Windows2000/private/ntos/dll/i386/emtrig.asm
2020-09-30 17:12:32 +02:00

864 lines
30 KiB
NASM

subttl emtrig.asm - Trig functions sine, cosine, tangent
page
; Copyright (c) Microsoft Corporation 1991
; All Rights Reserved
;emtrig.asm - Trig functions sine, cosine, tangent
; by Tim Paterson
;Purpose:
; FCOS, FPTAN, FSIN, FSINCOS instructions
;Inputs:
; edi = [CURstk]
;Revision History:
; [] 09/05/91 TP Initial 32-bit version.
;XPi is the 66-bit value of Pi from the Intel manual
XPiHi equ 0C90FDAA2H
XPiMid equ 02168C234H
XPiLo equ 0C0000000H ;Extension of pi
PiOver4exp equ -1 ;Pi/4 ~= 3/4, so exponent is -1
TinyAngleExp equ -32 ;Smallest angle we bother with
MaxAngleExp equ 63 ;Angle that's too big
Trig1Result:
;Trig function reduction routine used by functions returning 1 value
;(FSIN and FCOS)
;edi = [CURstk] = argument pointer
;Argument has already been checked for zero.
;ZF = (tag == bTAG_ZERO)
jb TrigPrem
;Tagged special
mov al,EMSEG:[edi].bTAG
cmp al,bTAG_DEN
jz TrigDenorm
add sp,4 ;Don't return to caller
cmp al,bTAG_INF
jnz SpclDestNotDen ;Check for Empty or NAN
mov EMSEG:[SWcc],C2 ;Can't reduce infinity
jmp ReturnIndefinite
TrigDenorm:
mov EMSEG:[CURerr],Denormal
test EMSEG:[CWmask],Denormal ;Is denormal exception masked?
jnz TrigPrem ;Yes, continue
add sp,4 ;Don't return to caller
TrigRet:
ret
Trig2Inf:
mov EMSEG:[SWcc],C2 ;Can't reduce infinity
jmp Trig2Indefinite
Trig2StackOver:
mov EMSEG:[SWcc],C1 ;Signal overflow
Trig2StackUnder:
mov EMSEG:[CURerr],Invalid+StackFlag
Trig2Indefinite:
add sp,4 ;Don't return to caller
call ReturnIndefinite
jz TrigRet ;Unmasked, don't change registers
;Produce masked response
mov EMSEG:[CURstk],esi ;Push stack
mov edi,esi
jmp ReturnIndefinite
Trig2Special:
cmp al,bTAG_DEN
jz TrigDenorm
cmp al,bTAG_INF
jz Trig2Inf
;Must be a NAN
add sp,4 ;Don't return to caller
call DestNAN
jz TrigRet ;Unmasked, don't change registers
;Produce masked response
mov EMSEG:[CURstk],esi ;Push stack
mov eax,EMSEG:[edi].ExpSgn
mov EMSEG:[esi].ExpSgn,eax
mov eax,EMSEG:[edi].lManHi
mov EMSEG:[esi].lManHi,eax
mov eax,EMSEG:[edi].lManLo
mov EMSEG:[esi].lManLo,eax
ret
Trig2Zero:
add sp,4 ;Don't return to caller
mov EMSEG:[CURstk],esi
mov edi,esi
;Amazing coincidence: both FSINCOS and FPTAN return the same result for
;a zero argument:
; FSINCOS returns ST(0) = cos(0) = 1, ST(1) = sin(0) = 0.
; FPTAN returns ST(0) = 1 always, ST(1) = tan(0) = 0.
;Return zero has same sign as argument zero, so we don't need to touch
;it -- just push +1.0.
jmp ReturnOne
TrigOutOfRange:
mov EMSEG:[SWcc],C2 ;Signal argument not reduced
add sp,4
ret
PrevStackWrap esi,Trig2 ;Tied to PrevStackElem below
Trig2Result:
;Trig function reduction routine used by functions returning 2 values
;(FSINCOS and FPTAN)
;edi = [CURstk] = argument pointer
mov esi,edi
PrevStackElem esi,Trig2 ;esi points to second result location
mov al,EMSEG:[edi].bTAG ;Get tag
cmp al,bTAG_EMPTY ;Stack underflow if empty
jz Trig2StackUnder
cmp EMSEG:[esi].bTAG,bTAG_EMPTY ;Stack overflow if not empty
jnz Trig2StackOver
cmp al,bTAG_ZERO ;Is it Special?
ja Trig2Special
jz Trig2Zero
;Fall into TrigPrem
;TrigPrem
;This routine reduces an angle in radians to the range [0, pi/4].
;Angles in odd-numbered octants have been subtracted from pi/4.
;It uses a 66-bit value for pi, as required by the 387.
;TrigPrem uses the same two-stage algorithm as FPREM (see
;emfprem.asm). However, it is limited to an argument < 2^63.
;Inputs:
; edi = [CURstk]
;Outputs:
; ebx:esi = remainder, normalized
; high ecx = exponent, cl = tag
; al = octant
; edi = [CURstk]
TrigPrem:
mov EMSEG:[Result],edi
mov eax,EMSEG:[edi].lManLo
mov edx,EMSEG:[edi].lManHi
movsx ebx,EMSEG:[edi].wExp
cmp ebx,MaxAngleExp
jge TrigOutOfRange
xor edi,edi ;Extend dividend
xor esi,esi ;Quotient, in case we skip stage 1
.erre PiOver4exp eq -1
inc ebx ;Subtract exponent of pi/4
jl ExitTrigPrem ;If dividend is smaller, return it.
;We now know that 0 <= ExpDif < 64, so it fits in bl.
cmp bl,31 ;Do we need to do stage 1?
jl FitPi ;No, start stage 2
;FPREM stage 1
;Exponent difference is at least 31. Use 32-bit division to compute
;quotient and exact remainder, reducing exponent difference by 31.
;edx:eax = dividend
;ebx = exponent difference
;Shift dividend right one bit to be sure DIV instruction won't overflow
;This means we'll be reducing the exponent difference by 31, not 32
xor ebp,ebp ;Dividend extension
shrd ebp,eax,1
shrd eax,edx,1
shr edx,1
sub bl,31 ;Exponent reduced
mov ecx,XPiHi
div ecx ;Guess a quotient "digit"
;Check out our guess.
;Currently, remainder in edx = (high dividend) - (quotient * high pi).
;(High dividend is the upper 64 bits--ebp has 1 bit.) The definition
;of remainder is (all dividend) - (quotient * all pi). So if we
;subtract (quotient * low pi) from edx:ebp, we'll get the true
;remainder. If it's negative, our guess was too big.
mov esi,eax ;Save quotient
mov ecx,edx ;Save remainder
;The pi/4 we use has two bits set below the first 64 bits. This means
;we must add another 3/4 of the quotient into the amount to subtract,
;which we'll compute by rounding the low 32 bits up 1, then subtracting
;1/4 of quotient. But since we're computing the amount to subtract from
;the remainder, we'll add the 1/4 of the quotient to the remainder instead
;of subtracting it from the amount to subtract.
.erre XPiLo eq (3 shl 30)
mov eax,XPiMid+1
mul esi ;Quotient * low pi
;Note that ebp is either 0 or 800...00H
shr ebp,30 ;Move down to low end
shld ebp,esi,30 ;Move back up, adding 1/4 of quotient
mov edi,esi ;Another copy of quotient
shl edi,30 ;Keep last two bits
;edx:eax has amount to subtract to get correct remainder from ecx:ebp:edi
sub ebp,eax
sbb ecx,edx ;Subtract from remainder
mov eax,ebp
mov edx,ecx ;Remainder back to edx:eax:edi
jnc TrigPremNorm ;Was quotient OK?
TrigCorrect:
dec esi ;Quotient was too big
add edi,XPiLo
adc eax,XPiMid ;Add divisor back into remainder
adc edx,XPiHi
jnc TrigCorrect ;Repeat if quotient is still too big
jmp TrigPremNorm
;FPREM stage 2
;Exponent difference is less than 32. Use restoring long division to
;compute quotient bits until exponent difference is zero. Note that we
;often get more than one bit/loop: BSR is used to scan off leading
;zeros each time around. Since the divisor is normalized, we can
;instantly compute a zero quotient bit for each leading zero bit.
TrigPremLoop:
;edx:eax:edi = dividend (remainder) minus pi/4
;esi = quotient
;ebx = exponent difference
;If D is current dividend and p is pi/4, then we have edx:eax:edi = D - p,
;which is negative. We want 2*D - p, which is positive.
;2*D - p = 2*(D - p) + p.
add edi,edi ;2*(D - p)
adc eax,eax
adc edx,edx
add edi,XPiLo ;2*(D-p) + p = 2*D - p
adc eax,XPiMid
adc edx,XPiHi
add esi,esi ;Double quotient too
dec ebx ;Decrement exponent difference
PiFit:
inc esi
TrigPremNorm:
bsr ecx,edx ;Find first 1 bit
jz TrigPremZero
not cl
and cl,1FH ;Convert bit no. to shift count
sub ebx,ecx ;Reduce exponent difference
jl TrigTooFar
shld edx,eax,cl
shld eax,edi,cl
shl edi,cl ;Finish normalize shift
shl esi,cl ;Shift quotient
FitPi:
;Dividend could be larger or smaller than divisor
sub edi,XPiLo
sbb eax,XPiMid
sbb edx,XPiHi
jnc PiFit
;Couldn't subtract pi/2 from dividend.
;edx:eax:edi = dividend - pi/4, which is negative
or ebx,ebx ;Is exponent difference zero?
jg TrigPremLoop
;If quotient (octant number) is odd, we have subtracted an odd number of
;pi/4's. However, simple angle reductions work in multiples of pi/2.
;We will keep the extra pi/4 we just subtracted if the octant was odd.
;This will give a result range of [-pi/4, pi/4].
test esi,1 ;Is octant odd?
jz EvenOctant
NegPremResult:
;-pi/4 < dividend < 0. Negate this since we use sign-magnitude representation.
not edx ;96-bit negation
not eax
neg edi
sbb eax,-1
sbb edx,-1
;May need to normalize
bsr ecx,edx
jz TrigNorm32
lea ebx,[ebx+ecx-31] ;Fix up exponent for normalization
not cl ;Convert bit no. to shift count
TrigShortNorm:
shld edx,eax,cl
shld eax,edi,cl
shl edi,cl ;Finish normalize shift
RoundPrem:
;Must round 66-bit result to 64 bits.
;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.
bt eax,0 ;Is mantissa even or odd? (set CY)
adc edi,(1 shl 31)-1 ;Sum LSB & sticky bits--CY if round up
adc eax,0
adc edx,0
ExitTrigPrem:
;edx:eax = remainder, normalized
;esi = quotient
;ebx = exponent difference, zero or less
.erre PiOver4exp eq -1
dec ebx ;True exponent
.erre bTAG_SNGL eq 0
shrd ecx,ebx,16 ;Exponent to high ecx
mov ebx,edx ;High mant. to ebx
xchg esi,eax ;Low mant. to esi, octant to eax
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
mov edi,EMSEG:[CURstk]
test EMSEG:[edi].bSgn,bSign ;Was angle negative?
jnz FlipOct ;Yes, flip octant over
ret
FlipOct:
;Angle was negative. Subtract octant from 7.
neg al
add al,7
ret
EvenOctant:
;Restore dividend
add edi,XPiLo
adc eax,XPiMid
adc edx,XPiHi
jmp RoundPrem
TrigTooFar:
;Exponent difference in ebx went negative when reduced by shift count in ecx.
;We need a quotient corresponding to exponent difference of zero.
add ecx,ebx ;Compute previous exponent difference
shl esi,cl ;Fix up quotient
sub ecx,ebx ;Restore shift count
test esi,1 ;Is octant odd?
jz TrigShortNorm ;No, go normalize
xor ebx,ebx ;Restore old exponent difference (zero)
SubPiOver4:
;We are here if exponent difference was zero and octant is odd.
;As noted above, we need to reduce the angle by a multiple of pi/2,
;not pi/4. We will subtract one more pi/4, which will make the
;result range [-pi/4, pi/4].
sub edi,XPiLo
sbb eax,XPiMid
sbb edx,XPiHi
jmp NegPremResult
TrigPremZero:
;High dword of remainder is all zero, so we've reduced exponent difference
;by 32 bits and overshot. We need a quotient corresponding to exponent
;difference of zero, so we just shift it by the original difference. Then
;we need to normalize the rest of the remainder.
mov ecx,ebx ;Get exponent difference
shl esi,cl ;Fix up quotient
test esi,1 ;Is octant odd?
jnz SubPiOver4 ;Yes, go subtract another pi/4
TrigNorm32:
bsr ecx,eax
jz TinyTrig
lea ebx,[ebx+ecx-31-32] ;Fix up exponent for normalization
mov edx,eax
mov eax,edi ;Shift left by 32 bits
not cl ;Convert bit no. to shift count
shld edx,eax,cl ;Normalize remainder
shl eax,cl
jmp ExitTrigPrem
TinyTrig:
;Upper 64 bits of remainder are all zero. We are assured that the extended
;remainder is never zero, though.
mov edx,edi ;Shift left 64 bits
bsr ecx,edi
lea ebx,[ebx+ecx-31-64] ;Fix up exponent for normalization
not cl ;Convert bit no. to shift count
shl edx,cl ;Normalize
jmp ExitTrigPrem
EM_ENTRY eFCOS
eFCOS:
and [esp].[OldLongStatus+4],NOT(C2 SHL 16) ;clear C2
cmp EMSEG:[edi].bTAG,bTAG_ZERO
jz ReturnOne
call Trig1Result
;ebx:esi,ecx = reduced argument
;eax = octant
mov ch,80H ;Assume negative
test al,110B ;Negative in octants 2 - 5
jpo @F ;Occurs when 1 of these bits are set
xor ch,ch ;Actually positve
@@:
test al,011B ;Look for octants 0,3,4,7
jpo TakeSine ;Use sine if not
TakeCosine:
cmp ecx,TinyAngleExp shl 16 ;Is angle really small?
jl CosReturnOne ;cos(x) = 1 for tiny x
CosNotTiny:
mov edi,offset tCosPoly
;Note that argument needs to be saved in ArgTemp (by EvalPolySetup) in case
;we were called from eFSINCOS and we'll need the arg for the sine. Argument
;is not needed for cosine, however (just its square).
call EvalPolySetup ;In emftran.asm
mov ch,EMSEG:[ArgTemp].bSgn ;Get sign we already figured out
TransUnround:
;The last operation performed a simple round nearest, without setting the
;C1 status bit if round up occured. We reverse this last rounding now
;so we can do the user's selected rounding mode. We also ensure that
;the answer is never exact.
sub eax,(1 shl 31)-1 ;Sum LSB & sticky bits--CY if round up
jz UnroundExact ;Answer looks exact, but it's not
sbb esi,0
sbb ebx,0
jns PolyDropExponent ;We had rounded up exponent too
FinalTransRound:
;A jump through [TransRound] is only valid if the number is known not to
;underflow. Unmasked underflow requires [RoundMode] be set.
mov edx,EMSEG:[TransRound]
mov EMSEG:[RoundMode],edx
call edx ;Perform user's rounding
RestoreRound:
;Restore rounding vectors
mov EMSEG:[ZeroVector],offset SaveResult
mov eax,EMSEG:[SavedRoundMode]
mov EMSEG:[RoundMode],eax
ret
UnroundExact:
inc eax ;Let's say our answer is a bit small
jmp FinalTransRound
PolyDropExponent:
sub ecx,1 shl 16 ;Decrement exponent
or ebx,1 shl 31 ;Set MSB
jmp FinalTransRound
SinRet:
ret
SaveTinySin:
;Argument in ebx:esi,ecx is small enough so that sin(x) = x, which happens
;when x - x^3/6 = x [or 1 - x^2/6 = 1]. Note that the infinitely precise
;result is slightly less than the argument. To get the correct answer for
;any rounding mode, we decrement the argument and set up for rounding.
mov eax,-1 ;Set up rounding bits
sub esi,1
sbb ebx,0 ;Drop mantissa by one
js FinalTransRound ;Still normalized?
;mantissa must have been 800..000H, set it to 0FFF...FFFH and drop exponent
mov ebx,eax ;ebx = -1
sub ecx,1 shl 16 ;Drop exponent by one
jmp FinalTransRound
EM_ENTRY eFSIN
eFSIN:
and [esp].[OldLongStatus+4],NOT(C2 SHL 16) ;clear C2
cmp EMSEG:[edi].bTAG,bTAG_ZERO
jz SinRet ;Return zero for zero argument
call Trig1Result
mov ch,al
shl ch,7-2 ;Move bit 2 to bit 7 as sign bit
ReducedSine:
;ebx:esi,ecx = reduced argument
;ch = correct sign
;eax = octant
test al,011B ;Look for octants 0,3,4,7
jpo TakeCosine ;Use cosine if not
TakeSine:
cmp ecx,TinyAngleExp shl 16 ;Is angle really small?
jl SaveTinySin ;sin(x) = x for tiny x
;The polynomial for sine is sin(x) = x * P(x^2). However, the degree zero
;coefficient of P() is 1, so P() = R() + 1, where R() has no degree zero
;term. Thus sin(x) = x * [R(x^2) + 1] = x * R(x^2) + x.
;What's important here is that adding 1 to R(x^2) can blow away a lot of
;precision just before we do that last multiply by x. Note that x < pi/4 < 1,
;so that x^2 is often << 1. The precision is lost when R(x^2) is shifted
;right to align its binary point with 1.0. This can cause a loss of at
;least 1 bit of precision after the final multiply by x in addition to
;rounding errors.
;To avoid this precision loss, we use the alternate form given above,
;sin(x) = x * R(x^2) + x. Instead of adding 1.0 and multiplying by x,
;we multiply by x and add x--exactly the same level of difficulty. But
;the mulitply has all of R(x^2)'s precision available.
;Because the polynomial R() has no zero-degree term, we give EvalPoly
;one degree less (so we don't have to add zero as the last term).
;Then we have to multiply once more by x^2 since we left the loop early.
SineNotTiny:
mov edi,offset tSinPoly
call EvalPolySetup ;In emftran.asm
SineFinish:
ifdef NT386
mov edi,YFloatTemp
else
mov edi,offset edata:FloatTemp
endif
call PolyMulDouble ;Last coefficient in R(x^2)
ifdef NT386
mov edi,YArgTemp ;Point to original x
else
mov edi,offset edata:ArgTemp ;Point to original x
endif
call PolyMulDouble ;Compute x * R(x^2)
ifdef NT386
mov edi,YArgTemp ;Point to original x
else
mov edi,offset edata:ArgTemp ;Point to original x
endif
push offset TransUnround
jmp PolyAddDouble ;Compute x * R(x^2) + x
EM_ENTRY eFPTAN
eFPTAN:
and [esp].[OldLongStatus+4],NOT(C2 SHL 16) ;clear C2
call Trig2Result
push offset TanPushOne ; Push 1.0 when we're all done
;ebx:esi,ecx = reduced argument
;eax = octant
mov ch,al
shl ch,7-1 ;Move bit 1 to bit 7 as sign bit
;Note that ch bit 6 now has even/odd octant, which we'll need when we're
;done to see if we should take reciprocal.
cmp ecx,TinyAngleExp shl 16 ;Is angle really small?
jl TinyTan
mov edi,offset tTanPoly
call Eval2Poly ;In emftran.asm
mov edi,EMSEG:[CURstk] ;Point to first result
push offset TransUnround ;Return address of divide
test EMSEG:[ArgTemp].bSgn,0C0H ;Check low 2 bits of octant
;Given the reduced input range, the result can never overflow or underflow.
;It is must then be safe to assume neither operand is zero.
jpe DivDouble ;Tan() octants 0,3,4,7
jmp DivrDouble ;CoTan()
TinyTan:
test ch,0C0H ;Check low 2 bits of octant
jpe SaveTinySin ;Octants 0,3,4,7: tan(x) = x for tiny x
;Need reciprocal of reduced argument
mov edi,esi
mov esi,ebx ;Mantissa in esi:edi
mov ebx,ecx ;ExpSgn to ebx
mov edx,1 shl 31 ;Load 1.0
xor eax,eax
.erre TexpBias eq 0
xor ecx,ecx ;Sign and exponent are zero
;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
push offset TransUnround ;Return address of divide
;Note that this can never overflow, because the reduced argument is never
;smaller than about 2^-65.
jmp DivDoubleReg
PrevStackWrap edi,Tan ;Tied to PrevStackElem below
TanPushOne:
PrevStackElem edi,Tan ;edi points to second result location
mov EMSEG:[CURstk],edi
ReturnOne:
mov EMSEG:[edi].lManLo,0
mov EMSEG:[edi].lManHi,1 shl 31
mov EMSEG:[edi].ExpSgn,(0-TexpBias) shl 16 + bTAG_SNGL
ret
PrevStackWrap edi,SinCos ;Tied to PrevStackElem below
eFSINCOS:
and [esp].[OldLongStatus+4],NOT(C2 SHL 16) ;clear C2
call Trig2Result
;Figure out signs
mov ch,al ;Start with sign of sine
shl ch,7-2 ;Move bit 2 to bit 7 as sign bit
mov ah,80H ;Assume sign of cosine is negative
test al,110B ;Negative in octants 2 - 5
jpo @F ;Occurs when 1 of these bits are set
xor ah,ah ;Actually positve
@@:
;ch = sign of sine
;ah = sign of cosine
cmp ecx,TinyAngleExp shl 16 ;Is angle really small?
jl TinySinCos
push eax ;Save octant and sign of cosine
call ReducedSine ;On exit, edi = [CURstk]
pop eax
;The Sin() funcion restored the rounding vectors to normal. Set them back.
mov EMSEG:[RoundMode],offset PolyRound
mov EMSEG:[ZeroVector],offset PolyZero
PrevStackElem edi,SinCos ;edi points to second result location
mov EMSEG:[CURstk],edi
mov EMSEG:[Result],edi
;Load x^2 back into registers
mov ecx,EMSEG:[FloatTemp].ExpSgn
mov ebx,EMSEG:[FloatTemp].lManHi
mov esi,EMSEG:[FloatTemp].lManLo
mov EMSEG:[ArgTemp].bSgn,ah ;Save sign
test al,011B ;Look for octants 0,3,4,7
jpo FastSine ;Use sine if not
mov edi,offset tCosPoly
call EvalPoly ;In emftran.asm
mov ch,EMSEG:[ArgTemp].bSgn ;Get sign we already figured out
jmp TransUnround
FastSine:
mov edi,offset tSinPoly
push offset SineFinish
jmp EvalPoly ;In emftran.asm
TinySinCos:
;ch = sign of sine
;ah = sign of cosine
;ebx:esi,high ecx = reduced argument
;edi = [CURstk]
test al,011B ;Look for octants 0,3,4,7
jpo TinyCosSin ;Take cosine first if not
push eax
call SaveTinySin ;For sine, arg is result
pop ecx
;edi = [CURstk]
;ch = sign of cosine
;Set cosine to 1.0
PrevStackElem edi,TinySinCos ;edi points to second result location
mov EMSEG:[CURstk],edi
mov EMSEG:[Result],edi
CosReturnOne:
;Cosine is nearly equal to 1.0. Put in next smaller value and round it.
mov ebx,-1
mov esi,ebx ;Set mantissa to -1
mov eax,ebx ;Set up rounding bits
.erre TexpBias eq 0
and ecx,bSign shl 8 ;Keep only sign
sub ecx,1 shl 16 ;Exponent of -1
;A jump through [TransRound] is only valid if the number is known not to
;underflow. Unmasked underflow requires [RoundMode] be set.
jmp EMSEG:[TransRound]
PrevStackWrap edi,TinySinCos
PrevStackWrap edi,TinyCosSin
TinyCosSin:
;Sine is nearly 1.0, cosine is argument
;ch = sign of sine
;ah = sign of cosine
;ebx:esi,high ecx = reduced argument
;edi = [CURstk]
xchg ah,ch ;Cosine sign to ch, sine sign to ah
push edi ;Save place for sine
PrevStackElem edi,TinyCosSin ;edi points to second result location
mov EMSEG:[CURstk],edi
mov EMSEG:[Result],edi
push eax
call SaveTinySin ;For sine, arg is result
pop ecx
;ch = sign of sine
pop EMSEG:[Result] ;Set up location for sine
jmp CosReturnOne
;********************* 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.
tCosPoly label word
;These constants are derived from Hart #3824: cos(x) = P(x^2),
;accurate to 19.45 digits over interval [0, pi/4]. The original
;constants in Hart required that the argument x be divided by pi/4.
;These constants have been scaled so this is no longer required.
;Scaling is done by multiplying the constant by a power of 4/pi.
;The power is given in the table.
dd 7 ;Degree seven
; Original Hart constant power Scaled constant
;-0.38577 62037 2 E-12 14 -0.113521232057839395845871741043E-10
;Hex value: 0.C7B56AF786699CF1BD13FD290 HFFDC
dq 0C7B56AF786699CF2H
dw (bSign shl 8)+bTAG_VALID,0FFDCH-1
;+0.11500 49702 4263 E-9 12 +0.208755551456778828747793797596E-8
;Hex value: 0.8F74AA3CCE49E68D6F5444A18 HFFE4
dq 08F74AA3CCE49E68DH
dw bTAG_VALID,0FFE4H-1
;-0.24611 36382 63700 5 E-7 10 -0.275573128656960822243472872247E-6
;Hex value: 0.93F27B7F10CC8A1703EFC8A04 HFFEB
dq 093F27B7F10CC8A17H
dw (bSign shl 8)+bTAG_VALID,0FFEBH-1
;+0.35908 60445 88581 953 E-5 8 +0.248015872828994630247806807317E-4
;Hex value: 0.D00D00CD6BB3ECD17E10D5830 HFFF1
dq 0D00D00CD6BB3ECD1H
dw bTAG_VALID,0FFF1H-1
;-0.32599 18869 26687 55044 E-3 6 -0.138888888888589604343951947246E-2
;Hex value: 0.B60B60B609B165894CFE522AC HFFF7
dq 0B60B60B609B16589H
dw (bSign shl 8)+bTAG_VALID,0FFF7H-1
;+0.15854 34424 38154 10897 54 E-1 4 +0.416666666666664302573692446873E-1
;Hex value: 0.AAAAAAAAAAA99A1AF53042B08 HFFFC
dq 0AAAAAAAAAAA99A1BH
dw bTAG_VALID,0FFFCH-1
;-0.30842 51375 34042 45242 414 E0 2 -0.499999999999999992843582920899E0
;Hex value: 0.FFFFFFFFFFFFFEF7F98D3BFA8 HFFFF
dq 0FFFFFFFFFFFFFEF8H
dw (bSign shl 8)+bTAG_VALID,0FFFFH-1
;+0.99999 99999 99999 99996 415 E0 0 (no change)
;Hex value 0.FFFFFFFFFFFFFFFF56B402618 H0
dq 0FFFFFFFFFFFFFFFFH
dw bTAG_VALID,00H-1
tSinPoly label word
;These constants are derived from Hart #3044: sin(x) = x * P(x^2),
;accurate to 20.73 digits over interval [0, pi/4]. The original
;constants in Hart required that the argument x be divided by pi/4.
;These constants have been scaled so this is no longer required.
;Scaling is done by multiplying the constant by a power of 4/pi.
;The power is given in the table.
dd 7-1 ;Degree seven, but the last coefficient
;is 1.0 and is not listed here.
; Original Hart constant power Scaled constant
;-0.20225 31292 93 E-13 15 -0.757786788401271156262125540409E-12
;Hex value: 0.D54C4AF2B524F0F2D6411C90A HFFD8
dq 0D54C4AF2B524F0F3H
dw (bSign shl 8)+bTAG_VALID,0FFD8H-1
;+0.69481 52035 0522 E-11 13 +0.160583476232246065559545749398E-9
;Hex value: 0.B0903AF085DA66030F16E43BC HFFE0
dq 0B0903AF085DA6603H
dw bTAG_VALID,0FFE0H-1
;-0.17572 47417 61708 06 E-8 11 -0.250521047382673309542092418731E-7
;Hex value: 0.D73229320D2AF05971AC96FF4 HFFE7
dq 0D73229320D2AF059H
dw (bSign shl 8)+bTAG_VALID,0FFE7H-1
;+0.31336 16889 17325 348 E-6 9 +0.275573192133901687156480447942E-5
;Hex value: 0.B8EF1D2984D2FBA28A9CC9DEE HFFEE
dq 0B8EF1D2984D2FBA3H
dw bTAG_VALID,0FFEEH-1
;-0.36576 20418 21464 00052 9 E-4 7 -0.198412698412531058609618529749E-3
;Hex value: 0.D00D00D00C3FDDD7916E5CB28 HFFF4
dq 0D00D00D00C3FDDD8H
dw (bSign shl 8)+bTAG_VALID,0FFF4H-1
;+0.24903 94570 19271 62752 519 E-2 5 +0.83333333333333203341753387264E-2
;Hex value: 0.8888888888884C95D619A0343 HFFFA
dq 08888888888884C96H
dw bTAG_VALID,0FFFAH-1
;-0.80745 51218 82807 81520 2582 E-1 3 -0.166666666666666666281276062229E0
;Hex value: 0.AAAAAAAAAAAAAA8E3AD80EAB8 HFFFE
dq 0AAAAAAAAAAAAAA8EH
dw (bSign shl 8)+bTAG_VALID,0FFFEH-1
;+0.78539 81633 97448 30961 41845 E0 1 +0.99999999999999999999812025812E0
;Hex value: 0.FFFFFFFFFFFFFFFFF71F88110 H0
; dq 8000000000000000H ;This constant of 1.0 omitted here.
; dw bTAG_VALID,0 ; It is handled in code.
tTanPoly label word
;These constants are derived from Hart #4286: tan(x) = x * P(x^2) / Q(x^2),
;accurate to 19.94 digits over interval [0, pi/4]. The original
;constants in Hart required that the argument x be divided by pi/4.
;These constants have been scaled so this is no longer required.
;Scaling is done by multiplying the constant by the same power of 4/pi
;as the power of x the constant is used on. However, the highest
;degree coefficient of Q() is 1, and after scaling this way it would
;become (4/pi)^8. In order to keep this coefficient equal to one,
;we scale everything again by (pi/4)^8. This scaling is partially
;canceled by the original scaling by powers of 4/pi, and the net
;resulting power of pi/4 is given in the table.
dd 3 ;First poly is degree 3
; Original Hart constant power Scaled constant
;-.45649 31943 86656 31873 96113 7 E2 1 -35.8528916474714232910463077546
;Hex value: 0.8F695C6D93AF6F97B6E022AB3 H6
dq 08F695C6D93AF6F98H
dw (bSign shl 8)+bTAG_VALID,06H-1
;+.14189 85425 27617 78388 00394 831 E5 3 +6874.60229709782436592720603503
;Hex value: 0.D6D4D181240D0D08C88DF4AA6 HD
dq 0D6D4D181240D0D09H
dw bTAG_VALID,0DH-1
;-.89588 84400 67680 41087 29639 541 E6 5 -267733.884797157298951145495276
;Hex value: 0.82BABC504220C62B1D0722684 H13
dq 082BABC504220C62BH
dw (bSign shl 8)+bTAG_VALID,013H-1
;+.10888 60043 72816 87521 38857 983 E8 7 +2007248.9111748838841548144685
;Hex value: 0.F506874A160EB9C0994AADD6A H15
dq 0F506874A160EB9C1H
dw bTAG_VALID,015H-1
dd 4 ;Second poly is degree 4
;NOTE: Eval2Poly assumes the first coefficient is 1.0, so it is omitted
; Original Hart constant power Scaled constant
;-.10146 56190 25288 53387 54401 947 E4 2 -625.890950057027419879480354834
;Hex value: 0.9C790553635355A95241A5324 HA
dq 09C790553635355A9H
dw (bSign shl 8)+bTAG_VALID,0AH-1
;+.13538 27128 05119 09382 89294 872 E6 4 +51513.6992033752080924797647367
;Hex value: 0.C939B2FEFE0DC585E649870FE H10
dq 0C939B2FEFE0DC586H
dw bTAG_VALID,010H-1
;-.39913 09518 03516 51504 43427 94 E7 6 -936816.855188785264866481436899
;Hex value: 0.E4B70DAEDA6F89E5A7CE626FA H14
dq 0E4B70DAEDA6F89E6H
dw (bSign shl 8)+bTAG_VALID,014H-1
;+.13863 79666 35676 29165 33913 361 E8 8 +2007248.91117488388417770850458
;Hex value: 0.F506874A160EB9C0CCD8313BC H15
dq 0F506874A160EB9C1H
dw bTAG_VALID,015H-1