e_atanhf.s

glibc 2.9,最新版的C语言库函数

data8 0x3fd237c1841a502e    //log(1/frcpa(1+196/256))/2
data8 0x3fd24cfce6f80d9a    //log(1/frcpa(1+197/256))/2
data8 0x3fd25b2c55cd5762    //log(1/frcpa(1+198/256))/2
data8 0x3fd2707f4d5f7c41    //log(1/frcpa(1+199/256))/2
data8 0x3fd285e0842ca384    //log(1/frcpa(1+200/256))/2
data8 0x3fd294294708b773    //log(1/frcpa(1+201/256))/2
data8 0x3fd2a9a2670aff0c    //log(1/frcpa(1+202/256))/2
data8 0x3fd2b7fb2c8d1cc1    //log(1/frcpa(1+203/256))/2
data8 0x3fd2c65a6395f5f5    //log(1/frcpa(1+204/256))/2
data8 0x3fd2dbf557b0df43    //log(1/frcpa(1+205/256))/2
data8 0x3fd2ea64c3f97655    //log(1/frcpa(1+206/256))/2
data8 0x3fd3001823684d73    //log(1/frcpa(1+207/256))/2
data8 0x3fd30e97e9a8b5cd    //log(1/frcpa(1+208/256))/2
data8 0x3fd32463ebdd34ea    //log(1/frcpa(1+209/256))/2
data8 0x3fd332f4314ad796    //log(1/frcpa(1+210/256))/2
data8 0x3fd348d90e7464d0    //log(1/frcpa(1+211/256))/2
data8 0x3fd35779f8c43d6e    //log(1/frcpa(1+212/256))/2
data8 0x3fd36621961a6a99    //log(1/frcpa(1+213/256))/2
data8 0x3fd37c299f3c366a    //log(1/frcpa(1+214/256))/2
data8 0x3fd38ae2171976e7    //log(1/frcpa(1+215/256))/2
data8 0x3fd399a157a603e7    //log(1/frcpa(1+216/256))/2
data8 0x3fd3afccfe77b9d1    //log(1/frcpa(1+217/256))/2
data8 0x3fd3be9d503533b5    //log(1/frcpa(1+218/256))/2
data8 0x3fd3cd7480b4a8a3    //log(1/frcpa(1+219/256))/2
data8 0x3fd3e3c43918f76c    //log(1/frcpa(1+220/256))/2
data8 0x3fd3f2acb27ed6c7    //log(1/frcpa(1+221/256))/2
data8 0x3fd4019c2125ca93    //log(1/frcpa(1+222/256))/2
data8 0x3fd4181061389722    //log(1/frcpa(1+223/256))/2
data8 0x3fd42711518df545    //log(1/frcpa(1+224/256))/2
data8 0x3fd436194e12b6bf    //log(1/frcpa(1+225/256))/2
data8 0x3fd445285d68ea69    //log(1/frcpa(1+226/256))/2
data8 0x3fd45bcc464c893a    //log(1/frcpa(1+227/256))/2
data8 0x3fd46aed21f117fc    //log(1/frcpa(1+228/256))/2
data8 0x3fd47a1527e8a2d3    //log(1/frcpa(1+229/256))/2
data8 0x3fd489445efffccc    //log(1/frcpa(1+230/256))/2
data8 0x3fd4a018bcb69835    //log(1/frcpa(1+231/256))/2
data8 0x3fd4af5a0c9d65d7    //log(1/frcpa(1+232/256))/2
data8 0x3fd4bea2a5bdbe87    //log(1/frcpa(1+233/256))/2
data8 0x3fd4cdf28f10ac46    //log(1/frcpa(1+234/256))/2
data8 0x3fd4dd49cf994058    //log(1/frcpa(1+235/256))/2
data8 0x3fd4eca86e64a684    //log(1/frcpa(1+236/256))/2
data8 0x3fd503c43cd8eb68    //log(1/frcpa(1+237/256))/2
data8 0x3fd513356667fc57    //log(1/frcpa(1+238/256))/2
data8 0x3fd522ae0738a3d8    //log(1/frcpa(1+239/256))/2
data8 0x3fd5322e26867857    //log(1/frcpa(1+240/256))/2
data8 0x3fd541b5cb979809    //log(1/frcpa(1+241/256))/2
data8 0x3fd55144fdbcbd62    //log(1/frcpa(1+242/256))/2
data8 0x3fd560dbc45153c7    //log(1/frcpa(1+243/256))/2
data8 0x3fd5707a26bb8c66    //log(1/frcpa(1+244/256))/2
data8 0x3fd587f60ed5b900    //log(1/frcpa(1+245/256))/2
data8 0x3fd597a7977c8f31    //log(1/frcpa(1+246/256))/2
data8 0x3fd5a760d634bb8b    //log(1/frcpa(1+247/256))/2
data8 0x3fd5b721d295f10f    //log(1/frcpa(1+248/256))/2
data8 0x3fd5c6ea94431ef9    //log(1/frcpa(1+249/256))/2
data8 0x3fd5d6bb22ea86f6    //log(1/frcpa(1+250/256))/2
data8 0x3fd5e6938645d390    //log(1/frcpa(1+251/256))/2
data8 0x3fd5f673c61a2ed2    //log(1/frcpa(1+252/256))/2
data8 0x3fd6065bea385926    //log(1/frcpa(1+253/256))/2
data8 0x3fd6164bfa7cc06b    //log(1/frcpa(1+254/256))/2
data8 0x3fd62643fecf9743    //log(1/frcpa(1+255/256))/2
LOCAL_OBJECT_END(atanhf_data2)


.section .text
GLOBAL_LIBM_ENTRY(atanhf)

{ .mfi
      getf.exp      rArgSExpb = f8
      fclass.m      p9,p0 = f8, 0x0b        // is arg denormal ?
      mov           rExpbMask = 0x1ffff
}
{ .mfi
      addl          DataPtr = @ltoff(atanhf_data), gp
      fnma.s1       fOneMx = f8, f1, f1     // 1 - x
      mov           rBias = 0xffff
}
;;

{ .mfi
      nop.m         0
      fclass.m      p7,p0 = f8, 0xc7        // is arg NaN or +/-0 ?
      mov           rNearZeroBound = 0xffeb // 2^(-20)
}
{ .mfi
      ld8           DataPtr = [DataPtr]
      fma.s1        fOnePx = f8, f1, f1     // 1 + x
      nop.i         0
}
;;

{ .mfb
      nop.m         0
      fnorm.s1      fNormX = f8                     // Normalize x
(p9)  br.cond.spnt  ATANH_UNORM                     // Branch if x=unorm
}
;;

ATANH_COMMON:
// Return here if x=unorm and not denorm
{ .mfi
      ldfpd         fP3, fP2 = [DataPtr], 16
      fma.s1        fX2 = f8, f8, f0        // x^2
      nop.i         0
}
{ .mfb
      nop.m         0
(p7)  fma.s.s0      f8 =  f8,f1,f8          // NaN or +/-0
(p7)  br.ret.spnt   b0
}
;;

{ .mfi
      ldfpd         fP1, fHalf = [DataPtr], 16
      frcpa.s1      fRcpM, p9 = f1, fOneMx  // rcpm = frcpa(1 - x)
      nop.i         0
}
;;

{ .mfi
      getf.exp      rExpbm = fOneMx
      frcpa.s1      fRcpP, p0 = f1, fOnePx  // rcpp = frcpa(1 + x)
      // biased exponent
      and           rArgExpb = rArgSExpb, rExpbMask
}
;;

{ .mmi
      getf.exp      rExpbp = fOnePx
      // is |x| < 2^(-20) ?
      cmp.gt        p8,p0 = rNearZeroBound, rArgExpb
      cmp.ge        p6,p0 = rArgExpb, rBias // is |x| >= 1 ?
}
;;

{ .mmb
      getf.sig      rSigm = fOneMx
      nop.m         0
(p6)  br.cond.spnt  atanhf_ge_one
}
;;

{ .mfb
      getf.sig      rSigp = fOnePx
(p8)  fma.s.s0      f8 =  fX2, f8, f8  // x + x^3
(p8)  br.ret.spnt   b0                 // Exit for MAX_DENORM_ABS < |x| < 2^-20
}
;;

{ .mfi
      ldfd          fLog2 = [DataPtr], 16
      fms.s1        fRm = fRcpM, fOneMx, f1 // rm = rcpm * (1 - x) - 1
      nop.i         0
}
;;

{ .mmf
      // (1 - x) is always positive here and we need not mask sign bit
      sub           rNm = rExpbm, rBias
      // (1 + x) is always positive here and we need not mask sign bit
      sub           rNp = rExpbp, rBias
      fms.s1        fRp = fRcpP, fOnePx, f1 // rp = rcpp * (1 + x) - 1
}
;;

{ .mmi
      setf.sig      fN4CvtM = rNm
      setf.sig      fN4CvtP = rNp
      extr.u        rIndm = rSigm,55,8                // Extract 8 bits
}
;;

{ .mmi
      shladd        RcpTablePtrM = rIndm, 3, DataPtr
      nop.m         0
      extr.u        rIndp = rSigp,55,8                // Extract 8 bits
}
;;

{ .mmi
      ldfd          fLogTm = [RcpTablePtrM]
      shladd        RcpTablePtrP = rIndp, 3, DataPtr
      nop.i         0
}
;;

{ .mfi
      ldfd          fLogTp = [RcpTablePtrP]
      fma.s1        fRm2 = fRm, fRm, f0     // rm^2
      nop.i         0
}
{ .mfi
      nop.m         0
      fma.s1        fP32m = fP3, fRm, fP2   // P3*rm + P2
      nop.i         0
}
;;

{ .mfi
      nop.m         0
      fma.s1        fRp2 = fRp, fRp, f0     // rp^2
      nop.i         0
}
{ .mfi
      nop.m         0
      fma.s1        fP10m = fP1, fRm, fHalf   // P1*rm + 1
      nop.i         0
}
;;

{ .mfi
      nop.m         0
      fma.s1        fP32p = fP3, fRp, fP2   // P3*rp + P2
      nop.i         0
}
{ .mfi
      nop.m         0
      fma.s1        fP10p = fP1, fRp, fHalf   // P1*rp + 1
      nop.i         0
}
;;

{ .mfi
      nop.m         0
      fcvt.xf       fNm = fN4CvtM
      nop.i         0
}
{ .mfi
      nop.m         0
      fcvt.xf       fNp = fN4CvtP
      nop.i         0
}
;;

{ .mfi
      nop.m         0
      // (P3*rm + P2)*rm^2 + (P1*rm + 1)
      fma.s1        fP32m = fP32m, fRm2, fP10m
      nop.i         0
}
{ .mfi
      nop.m         0
      // (P3*rp + P2)*rp^2 + (P1*rp + 1)
      fma.s1        fP32p = fP32p, fRp2, fP10p
      nop.i         0
}
;;

{ .mfi
      nop.m         0
      // Nm*ln(2)/2 + Tm/2
      fma.s1        fLogTm = fNm, fLog2, fLogTm
      nop.i         0
}
{ .mfi
      nop.m         0
      // Np*ln(2)/2 + Tp/2
      fma.s1        fLogTp = fNp, fLog2, fLogTp
      nop.i         0
}
;;

{ .mfi
      nop.m         0
      // ((P3*rm + P2)*rm^2 + (P3*rm + 1))*0.5*rm + (Nm*ln(2)/2 + Tm/2)
      fma.d.s1      fP32m = fP32m, fRm, fLogTm
      nop.i         0
}
{ .mfi
      nop.m         0
      // ((P3*rp + P2)*rp^2 + (P3*rp + 1))*0.5*rp + (Np*ln(2)/2 + Tp/2)
      fma.d.s1      fP32p = fP32p, fRp, fLogTp
      nop.i         0
}
;;

{ .mfb
      nop.m         0
      // atanhf(x) = 0.5 * (log(1 + x) - log(1 - x))
      fnma.s.s0     f8 = fP32m, f1, fP32p
      br.ret.sptk   b0                      // Exit for 2^(-20) <= |x| < 1.0
}
;;


ATANH_UNORM:
// Here if x=unorm
{ .mfi
      getf.exp      rArgSExpb = fNormX           // Recompute if x unorm
      fclass.m      p0,p9 = fNormX, 0x0b         // Test x denorm
      nop.i         0
}
;;

{ .mfb
      nop.m         0
      fcmp.lt.s0    p10,p11 = f8, f0      // Set denormal flag
(p9)  br.cond.sptk  ATANH_COMMON          // Continue if x unorm and not denorm
}
;;

.pred.rel "mutex",p6,p7
{ .mfi
      nop.m         0
(p6)  fnma.s.s0     f8 = f8,f8,f8                // Result x-x^2 if x=-denorm
      nop.i         0
}
{ .mfb
      nop.m         0
(p7)  fma.s.s0      f8 = f8,f8,f8                // Result x+x^2 if x=+denorm
      br.ret.spnt   b0                           // Exit if denorm
}
;;

// Here if |x| >= 1.0
atanhf_ge_one:
{ .mfi
      alloc         r32 = ar.pfs,1,3,4,0
      fmerge.s      fArgAbs = f0, f8        // Form |x|
      nop.i         0
}
;;

{ .mfi
      nop.m         0
      fmerge.s      f10 = f8, f8            // Save input for error call
      nop.i         0
}
;;

{ .mfi
      nop.m         0
      fcmp.eq.s1    p6,p7 = fArgAbs, f1     // Test for |x| = 1.0
      nop.i         0
}
;;

// Set error tag and result, and raise invalid flag if |x| > 1.0
{ .mfi
(p7)  mov           atanh_GR_tag = 133
(p7)  frcpa.s0      f8, p0 = f0, f0         // Get QNaN, and raise invalid
      nop.i         0
}
;;

// Set error tag and result, and raise Z flag if |x| = 1.0
{ .mfi
      nop.m         0
(p6)  frcpa.s0      fRm, p0 = f1, f0        // Get inf, and raise Z flag
      nop.i         0
}
;;

{ .mfb
(p6)  mov           atanh_GR_tag = 134
(p6)  fmerge.s      f8 = f8, fRm            // result is +-inf
      br.cond.sptk  __libm_error_region     // Exit if |x| >= 1.0
}
;;

GLOBAL_LIBM_END(atanhf)


LOCAL_LIBM_ENTRY(__libm_error_region)
.prologue

{ .mfi
      add           GR_Parameter_Y=-32,sp   // Parameter 2 value
      nop.f         0
.save   ar.pfs,GR_SAVE_PFS
      mov           GR_SAVE_PFS=ar.pfs      // Save ar.pfs
}
{ .mfi
.fframe 64
      add sp=-64,sp                         // Create new stack
      nop.f 0
      mov GR_SAVE_GP=gp                     // Save gp
};;

{ .mmi
      stfs [GR_Parameter_Y] = f1,16         // STORE Parameter 2 on stack
      add GR_Parameter_X = 16,sp            // Parameter 1 address
.save   b0, GR_SAVE_B0
      mov GR_SAVE_B0=b0                     // Save b0
};;

.body
{ .mib
      stfs [GR_Parameter_X] = f10           // STORE Parameter 1 on stack
      // Parameter 3 address
      add   GR_Parameter_RESULT = 0,GR_Parameter_Y
      nop.b 0
}
{ .mib
      stfs [GR_Parameter_Y] = f8            // STORE Parameter 3 on stack
      add   GR_Parameter_Y = -16,GR_Parameter_Y
      br.call.sptk b0=__libm_error_support# // Call error handling function
};;

{ .mmi
      add   GR_Parameter_RESULT = 48,sp
      nop.m 0
      nop.i 0
};;

{ .mmi
      ldfs  f8 = [GR_Parameter_RESULT]      // Get return result off stack
.restore sp
      add   sp = 64,sp                      // Restore stack pointer
      mov   b0 = GR_SAVE_B0                 // Restore return address
};;

{ .mib
      mov   gp = GR_SAVE_GP                 // Restore gp
      mov   ar.pfs = GR_SAVE_PFS            // Restore ar.pfs
      br.ret.sptk     b0                    // Return
};;

LOCAL_LIBM_END(__libm_error_region)


.type   __libm_error_support#,@function
.global __libm_error_support#