s_asinh.s

glibc 2.9,最新版的C语言库函数

;;

{ .mfi
      getf.sig   log_GR_significand_f8 = log_arg_early
      // (0.5*z)*(3-(y*z)*z)*y + |x|
      fma.s1     log_arg = log_y_rs_iter1,log_y_rs,asinh_f8
      //to get third table address
      adds       log_table_address3 = 0x70, NR_table_address
}
;;

///////////////////////////////// The end NR iterations /////////////////////
{ .mfi
      nop.m      0
      nop.f      0
      //significant bit destruction
      and        log_GR_exp_f8 = log_GR_signexp_f8, log_GR_exp_17_ones
}
;;

{ .mfi
      //BIAS subtraction
      sub        log_GR_true_exp_f8 = log_GR_exp_f8, log_GR_exp_16_ones
(p7)  fnma.s1    log2 = log2,f1,f0
      nop.i      0
}
;;

{ .mfi
      setf.sig   log_int_Nfloat = log_GR_true_exp_f8
      fms.s1     log_r = log_C,log_arg,f1  // C = frcpa(x); r = C * x - 1
      extr.u     log_GR_index = log_GR_significand_f8,55,8 //Extract 8 bits
}
;;

{ .mmi
      //pre-index*16 + index
      shladd     log_table_address3 = log_GR_index,4,log_table_address3
;;
      ldfe       log_T = [log_table_address3]
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     log_rsq = log_r, log_r, f0          //r^2
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     log_rp_p4 = log_P5, log_r, log_P4   //P5*r + P4
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     log_rp_p32 = log_P3, log_r, log_P2  //P3*r + P2
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      //convert N to the floating-point format
      fcvt.xf    log_Nfloat = log_int_Nfloat
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     log_rcube = log_rsq, log_r, f0      //r^3
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     log_rp_p10 = log_rsq, log_P1, log_r //P1*r^2 + r
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      //(P5*r + P4)*r^2 + P3*r + P2
      fma.s1     log_rp_p2 = log_rp_p4, log_rsq, log_rp_p32
      nop.i      0
}
;;

.pred.rel "mutex",p7,p11
{ .mfi
      nop.m      0
(p11) fma.s1     log_T_plus_Nlog2 = log_Nfloat,log2,log_T  //N*log2 + T if x>0
      nop.i      0
}
{ .mfi
      nop.m      0
(p7)  fms.s1     log_T_plus_Nlog2 = log_Nfloat,log2,log_T  //N*log2 - T if x<0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      //((P5*r + P4)*r^2 + P3*r + P2)*w^3 + P1*r^2 + r
      fma.s1     log_r2P_r = log_rp_p2, log_rcube, log_rp_p10
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      //  N*log2 + T + ((P5*r + P4)*r^2 + P3*r + P2)*r^3 + P1*r^2 + r
(p11) fadd.d.s0  f8 = log_T_plus_Nlog2,log_r2P_r
      nop.i      0
}
{ .mfb
      nop.m      0
      // -N*log2 - T - ((P5*r + P4)*r^2 + P3*r + P2)*r^3 + P1*r^2 + r
(p7)  fsub.d.s0  f8 = log_T_plus_Nlog2,log_r2P_r
      br.ret.sptk b0           // Exit main path, path 3: 2^-3 <= |x| < 2^63
}
;;

// Here if path 4, |x| >= 2^63
LOG_COMMON1:
{ .mfi
      ldfpd      log_P5,log_P4 = [NR_table_address],16
      nop.f      0
      nop.i      0
}
;;

{ .mfi
      ldfpd      log_P3,log_P2 = [NR_table_address],16
      frcpa.s1   log_C,p0 = f1,log_arg
      nop.i      0
}
;;

{ .mmi
      getf.exp   log_GR_signexp_f8 = log_arg
      ldfd       log_P1 = [NR_table_address],16
      nop.i      0
}
;;

{ .mmi
      getf.sig   log_GR_significand_f8 = log_arg
      ldfe       log2 = [NR_table_address],16
      nop.i      0
}
;;

{ .mfi
      adds       log_table_address3 = 0x70, NR_table_address
      nop.f      0
      //significant bit destruction
      and        log_GR_exp_f8 = log_GR_signexp_f8, log_GR_exp_17_ones
}
;;

{ .mmf
      nop.m      0
      //BIAS subtraction
      sub        log_GR_true_exp_f8 = log_GR_exp_f8, log_GR_exp_16_ones
      fms.s1     log_r = log_C,log_arg,f1  //C = frcpa(x); r = C * x - 1
}
;;

{ .mfi
      setf.sig   log_int_Nfloat = log_GR_true_exp_f8
      nop.f      0
      extr.u     log_GR_index = log_GR_significand_f8,55,8 //Extract 8 bits
}
;;

{ .mmi
      //pre-index*16 + index
      shladd     log_table_address3 = log_GR_index,4,log_table_address3
;;
      ldfe       log_T = [log_table_address3]
      nop.i      0

}
;;

{ .mfi
      nop.m      0
      fma.s1     log_rsq = log_r, log_r, f0          //r^2
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     log_rp_p4 = log_P5, log_r, log_P4   //P5*r + P4
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     log_rp_p32 = log_P3, log_r, log_P2  //P3*r + P2
      nop.i      0
}
{ .mfi
      nop.m      0
(p7)  fnma.s1    log2 = log2,f1,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     log_rcube = log_rsq, log_r, f0      //r^3
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     log_rp_p10 = log_rsq, log_P1, log_r //P1*r^2 + r
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      //convert N to the floating-point format
      fcvt.xf    log_Nfloat = log_int_Nfloat
      nop.i      0
}
{ .mfi
      nop.m      0
      //(P5*r + P4)*r^2 + P3*r + P2
      fma.s1     log_rp_p2 = log_rp_p4, log_rsq, log_rp_p32
      nop.i      0
}
;;

{ .mfi
      nop.m      0
(p7)  fnma.s1    log_T = log_T,f1,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     log_T_plus_Nlog2 = log_Nfloat,log2,log_T    //N*log2 + T
      nop.i      0
}
{ .mfi
      nop.m      0
      //((P5*r + P4)*r^2 + P3*r + P2)*w^3 + P1*r^2 + r
      fma.s1     log_r2P_r = log_rp_p2, log_rcube, log_rp_p10
      nop.i      0
}
;;

.pred.rel "mutex",p7,p11
{ .mfi
      nop.m      0
      //  N*log2 + T + ((P5*r + P4)*r^2 + P3*r + P2)*r^3 + P1*r^2 + r
(p11) fadd.d.s0  f8 = log_T_plus_Nlog2,log_r2P_r
      nop.i      0
}
{ .mfb
      nop.m      0
      // -N*log2 - T - ((P5*r + P4)*r^2 + P3*r + P2)*r^3 + P1*r^2 + r
(p7)  fsub.d.s0  f8 = log_T_plus_Nlog2,log_r2P_r
      br.ret.sptk b0              // Exit path 4, |x| >= 2^63
}
;;

// Here is path 2, 0 < |x| < 2^-3
ASINH_NEAR_ZERO:
{ .mfi
      ldfe       log_C9 = [log_table_address2],16
      fma.s1     asinh_w_cube = asinh_w_sq,fNormX,f0
      nop.i      0
}
;;

{ .mfi
      ldfe       log_C7 = [log_table_address2],16
      fma.s1     asinh_w_four = asinh_w_sq,asinh_w_sq,f0
      nop.i      0
}
;;

{ .mfi
      ldfe       log_C5 = [log_table_address2],16
      nop.f      0
      nop.i      0
}
;;

{ .mfi
      ldfe       log_C3 = [log_table_address2],16
      nop.f      0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     asinh_w_13 = log_C13,asinh_w_sq,log_C11
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     asinh_w_9 = log_C9,asinh_w_sq,log_C7
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     asinh_w_3 = log_C5,asinh_w_sq,log_C3
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     asinh_w_seven = asinh_w_four,asinh_w_cube,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     asinh_w_7 = asinh_w_13,asinh_w_four,asinh_w_9
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     asinh_w_5 = asinh_w_3,asinh_w_cube,fNormX
      nop.i      0
}
;;

{ .mfb
      nop.m      0
      fma.d.s0   f8 = asinh_w_7,asinh_w_seven,asinh_w_5
      br.ret.sptk b0                   // Exit path 2 (0.0 <|x| < 2^(-3))
}
;;

ASINH_UNORM:
// Here if x=unorm
{ .mfi
      getf.exp   asinh_GR_f8 = fNormX  // Recompute if x unorm
      fclass.m   p0,p13 = fNormX, 0x0b // Test x denorm
      nop.i      0
}
;;

{ .mfb
      nop.m      0
      fcmp.eq.s0 p14,p0 = f8, f0       // Dummy to set denormal flag
(p13) br.cond.sptk ASINH_COMMON        // Continue if x unorm and not denorm
}
;;

.pred.rel "mutex",p7,p11
{ .mfi
      nop.m      0
(p7)  fma.d.s0   f8 = f8,f8,f8         // Result x+x^2 if x=-denorm
      nop.i      0
}
{ .mfb
      nop.m      0
(p11) fnma.d.s0  f8 = f8,f8,f8         // Result x-x^2 if x=+denorm
      br.ret.spnt b0                   // Exit if denorm
}
;;

GLOBAL_LIBM_END(asinh)