libm_lgammaf.s

glibc 2.9,最新版的C语言库函数

}
{ .mfb
      nop.m         0
(p7)  fma.s.s0      f8 = FR_A1,FR_x,FR_A0
      br.ret.sptk   b0
};;
// branch for calculating of ln(GAMMA(x)) for -9 < x < 1
//---------------------------------------------------------------------
.align 32
lgammaf_negrecursion:
{ .mfi
      getf.sig      GR_N = FR_int_Ntrunc
      fms.s1        FR_1pXf = FR_Xp2,f1,FR_Ntrunc // 1 + (x+1) - [x]
      mov           GR_Neg2 = 2
}
{ .mfi
      add           GR_ad_Co = 0xCE0,GR_ad_Data
      fms.s1        FR_Xf = FR_Xp1,f1,FR_Ntrunc // (x+1) - [x]
      mov           GR_Neg4 = 4
};;
{ .mfi
      add           GR_ad_Ce = 0xD00,GR_ad_Data
      fma.s1        FR_Xp6 = FR_2,FR_2,FR_Xp2 // (x+6)
      add           GR_ad_C0 = 0xB30,GR_ad_Data
}
{ .mfi
      sub           GR_Neg2 = r0,GR_Neg2
      fma.s1        FR_Xp7 = FR_2,FR_3,FR_Xp1 // (x+7)
      sub           GR_Neg4 = r0,GR_Neg4
};;
{ .mfi
      cmp.ne        p8,p0 = r0,GR_N
      fcmp.eq.s1    p13,p0 = FR_NormX,FR_Ntrunc
      and           GR_IntNum = 0xF,GR_N
}
{ .mfi
      cmp.lt        p6,p0 = GR_N,GR_Neg2
      fma.s1        FR_Xp8 = FR_2,FR_3,FR_Xp2 // (x+8)
      cmp.lt        p7,p0 = GR_N,GR_Neg4
};;
{ .mfi
      getf.d        GR_Arg = FR_NormX
(p6)  fma.s1        FR_Xp2 = FR_Xp2,FR_Xp3,f0
(p8)  tbit.z.unc    p14,p15 = GR_IntNum,0
}
{ .mfi
      sub           GR_RootInd = 0xE,GR_IntNum
(p7)  fma.s1        FR_Xp4 = FR_Xp4,FR_Xp5,f0
      add           GR_ad_Root = 0xDE0,GR_ad_Data
};;
{ .mfi
      shladd        GR_ad_Root = GR_RootInd,3,GR_ad_Root
      fms.s1        FR_x = FR_Xp1,f1,FR_Ntrunc // (x+1) - [x]
      nop.i         0
}
{ .mfb
      nop.m         0
      nop.f         0
(p13) br.cond.spnt  lgammaf_singularity
};;
.pred.rel "mutex",p14,p15
{ .mfi
      cmp.gt        p6,p0 = 0xA,GR_IntNum
(p14) fma.s1        FR_Req = FR_Req,FR_Xf,f0
      cmp.gt        p7,p0 = 0xD,GR_IntNum
}
{ .mfi
(p15) mov           GR_SignOfGamma = 1 // set sign of gamma(x) to 1
(p15) fnma.s1       FR_Req = FR_Req,FR_Xf,f0
      cmp.leu       p0,p13 = 2,GR_RootInd
};;
{ .mfi
      nop.m         0
(p6)  fma.s1        FR_Xp6 = FR_Xp6,FR_Xp7,f0
(p13) add           GR_ad_RootCo = 0xE00,GR_ad_Data
};;
{ .mfi
      nop.m         0
      fcmp.eq.s1    p12,p11 = FR_1pXf,FR_2
      nop.i         0
};;
{ .mfi
      getf.sig      GR_Sig = FR_1pXf
      fcmp.le.s1    p9,p0 = FR_05,FR_Xf
      nop.i         0
}
{ .mfi
(p13) shladd        GR_RootInd = GR_RootInd,4,r0
(p7)  fma.s1        FR_Xp2 = FR_Xp2,FR_Xp4,f0
(p8)  cmp.gt.unc    p10,p0 = 0x9,GR_IntNum
};;
.pred.rel "mutex",p11,p12
{ .mfi
      nop.m         0
(p10) fma.s1        FR_Req = FR_Req,FR_Xp8,f0
(p11) extr.u        GR_Ind = GR_Sig,61,2
}
{ .mfi
(p13) add           GR_RootInd = GR_RootInd,GR_RootInd
      nop.f         0
(p12) mov           GR_Ind = 3
};;
{ .mfi
      shladd        GR_IndX2 = GR_Ind,1,r0
      nop.f         0
      cmp.gt        p14,p0 = 2,GR_Ind
}
{ .mfi
      shladd        GR_IndX8 = GR_Ind,3,r0
      nop.f         0
      cmp.eq        p6,p0 = 1,GR_Ind
};;
.pred.rel "mutex",p6,p9
{ .mfi
      shladd        GR_ad_Co = GR_IndX8,3,GR_ad_Co
(p6)  fms.s1        FR_x = FR_Xf,f1,FR_LocalMin
      cmp.gt        p10,p0 = 0xB,GR_IntNum
}
{ .mfi
      shladd        GR_ad_Ce = GR_IndX8,3,GR_ad_Ce
(p9)  fma.s1        FR_x = f0,f0,FR_1pXf
      shladd        GR_ad_C0 = GR_IndX2,4,GR_ad_C0
};;
{ .mfi
      // load coefficients of polynomial approximation
      // of ln(GAMMA(x)), 1 <= x < 2
      ldfpd         FR_A8,FR_A7 = [GR_ad_Co],16
(p10) fma.s1        FR_Xp2 = FR_Xp2,FR_Xp6,f0
      add           GR_ad_C0 = 8,GR_ad_C0
}
{ .mfi
      ldfpd         FR_A6,FR_A5 = [GR_ad_Ce],16
      nop.f         0
(p14) add           GR_ad_Root = 0x10,GR_ad_Root
};;
{ .mfi
      ldfpd         FR_A4,FR_A3 = [GR_ad_Co],16
      nop.f         0
      add           GR_ad_RootCe = 0xE10,GR_ad_Data
}
{ .mfi
      ldfpd         FR_A2,FR_A1 = [GR_ad_Ce],16
      nop.f         0
(p14) add           GR_RootInd = 0x40,GR_RootInd
};;
{ .mmi
      ldfd          FR_A0 = [GR_ad_C0]
(p13) add           GR_ad_RootCo = GR_ad_RootCo,GR_RootInd
(p13) add           GR_ad_RootCe = GR_ad_RootCe,GR_RootInd
};;
{ .mmi
(p13) ld8           GR_Root = [GR_ad_Root]
(p13) ldfd          FR_Root = [GR_ad_Root]
      mov           GR_ExpBias = 0xffff
};;
{ .mfi
      nop.m         0
      fma.s1        FR_x2 = FR_x,FR_x,f0
      nop.i         0
}
{ .mlx
(p8)  cmp.gt.unc    p10,p0 = 0xF,GR_IntNum
      movl          GR_Dx = 0x000000014F8B588E
};;
{ .mfi
      // load coefficients of polynomial approximation
      // of ln(GAMMA(x)), x is close to one of negative roots
(p13) ldfpd         FR_R3,FR_R2 = [GR_ad_RootCo]
      // argumenth for logarithm
(p10) fma.s1        FR_Req = FR_Req,FR_Xp2,f0
      mov           GR_ExpMask = 0x1ffff
}
{ .mfi
(p13) ldfpd         FR_R1,FR_R0 = [GR_ad_RootCe]
      nop.f         0
      // set p9 if signgum is 32-bit int
      // set p8 if signgum is 64-bit int
      cmp.eq        p8,p9 = 8,r34
};;
.pred.rel "mutex",p9,p8
{ .mfi
(p9)  st4           [r33] = GR_SignOfGamma // as 32-bit int
      fma.s1        FR_A7 = FR_A8,FR_x,FR_A7
(p13) sub           GR_Root = GR_Arg,GR_Root
}
{ .mfi
(p8)  st8           [r33] = GR_SignOfGamma // as 64-bit int
      fma.s1        FR_A5 = FR_A6,FR_x,FR_A5
      nop.i         0
};;
{ .mfi
      nop.m         0
      fms.s1        FR_w = FR_Req,f1,f1
(p13) add           GR_Root = GR_Root,GR_Dx
}
{ .mfi
      nop.m         0
      nop.f         0
(p13) add           GR_2xDx = GR_Dx,GR_Dx
};;
{ .mfi
      nop.m         0
      fma.s1        FR_A3 = FR_A4,FR_x,FR_A3
      nop.i         0
}
{ .mfi
      nop.m         0
      fma.s1        FR_A1 = FR_A2,FR_x,FR_A1
(p13) cmp.leu.unc   p10,p0 = GR_Root,GR_2xDx
};;
{ .mfi
      nop.m         0
      frcpa.s1      FR_InvX,p0 = f1,FR_Req
      nop.i         0
}
{ .mfi
      nop.m         0
(p10) fms.s1        FR_rx = FR_NormX,f1,FR_Root
      nop.i         0
};;
{ .mfi
      getf.exp      GR_SignExp = FR_Req
      fma.s1        FR_x4 = FR_x2,FR_x2,f0
      nop.i         0
};;
{ .mfi
      getf.sig      GR_Sig = FR_Req
      fma.s1        FR_A5 = FR_A7,FR_x2,FR_A5
      nop.i         0
};;
{ .mfi
      sub           GR_PureExp = GR_SignExp,GR_ExpBias
      fma.s1        FR_w2 = FR_w,FR_w,f0
      nop.i         0
}
{ .mfi
      nop.m         0
      fma.s1        FR_Q32 = FR_P3,FR_w,FR_P2
      nop.i         0
};;
{ .mfi
      setf.sig      FR_int_Ln = GR_PureExp
      fma.s1        FR_A1 = FR_A3,FR_x2,FR_A1
      extr.u        GR_Ind4T = GR_Sig,55,8
}
{ .mfi
      nop.m         0
      fma.s1        FR_Q10 = FR_P1,FR_w,f1
      nop.i         0
};;
{ .mfi
      shladd        GR_ad_T = GR_Ind4T,3,GR_ad_Data
      fms.s1        FR_r = FR_InvX,FR_Req,f1
      nop.i         0
}
{ .mfi
      nop.m         0
(p10) fms.s1        FR_rx2 = FR_rx,FR_rx,f0
      nop.i         0
};;
{ .mfi
      ldfd          FR_T = [GR_ad_T]
(p10) fma.s1        FR_R2 = FR_R3,FR_rx,FR_R2
      nop.i         0
}
{ .mfi
      nop.m         0
(p10) fma.s1        FR_R0 = FR_R1,FR_rx,FR_R0
      nop.i         0
};;
{ .mfi
      getf.exp      GR_Exp = FR_w
      fma.s1        FR_A1 = FR_A5,FR_x4,FR_A1
      mov           GR_ExpMask = 0x1ffff
}
{ .mfi
      nop.m         0
      fma.s1        FR_Q32 = FR_Q32, FR_w2,FR_Q10
      nop.i         0
};;
{ .mfi
      nop.m         0
      fma.s1        FR_r2 = FR_r,FR_r,f0
      mov           GR_fff7 = 0xFFF7
}
{ .mfi
      nop.m         0
      fma.s1        FR_P32 = FR_P3,FR_r,FR_P2
      nop.i         0
};;
{ .mfi
      nop.m         0
      fma.s1        FR_P10 = FR_P1,FR_r,f1
      and           GR_Exp = GR_ExpMask,GR_Exp
}
{ .mfb
      nop.m         0
(p10) fma.s.s0      f8 = FR_R2,FR_rx2,FR_R0
(p10) br.ret.spnt   b0 // exit for arguments close to negative roots
};;
{ .mfi
      nop.m         0
      fcvt.xf       FR_N = FR_int_Ln
      nop.i         0
}
{ .mfi
      cmp.ge        p14,p15 = GR_Exp,GR_fff7
      nop.f         0
      nop.i         0
};;
{ .mfi
      nop.m         0
      fma.s1        FR_A0 = FR_A1,FR_x,FR_A0
      nop.i         0
}
{ .mfi
      nop.m         0
(p15) fma.s1        FR_Ln = FR_Q32,FR_w,f0
      nop.i         0
};;
{ .mfi
      nop.m         0
(p14) fma.s1        FR_P32 = FR_P32,FR_r2,FR_P10
      cmp.eq        p6,p7 = 0,GR_Ind
};;
{ .mfi
      nop.m         0
(p14) fma.s1        FR_T = FR_N,FR_Ln2,FR_T
      nop.i         0
};;
{ .mfi
      nop.m         0
(p14) fma.s1        FR_Ln = FR_P32,FR_r,FR_T
      nop.i         0
};;
.pred.rel "mutex",p6,p7
{ .mfi
      nop.m         0
(p6)  fms.s.s0      f8 = FR_A0,FR_x,FR_Ln
      nop.i         0
}
{ .mfb
      nop.m         0
(p7)  fms.s.s0      f8 = FR_A0,f1,FR_Ln
      br.ret.sptk   b0
};;

// branch for calculating of ln(GAMMA(x)) for x < -2^13
//---------------------------------------------------------------------
.align 32
lgammaf_negstirling:
{ .mfi
      shladd        GR_ad_T = GR_Ind4T,3,GR_ad_Data
      fms.s1        FR_Xf = FR_NormX,f1,FR_N  // xf = x - [x]
      mov           GR_SingBound = 0x10016
}
{ .mfi
      add           GR_ad_Co = 0xCA0,GR_ad_Data
      fma.s1        FR_P32 = FR_P3,FR_r,FR_P2
      nop.i         0
};;
{ .mfi
      ldfd          FR_T = [GR_ad_T]
      fcvt.xf       FR_int_Ln = FR_int_Ln
      cmp.le        p6,p0 = GR_SingBound,GR_Exp
}
{ .mfb
      add           GR_ad_Ce = 0x20,GR_ad_Co
      fma.s1        FR_r2 = FR_r,FR_r,f0
(p6)  br.cond.spnt  lgammaf_singularity
};;
{ .mfi
      // load coefficients of polynomial approximation
      // of ln(sin(Pi*xf)/(Pi*xf)), |xf| <= 0.5
      ldfpd         FR_S16,FR_S14 = [GR_ad_Co],16
      fma.s1        FR_P10 = FR_P1,FR_r,f1
      nop.i         0
}
{ .mfi
      ldfpd         FR_S12,FR_S10 = [GR_ad_Ce],16
      fms.s1        FR_xm05 = FR_NormX,f1,FR_05
      nop.i         0
};;
{ .mmi
      ldfpd         FR_S8,FR_S6 = [GR_ad_Co],16
      ldfpd         FR_S4,FR_S2 = [GR_ad_Ce],16
      nop.i         0
};;
{ .mfi
      getf.sig      GR_N = FR_int_Ntrunc // signgam calculation
      fma.s1        FR_Xf2 = FR_Xf,FR_Xf,f0
      nop.i         0
};;
{ .mfi
      nop.m         0
      frcpa.s1      FR_InvXf,p0 = f1,FR_Xf
      nop.i         0
};;
{ .mfi
      getf.d        GR_Arg = FR_Xf
      fcmp.eq.s1    p6,p0 = FR_NormX,FR_N
      mov           GR_ExpBias = 0x3FF
};;
{ .mfi
      nop.m         0
      fma.s1        FR_T = FR_int_Ln,FR_Ln2,FR_T
      extr.u        GR_Exp = GR_Arg,52,11
}
{ .mfi
      nop.m         0
      fma.s1        FR_P32 = FR_P32,FR_r2,FR_P10
      nop.i         0
};;
{ .mfi
      sub           GR_PureExp = GR_Exp,GR_ExpBias
      fma.s1        FR_S14 = FR_S16,FR_Xf2,FR_S14
      extr.u        GR_Ind4T = GR_Arg,44,8
}
{ .mfb
      mov           GR_SignOfGamma = 1 // set signgam to -1
      fma.s1        FR_S10 = FR_S12,FR_Xf2,FR_S10
(p6)  br.cond.spnt  lgammaf_singularity
};;
{ .mfi
      setf.sig      FR_int_Ln = GR_PureExp
      fms.s1        FR_rf = FR_InvXf,FR_Xf,f1
      // set p14 if GR_N is even
      tbit.z        p14,p0 = GR_N,0
}
{ .mfi