edi,Sqrt2Lo mov eax,Sqrt2exp-1 shl 16 ;sqrt(2)/2 call MulDoubleReg mov edx,TwoMinusSqrt2Hi mov edi,TwoMinusSqrt2Lo mov eax,(TwoMinusSqrt2Exp-1) shl 16 + bSign shl 8 ;(2-sqrt(2))/2 jmp AddDoubleReg
ExpReducedMinusHalf: ;Correct result is P * sqrt(2) + (sqrt(2)-1) mov edx,Sqrt2Hi mov edi,Sqrt2Lo mov eax,Sqrt2exp shl 16 call MulDoubleReg mov edx,Sqrt2m1Hi mov edi,Sqrt2m1Lo mov eax,Sqrt2m1Exp shl 16 jmp AddDoubleReg
;*******************************************************************************
;Dispatch table for log(x+1) ; ;One operand has been loaded into ecx:ebx:esi ("source"), the other is ;pointed to by edi ("dest"). ; ;Tag of source is shifted. Tag values are as follows:
.erre TAG_SNGL eq 0 ;SINGLE: low 32 bits are zero .erre TAG_VALID eq 1 .erre TAG_ZERO eq 2 .erre TAG_SPCL eq 3 ;NAN, Infinity, Denormal, Empty
;Any special case routines not found in this file are in emarith.asm
tFyl2xp1Disp label dword ;Source (ST(0)) Dest (*[di] = ST(1)) dd LogP1Double ;single single dd LogP1Double ;single double dd LogP1ZeroDest ;single zero dd LogP1SpclDest ;single special dd LogP1Double ;double single dd LogP1Double ;double double dd LogP1ZeroDest ;double zero dd LogP1SpclDest ;double special dd XorSourceSign ;zero single dd XorSourceSign ;zero double dd XorDestSign ;zero zero dd LogP1SpclDest ;zero special dd LogSpclSource ;special single dd LogSpclSource ;special double dd LogSpclSource ;special zero dd TwoOpBothSpcl ;special special dd LogTwoInf ;Two infinites
LogP1Double: ;st(0) mantissa in ebx:esi, exponent in high ecx, sign in ch bit 7 ;[edi] points to st(1), where result is returned ; ;This instruction is defined only for x+1 in the range [1/sqrt(2), sqrt(2)] ;The approximation used (valid over exactly this range) is ; log2(x) = z * P(z^2) / Q(z^2), z = (x-1) / (x+1), which is ; log2(x+1) = r * P(r^2) / Q(r^2), r = x / (x+2) ; ;We're not too picky about this range check because the function is simply ;"undefined" if out of range--EXCEPT, we're supposed to check for -1 and ;signal Invalid if less, -infinity if equal. or ecx,ecx ;abs(x) >= 1.0? jge LogP1OutOfRange ;Valid range is approx [-0.3, +0.4] mov EMSEG:[Result],edi mov EMSEG:[RoundMode],offset PolyRound mov EMSEG:[ZeroVector],offset PolyZero mov eax,1 shl 16 ;Exponent of 1 for adding 2.0 push offset TotalLog ;Return address for BasicLog ; jmp BasicLog ;Fall into BasicLog ;.erre BasicLog eq $
;BasicLog is used by eFYL2X and eFYL2XP1. ;eax has exponent and sign to add 1.0 or 2.0 to argument ;ebx:esi,ecx has argument, non-zero, tag not set ;ST has argument to take log2 of, minus 1. (This is the actual argument ;of eFYL2XP1, or argument minus 1 of eFYL2X.)
BasicLog: mov edx,1 shl 31 xor edi,edi ;edx:edi,eax = +1.0 or +2.0 call AddDoubleReg mov edi,EMSEG:[CURstk] ;Point to x-1 call DivDouble ;Compute (x-1) / (x+1) ;Result in registers is z = (x-1)/(x+1). For tiny z, ln(x) = 2*z, so ; log2(x) = 2 * log2(e) * z. Tiny z is such that z + z^3/3 = z. cmp ecx,-32 shl 16 ;Smallest exponent to bother with jl LogSkipPoly mov edi,offset tLogPoly call Eval2Poly mov edi,EMSEG:[CURstk] ;Point to first result, r * P(r^2) jmp DivDouble ;Compute r * P(r^2) / Q(r^2)
LogSkipPoly: ;Multiply r by 2 * log2(e) mov edx,Log2OfEHi mov edi,Log2OfELo mov eax,(Log2OfEexp+1) shl 16 jmp MulDoubleReg
LogP1OutOfRange: ;Input range isn't valid, so we can return anything we want--EXCEPT, for ;numbers < -1 we must signal Invalid Operation, and Divide By Zero for ;-1. Otherwise, we return an effective log of one by just leaving the ;second operand as the return value. ; ;Exponent in ecx >= 0 ( abs(x) >= 1 ) or ch,ch ;Is it positive? jns LogP1Ret ;If so, skip it and ecx,0FFFFH shl 16 ;Look at exponent only: 0 for -1.0 sub ebx,1 shl 31 ;Kill MSB or ebx,esi or ebx,ecx jnz ReturnIndefinite ;Must be < -1.0 jmp DivideByMinusZero
LogP1Ret: ret ;*** LogP1ZeroDest: or ch,ch ;Is it negative? jns LogP1Ret ;If not, just leave it zero or ecx,ecx ;abs(x) >= 1.0? jl XorDestSign ;Flip sign of zero ;Argument is <= -1 jmp ReturnIndefinite ;Have 0 * log( <=0 )
;*** LogP1SpclDest: mov al,EMSEG:[edi].bTag ;Pick up tag cmp al,bTAG_INF ;Is argument infinity? jnz SpclDest ;In emarith.asm ;Multiplying log(x+1) * infinity. ;If x > 0, return original infinity. ;If -1 <= x < 0, return infinity with sign flipped. ;If x < -1 or x == 0, invalid operation. cmp cl,bTAG_ZERO jz ReturnIndefinite or ch,ch ;Is it positive? jns LogP1Ret test ecx,0FFFFH shl 16 ;Is exponent zero? jl XorDestSign jg ReturnIndefinite sub ebx,1 shl 31 ;Kill MSB or ebx,esi jnz ReturnIndefinite ;Must be < -1.0 jmp XorDestSign
;*** LogSpclSource: cmp cl,bTAG_INF ;Is argument infinity? jnz SpclSource ;in emarith.asm or ch,ch ;Is it negative infinity? js ReturnIndefinite jmp MulByInf
;*** LogTwoInf: or ch,ch ;Is it negative infinity? js ReturnIndefinite jmp XorDestSign
;*******************************************************************************
;Dispatch table for log(x) ; ;One operand has been loaded into ecx:ebx:esi ("source"), the other is ;pointed to by edi ("dest"). ; ;Tag of source is shifted. Tag values are as follows:
.erre TAG_SNGL eq 0 ;SINGLE: low 32 bits are zero .erre TAG_VALID eq 1 .erre TAG_ZERO eq 2 .erre TAG_SPCL eq 3 ;NAN, Infinity, Denormal, Empty
;Any special case routines not found in this file are in emarith.asm
tFyl2xDisp label dword ;Source (ST(0)) Dest (*[di] = ST(1)) dd LogDouble ;single single dd LogDouble ;single double dd LogZeroDest ;single zero dd LogSpclDest ;single special dd LogDouble ;double single dd LogDouble ;double double dd LogZeroDest ;double zero dd LogSpclDest ;double special dd DivideByMinusZero ;zero single dd DivideByMinusZero ;zero double dd ReturnIndefinite ;zero zero dd LogSpclDest ;zero special dd LogSpclSource ;special single dd LogSpclSource ;special double dd LogSpclSource ;special zero dd TwoOpBothSpcl ;special special dd LogTwoInf ;Two infinites
LogDouble: ;st(0) mantissa in ebx:esi, exponent in high ecx, sign in ch bit 7 ;[edi] points to st(1), where result is returned ; ;Must reduce the argument to the range [1/sqrt(2), sqrt(2)] or ch,ch ;Is it positive? js ReturnIndefinite ;Can't take log of negative number mov EMSEG:[Result],edi mov EMSEG:[RoundMode],offset PolyRound mov EMSEG:[ZeroVector],offset PolyZero shld eax,ecx,16 ;Save exponent in ax as int part of log2 xor ecx,ecx ;Zero exponent: 1 <= x < 2 cmp ebx,Sqrt2Hi ;x > sqrt(2)? jb LogReduced ja LogReduceOne cmp esi,Sqrt2Lo jb LogReduced LogReduceOne: sub ecx,1 shl 16 ;1/sqrt(2) < x < 1 inc eax LogReduced: push eax ;Save integer part of log2 mov ebp,ecx ;Save reduced exponent (tag is wrong!) mov edx,1 shl 31 mov eax,bSign shl 8 ;Exponent of 0, negaitve xor edi,edi ;edx:edi,eax = -1.0 call AddDoubleReg cmp cl,bTAG_ZERO ;Was it exact power of two? jz LogDone ;Skip log if power of two ;Save (x - 1), reload x with reduced exponent mov edi,EMSEG:[CURstk] ;Point to original x again xchg EMSEG:[edi].lManHi,ebx xchg EMSEG:[edi].lManLo,esi mov EMSEG:[edi].ExpSgn,ecx mov ecx,ebp ;Get reduced exponent xor eax,eax ;Exponent of 0, positive call BasicLog LogDone: pop eax ;Get integer part back cwde or eax,eax ;Is it zero? jz TotalLog ;Next 3 instructions take abs() of integer cdq ;Extend sign through edx xor eax,edx ;Complement... sub eax,edx ; and increment if negative bsr dx,ax ;Look for MSB to normalize integer ;Bit number in dx ranges from 0 to 15 mov cl,dl not cl ;Convert to shift count shl eax,cl ;Normalize .erre TexpBias eq 0 rol edx,16 ;Move exponent high, sign low or ebx,ebx ;Was log zero? jz ExactPower xchg edx,eax ;Exp/sign to eax, mantissa to edx xor edi,edi ;Extend with zero call AddDoubleReg TotalLog: ;Registers could be zero if input was exactly 1.0 cmp cl,bTAG_ZERO jz ZeroLog TotalLogNotZero: mov edi,EMSEG:[Result] ;Point to second arg push offset TransUnround jmp MulDouble
ExactPower: ;Arg was a power of two, so log is exact (but not zero). mov ebx,eax ;Mantissa to ebx mov ecx,edx ;Exponent to ecx xor esi,esi ;Extend with zero ;Exponent of arg [= log2(arg)] is now normalized in ebx:esi,ecx ; ;The result log is exact, so we don't want TransUnround, which is designed ;to ensure the result is never exact. Instead we set the [RoundMode] ;vector to [TransRound] before the final multiply. mov eax,EMSEG:[TransRound] mov EMSEG:[RoundMode],eax mov edi,EMSEG:[Result] ;Point to second arg push offset RestoreRound ;Return addr. for MulDouble in emtrig.asm jmp MulDouble
ZeroLog: mov eax,EMSEG:[SavedRoundMode] mov EMSEG:[RoundMode],eax mov EMSEG:[ZeroVector],offset SaveResult jmp SaveResult
;*** LogZeroDest: or ch,ch ;Is it negative? js ReturnIndefinite ;Can't take log of negative numbers ;See if log is + or - so we can get correct sign of zero or ecx,ecx ;Is exponent >= 0? jge LogRet ;If so, keep present zero sign FlipDestSign: not EMSEG:[edi].bSgn ret
;*** LogSpclDest: mov al,EMSEG:[edi].bTag ;Pick up tag cmp al,bTAG_INF ;Is argument infinity? jnz SpclDest ;In emarith.asm ;Multiplying log(x) * infinity. ;If x > 1, return original infinity. ;If 0 <= x < 1, return infinity with sign flipped. ;If x < 0 or x == 1, invalid operation. cmp cl,bTAG_ZERO jz FlipDestSign or ch,ch ;Is it positive? js ReturnIndefinite test ecx,0FFFFH shl 16 ;Is exponent zero? jg LogRet ;x > 1, just return infinity jl FlipDestSign sub ebx,1 shl 31 ;Kill MSB or ebx,esi jz ReturnIndefinite ;x == 1.0 LogRet: ret
|