s_erfcf.s

glibc 2.9,最新版的C语言库函数

}
{ .mfi
      setf.d         EXP_RSHF = exp_GR_rshf   // Form right shift 1.100 * 2^63
(p7)  fms.s1         FR_AbsArg = f1, f0, f8                   // |x|  if x < 0
      mov            exp_TB1_size  = 0x100 
}
;;

// Form pointer GR_P_POINT_3 to the beginning of erfc_p_table 
{ .mfi
      setf.d         FR_05 = GR_05
      nop.f          0
      sub            GR_ShftPi = GR_ShftPi,GR_ShftPi_bias
}
{ .mfb
      add            GR_P_POINT_3 = GR_P_POINT_3, EXP_AD_TB1
      nop.f          0
(p9)  br.cond.spnt   SPECIAL                  // For x = 0,+inf,-inf,nan,unnorm
}
;;

{ .mfi
      add            GR_P_POINT_1 = GR_P_POINT_3, GR_ShftPi
      nop.f          0
      add            GR_P_POINT_2 = GR_P_POINT_3, GR_ShftPi
}
{ .mfi
      ldfe           exp_ln2_by_128_hi  = [EXP_AD_TB1],16 
      fma.s1         FR_NormX = f8,f1,f0
      add            GR_P_POINT_3 = GR_P_POINT_3, GR_ShftPi
}
;;

// Load coefficients for polynomial P15(x)
{ .mfi
      ldfpd          FR_A15, FR_A14 = [GR_P_POINT_1], 16
      nop.f          0      
      add            GR_P_POINT_3 = 0x30, GR_P_POINT_3
}
{ .mfi
      ldfe           exp_ln2_by_128_lo  = [EXP_AD_TB1], 16
      nop.f          0      
      add            GR_P_POINT_2 = 0x20, GR_P_POINT_2 
}
;;

// Now EXP_AD_TB1 points to the beginning of table 1
{ .mlx
      ldfpd          FR_A13, FR_A12 = [GR_P_POINT_1] 
      movl           GR_1_by_6 = 0x3FC5555555555555
}
{ .mfi
      add            GR_P_POINT_4 = 0x30, GR_P_POINT_2
      nop.f          0
      nop.i          0
}
;;

{ .mfi
      ldfpd          FR_A11, FR_A10 = [GR_P_POINT_2]
      fma.s1         FR_2 = f1, f1, f1
      mov            exp_TB2_size  = 0x80
}
{ .mfi
      ldfpd          FR_A9, FR_A8 = [GR_P_POINT_3],16
      nop.f          0
      add            GR_P_POINT_1 = 0x60 ,GR_P_POINT_1
}
;;

// W = X * Inv_log2_by_128
// By adding 1.10...0*2^63 we shift and get round_int(W) in significand.
// We actually add 1.10...0*2^56 to X * Inv_log2 to do the same thing.
{ .mfi
      ldfpd          FR_A7, FR_A6 = [GR_P_POINT_3] 
      fma.s1     EXP_W_2TO56_RSH = EXP_NORM_f8,EXP_INV_LN2_2TO63,EXP_RSHF_2TO56
      add            EXP_AD_TB2 = exp_TB1_size, EXP_AD_TB1
      
}
{ .mfi
      ldfpd          FR_A5, FR_A4 = [GR_P_POINT_4], 16
      nop.f          0
      nop.i          0
}
;;

{ .mfi
      ldfpd          FR_A3, FR_A2 = [GR_P_POINT_4]
      fmerge.s       FR_X = f8,f8
      nop.i          0
}
{ .mfi 
      ldfpd          FR_A1, FR_A0 = [GR_P_POINT_1]
      nop.f          0
      nop.i          0
}
;;

//p14: x < - NEG_ARG_ASYMP = -4.4 -> erfcf(x) ~=~ 2.0
{ .mfi
      setf.d         FR_1_by_6  = GR_1_by_6
(p7)  fcmp.gt.unc.s1 p14,p0 = FR_AbsArg, FR_NEG_ARG_ASYMP          //p7: x < 0
      nop.i          0
}
;;

//p15: x > POS_ARG_ASYMP = 10.06 -> erfcf(x) ~=~ 0.0
{ .mfi
      nop.m          0
(p6)  fcmp.gt.unc.s1 p15,p0 = FR_AbsArg, FR_POS_ARG_ASYMP          //p6: x > 0
      nop.i          0
}                       
;;

{ .mfi
      nop.m          0
      fcmp.le.s1     p8,p0 = FR_NormX, FR_UnfBound        // p8: x <= UnfBound
      nop.i          0
}
{ .mfb
      nop.m          0
(p14) fnma.s.s0      FR_RESULT = FR_EpsNorm, FR_EpsNorm, FR_2//y = 2 if x <-4.4
(p14) br.ret.spnt    b0
}
;;

// Nfloat = round_int(W) 
// The signficand of EXP_W_2TO56_RSH contains the rounded integer part of W,
// as a twos complement number in the lower bits (that is, it may be negative).
// That twos complement number (called N) is put into exp_GR_N.

// Since EXP_W_2TO56_RSH is scaled by 2^56, it must be multiplied by 2^-56
// before the shift constant 1.10000 * 2^63 is subtracted to yield EXP_Nfloat.
// Thus, EXP_Nfloat contains the floating point version of N

{ .mfi
      nop.m          0
      fms.s1         EXP_Nfloat = EXP_W_2TO56_RSH, EXP_2TOM56, EXP_RSHF
      nop.i          0
} 
{ .mfb
(p15) mov            GR_Parameter_TAG = 209
(p15) fma.s.s0       FR_RESULT = FR_EpsNorm,FR_EpsNorm,f0 //Result.for x>10.06
(p15) br.cond.spnt   __libm_error_region
}                   
;;

// Now we can calculate polynomial P15(x)
{ .mfi
      nop.m          0
      fma.s1         FR_P15_1_1 = FR_AbsArg, FR_AbsArg, f0             // x ^2
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         FR_P15_0_1 = FR_A15, FR_AbsArg, FR_A14
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         FR_P15_1_2 = FR_A13, FR_AbsArg, FR_A12 
      nop.i          0 
}
;;

{ .mfi
      getf.sig       exp_GR_N        = EXP_W_2TO56_RSH 
      fma.s1         FR_P15_2_1 = FR_A9, FR_AbsArg, FR_A8 
      nop.i          0 
}
{ .mfi
      nop.m          0
      fma.s1         FR_P15_2_2 = FR_A11, FR_AbsArg, FR_A10 
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         FR_P15_3_1 = FR_A5, FR_AbsArg, FR_A4
      nop.i          0                   
}
{ .mfi
      nop.m          0
      fma.s1         FR_P15_3_2 = FR_A7, FR_AbsArg, FR_A6
      nop.i          0
}
;;

// exp_GR_index_1 has index_1
// exp_GR_index_2_16 has index_2 * 16
// exp_GR_biased_M has M
// exp_GR_index_1_16 has index_1 * 16

// r2 has true M
{ .mfi
      and            exp_GR_index_1 = 0x0f, exp_GR_N
      fma.s1         FR_P15_4_1 = FR_A1, FR_AbsArg, FR_A0
      shr            r2 = exp_GR_N,  0x7
       
}
{ .mfi
      and            exp_GR_index_2_16 = 0x70, exp_GR_N
      fma.s1         FR_P15_4_2 = FR_A3, FR_AbsArg, FR_A2
      nop.i          0
}
;;

// EXP_AD_T1 has address of T1                           
// EXP_AD_T2 has address if T2                            

{ .mfi
      add            EXP_AD_T2 = EXP_AD_TB2, exp_GR_index_2_16
      nop.f          0                      
      shladd         EXP_AD_T1 = exp_GR_index_1, 4, EXP_AD_TB1
}
{ .mfi
      addl           exp_GR_biased_M = 0xffff, r2
      fnma.s1        exp_r   = EXP_Nfloat, exp_ln2_by_128_hi, EXP_NORM_f8
      nop.i          0
}
;;

// Create Scale = 2^M
// r = x - Nfloat * ln2_by_128_hi 
 
{ .mfi
      setf.exp       EXP_2M = exp_GR_biased_M
      fma.s1         FR_P15_7_1 = FR_P15_0_1, FR_P15_1_1, FR_P15_1_2
      nop.i          0
}
{ .mfi
      ldfe           exp_T2  = [EXP_AD_T2]      
      nop.f          0
      nop.i          0
}
;;

// Load T1 and T2

{ .mfi
      ldfe           exp_T1  = [EXP_AD_T1]
      fma.s1         FR_P15_7_2 = FR_P15_1_1, FR_P15_1_1, f0            // x^4
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         FR_P15_8_1 = FR_P15_1_1, FR_P15_2_2, FR_P15_2_1 
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         FR_P15_9_1 = FR_P15_1_1, FR_P15_4_2, FR_P15_4_1 
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         FR_P15_9_2 = FR_P15_1_1, FR_P15_3_2, FR_P15_3_1 
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         exp_P = FR_1_by_6, exp_r, FR_05 
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         exp_rsq = exp_r, exp_r, f0 
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         FR_P15_13_1 = FR_P15_7_2, FR_P15_7_1, FR_P15_8_1 
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         FR_P15_14_1 = FR_P15_7_2, FR_P15_9_2, FR_P15_9_1 
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         FR_P15_14_2 = FR_P15_7_2, FR_P15_7_2, f0           // x^8
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         exp_P     = exp_P, exp_rsq, exp_r
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         exp_S1  = EXP_2M, exp_T2, f0   
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         FR_Pol = FR_P15_14_2, FR_P15_13_1, FR_P15_14_1  // P15(x)
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         exp_S   = exp_S1, exp_T1, f0
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         FR_Exp = exp_S, exp_P, exp_S                 // exp(-x^2)
      nop.i          0          
}
;;   

{ .mfi
      nop.m          0
      fma.s.s0       FR_Tmpf = f8, f1, f0                          //  Flag  d
      nop.i          0 
}
;;

//p6: result for     0 < x < = POS_ARG_ASYMP 
//p7: result for   - NEG_ARG_ASYMP  <= x < 0
//p8: exit   for   - NEG_ARG_ASYMP <= x <= UnfBound, x!=0
.pred.rel "mutex",p6,p7
{ .mfi
      nop.m          0
(p6)  fma.s.s0       f8 = FR_Exp, FR_Pol, f0 
      nop.i          0         
}
{ .mfb
      mov            GR_Parameter_TAG = 209
(p7)  fnma.s.s0      f8 = FR_Exp, FR_Pol, FR_2
(p8)  br.ret.sptk    b0 
}
;;

//p10: branch for  UnfBound < x < = POS_ARG_ASYMP
{ .mfb
      nop.m          0
      nop.f          0
(p10) br.cond.spnt   __libm_error_region  
}
;;

//Only via (p9)  br.cond.spnt   SPECIAL  for x = 0,+inf,-inf,nan,unnorm
SPECIAL:

{ .mfi
      nop.m          0
      fclass.m.unc   p10,p0 = f8,0x07                            // p10: x = 0
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fclass.m.unc   p11,p0 = f8,0x21                         // p11: x = +inf
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fclass.m.unc   p12,p0 = f8,0x22                          // p12 x = -inf
      nop.i          0
}
{ .mfb
      nop.m          0
(p10) fma.s.s0       f8 = f1, f1, f0
(p10) br.ret.sptk    b0                                // Quick exit for x = 0
}
;;

{ .mfi
      nop.m          0
      fclass.m.unc   p13,p0 = f8,0xc3                          // p13: x = nan
      nop.i          0
}
{ .mfb
      nop.m          0
(p11) fma.s.s0       f8 = f0, f1, f0
(p11) br.ret.spnt    b0                             // Quick exit for x = +inf
}
;;
{ .mfi
      nop.m          0
      fclass.m.unc   p14,p0 = f8,0x0b                 // P14: x = unnormalized
      nop.i          0
}
{ .mfb
      nop.m          0
(p12) fma.s.s0       f8 = f1, f1, f1
(p12) br.ret.spnt    b0                             // Quick exit for x = -inf
}
;;

{ .mfb
      nop.m          0
(p13) fma.s.s0       f8 = f8, f1, f0
(p13) br.ret.sptk    b0                              // Quick exit for x = nan
}
;;

{ .mfb
      nop.m          0
(p14) fnma.s.s0      f8 = f8, f1, f1
(p14) br.ret.sptk    b0                     // Quick exit for x = unnormalized
}
;;

GLOBAL_LIBM_END(erfcf)


// Call via (p10) br.cond.spnt   __libm_error_region
//          for  UnfBound < x < = POS_ARG_ASYMP 
// and
// 
// call via (p15) br.cond.spnt   __libm_error_region
//          for  x > POS_ARG_ASYMP

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] = FR_Y,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] = FR_X             // STORE Parameter 1 on stack
        add   GR_Parameter_RESULT = 0,GR_Parameter_Y    // Parameter 3 address
        nop.b 0                                      
}
{ .mib
        stfs [GR_Parameter_Y] = FR_RESULT        // 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
        nop.m 0
        nop.m 0
        add   GR_Parameter_RESULT = 48,sp
};;
{ .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#