e_acosh.s

glibc 2.9,最新版的C语言库函数

      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 log_Nfloat
      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
}
;;

{ .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)*r^3 + P1*r^2 + r
      fma.s1     log_r2P_r = log_rp_p2, log_rcube, log_rp_p10
      nop.i      0
}
;;

{ .mfb
      nop.m      0
      // N*log2 + T + ((P5*r + P4)*r^2 + P3*r + P2)*w^3 + P1*r^2 + r
      fadd.d.s0  f8 = log_T_plus_Nlog2, log_r2P_r
      br.ret.sptk b0           // Exit main path, path 3: 1.0005 <= x < 2^63
}
;;

// Here if path 2, 1.0 < x < 1.0005
ACOSH_NEAR_ONE:
// The first NR iteration
{ .mfi
      ldfe       log_C1 = [log_table_address2],16
      fma.s1     acosh_w_iter1 = acosh_w_rs,log_w,f0  //t*w
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     acosh_w_1 = f8,log_C4,log_C3         //x*C4 + C3
      nop.i      0
}
;;

{ .mfi
      ldfe       log_C0 = [log_table_address2],16
      fma.s1     acosh_w_iter2 = acosh_w_rs,NR1,f0    //t*0.5
      nop.i      0
}
{ .mfi
      nop.m      0
      fnma.s1    acosh_w_iter1 = acosh_w_iter1,acosh_w_rs,NR2 //3-t*t*w
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      //(3-t*t*w)*t*0.5
      fma.s1     acosh_w_iter2 = acosh_w_iter2,acosh_w_iter1,f0
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     acosh_w_1 = acosh_w_1,log_w,log_C2 //(x*C4 + C3)*(x-1) + C2
      nop.i      0
}
;;

// The second NR iteration
{ .mfi
      nop.m      0
      fma.s1     acosh_w_rs = acosh_w_iter2,log_w,f0  //t*w
      nop.i      0
}
{ .mfi
      nop.m      0
      //((x*C4 + C3)*(x-1) + C2)*(x-1) + C1
      fma.s1     acosh_w_1 = acosh_w_1,log_w,log_C1
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fnma.s1    acosh_w_iter1 = acosh_w_iter2,acosh_w_rs,NR2
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     acosh_w_iter2 = acosh_w_iter2,NR1,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     acosh_w_iter2 = acosh_w_iter2,acosh_w_iter1,f0
      nop.i      0
}
{ .mfi
      nop.m      0
      //(((x*C4 + C3)*(x-1) + C2)*(x-1) + C1)*(x-1) + C0
      fma.s1     acosh_w_1 = acosh_w_1,log_w,log_C0
      nop.i      0
}
;;

//The third NR iteration
{ .mfi
      nop.m      0
      fma.s1     acosh_w_rs = acosh_w_iter2,log_w,f0  //t*w
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fnma.s1    acosh_w_iter1 = acosh_w_iter2,acosh_w_rs,NR2
      nop.i      0
}
{ .mfi
      nop.m      0
      fma.s1     acosh_w_iter2 = acosh_w_iter2,NR1,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     acosh_w_iter2 = acosh_w_iter2,acosh_w_iter1,f0
      nop.i      0
}
;;

{ .mfi
      nop.m      0
      fma.s1     acosh_w_sqrt = acosh_w_iter2,log_w,f0
      nop.i      0
}
;;

{ .mfb
      nop.m      0
      fma.d.s0   f8 = acosh_w_1,acosh_w_sqrt,f0
      br.ret.sptk b0               // Exit path 2, 1.0 < x < 1.0005
}
;;

// Here if path 4, x >= 2^63
LOG_COMMON1:
{ .mfi
      ldfpd      acosh_comp,log_P5 = [NR_table_address],16
      frcpa.s1   log_C,p0 = f1,log_arg
      nop.i      0
}
;;

{ .mmi
      getf.exp   log_GR_signexp_f8 = log_arg
      ldfpd      log_P4,log_P3 = [NR_table_address],16
      nop.i      0
}
;;

{ .mmi
      getf.sig   log_GR_significand_f8 = log_arg
      ldfpd      log_P2,log_P1 = [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
      ldfe       log2 = [NR_table_address],16
      //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
      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 log_Nfloat
      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
      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
}
;;

{ .mfb
      nop.m      0
      //  N*log2 + T + ((P5*r + P4)*r^2 + P3*r + P2)*w^3 + P1*r^2 + r
      fadd.d.s0  f8 = log_T_plus_Nlog2, log_r2P_r
      br.ret.sptk b0              // Exit path 4, x >= 2^63
}
;;

// Here if path 7, x < 1.0
ACOSH_LESS_ONE:
{ .mfi
      alloc      r32 = ar.pfs,1,3,4,0
      fmerge.s   f10 = f8,f8
      nop.i      0
}
;;

{ .mfb
      mov        acosh_GR_tag = 136
      frcpa.s0   f8,p0 = f0,f0
      br.cond.sptk __libm_error_region
}
;;

GLOBAL_LIBM_END(acosh)


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
        stfd [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
        stfd [GR_Parameter_X] = f10           // STORE Parameter 1 on stack
        add   GR_Parameter_RESULT = 0,GR_Parameter_Y  // Parameter 3 address
        nop.b 0
}
{ .mib
        stfd [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
        ldfd  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#