e_powl.s

Glibc 2.3.2源代码(解压后有100多M)

	nop.i 999 ;;
}
{ .mfi
(p0)   and GR_Expo_X = GR_Expo_X, GR_Mask  
//
//     Calculate unBIASed exponent of X
//     Point to Table of W1s
//     Point to Table of W2s
//
(p0)   fcvt.fx.s1 FR_N = FR_float_N 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p0)   fadd.s1 FR_P_lo = FR_P_lo, FR_X_lo 
//
//     Float_N = X * L_Inv
//     Create exponent BIAS
//     Get BIASed exponent of X
//
(p0)   sub GR_Expo_X = GR_Expo_X, GR_BIAS ;;  
}
{ .mib
(p0)   cmp.gt.unc  p9, p0  =  -6, GR_Expo_X 
	nop.i 999
//
//     N = fcvt.fx(float_N)
//     If -6 > Expo_X, set P9
//
(p9)   br.cond.spnt L(EXPL_SMALL) 
}
;;

//
//     If expo_X < -6 goto exp_small
//
{ .mmi
      nop.m 999
(p0)  addl           GR_T1_ptr   = @ltoff(Constants_exp_64_T1#), gp
(p0)  cmp.lt.unc  p10, p0 =  14, GR_Expo_X 
}
;;

{ .mmi
      ld8 GR_T1_ptr = [GR_T1_ptr]
      nop.m 999
      nop.i 999
}
;;

{ .mib
	nop.m 999
	nop.i 999
//
//     If 14 < Expo_X, set P10
//     Create pointer to T1 table
//  
(p10)  br.cond.spnt L(EXPL_HUGE) ;;
}


{ .mmi
(p0)  addl           GR_Table_Ptr   = @ltoff(Constants_exp_64_Exponents#), gp
(p0)  addl           GR_T2_ptr   = @ltoff(Constants_exp_64_T2#), gp
      nop.i 999
}
;;

{ .mmi
      ld8 GR_Table_Ptr = [GR_Table_Ptr]
      ld8 GR_T2_ptr = [GR_T2_ptr]
      nop.i 999
}
;;


{ .mmi
(p0)   shladd GR_Table_Ptr = GR_Expo_Range,4,GR_Table_Ptr ;;  
//
//     Adjust T1_ptr by x 4 for single-precision values
//     Adjust T2_ptr by x 4 for single-precision values
//
(p0)   ld8 GR_Big_Pos_Exp = [GR_Table_Ptr],8
	nop.i 999 ;;
}
//
//     Load double W1
//     Load +max exponent
//
{ .mfi
(p0)   ld8 GR_Big_Neg_Exp = [GR_Table_Ptr],0
//
//     If 14 < Expo_X, goto exp_huge
//
(p0)   fcvt.xf FR_float_N = FR_N 
	nop.i 999 
}
;;

//
//     Load double W2
//     Load -max exponent
//     Load ptr to A's
//

{ .mmi
(p0)  getf.sig GR_N_fix = FR_N 
(p0)  addl           GR_Table_Ptr   = @ltoff(Constants_exp_64_A#), gp
      nop.i 999
}
;;

{ .mmi
      ld8 GR_Table_Ptr = [GR_Table_Ptr]
      nop.m 999
      nop.i 999
}
;;

//
//     Load single T1
//     Load single T2
//     W_1_p1 = W_1 + 1
//
{ .mmi
(p0)   ldfe FR_A_3 = [GR_Table_Ptr],16 ;; 
//
//     Load A_3
//     if k > big_pos_exp, set p14 and Safe=False
//
(p0)   ldfe FR_A_2 = [GR_Table_Ptr],16 
(p0)   extr.u GR_M1 = GR_N_fix, 6, 6  
}
{ .mmi
	nop.m 999 ;;
(p0)   shladd GR_W1_ptr = GR_M1,3,GR_W1_ptr  
//
//     float_N = fcvt.xf(N) 
//     N_fix = significand of N
//     Create pointer to T2 table
//
(p0)   extr.u GR_M2 = GR_N_fix, 0, 6  
}
//
//     r = r + X_cor
//     Adjust W1_ptr by x 8 for double-precision values
//     Adjust W2_ptr by x 8 for double-precision values
//     Adjust Table_ptr by Expo_Rangex16 
//
{ .mmi
(p0)   shladd GR_T1_ptr = GR_M1,2,GR_T1_ptr ;;  
(p0)   ldfd  FR_W1 = [GR_W1_ptr],0 
(p0)   shladd GR_W2_ptr = GR_M2,3,GR_W2_ptr  
}
//
//     Load ptr to A's
//
{ .mfi
(p0)   ldfs  FR_T1 = [GR_T1_ptr],0 
(p0)   fnma.s1 FR_r = FR_L_hi, FR_float_N, FR_X 
(p0)   shladd GR_T2_ptr = GR_M2,2,GR_T2_ptr ;; 
}
{ .mmi
(p0)   ldfd  FR_W2 = [GR_W2_ptr],0 
(p0)   ldfs  FR_T2 = [GR_T2_ptr],0 
//
//     r = x - L_hi * float_N
//     M2 = extr.u(N_fix,0,6)
//     M1 = extr.u(N_fix,6,6)
//
(p0)   extr GR_k = GR_N_fix, 12, 52 ;;  
}
//
//     Load A_1
//     poly = A_3 * r + A_2
//     rsq = r*r
//
{ .mii
(p0)   add GR_BIAS_p_k = GR_BIAS, GR_k  
(p0)   cmp.gt.unc  p14,p15 = GR_k,GR_Big_Pos_Exp ;; 
(p15)  cmp.lt p14,p15 = GR_k,GR_Big_Neg_Exp
}
//
//     BIAS_p_K = BIAS + k
//     T = T1 * T2
//
{ .mfi
(p0)   setf.exp FR_Scale = GR_BIAS_p_k  
	nop.f 999
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p0)   fnma.s1 FR_r = FR_L_lo, FR_float_N, FR_r 
	nop.i 999
}
//
//     W = W_1_p1 * W2 + W1
//
{ .mfi
(p0)   ldfe FR_A_1 = [GR_Table_Ptr],16 
	nop.f 999
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p0)   fadd.s1 FR_W_1_p1 = FR_W1, f1 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     k = extr.u(N_fix,0,6)
//     r = r - N * L_lo
//     Load ptr to Table of exponent thresholds.
//
(p0)   fadd.s1 FR_r = FR_r, FR_X_cor 
	nop.i 999
}
{ .mfi
	nop.m 999
(p0)   fmpy.s1 FR_T = FR_T1, FR_T2 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     if k < big_neg_exp, set p14 and Safe=False
//     Load A_2
//
(p0)   fma.s1 FR_W = FR_W2, FR_W_1_p1, FR_W1 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p0)   fma.s1 FR_poly = FR_r, FR_A_3, FR_A_2 
	nop.i 999
}
{ .mfi
	nop.m 999
(p0)   fmpy.s1 FR_rsq = FR_r, FR_r 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p0)   mov FR_Y_hi = FR_T 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     Scale = set_exp(BIAS_p_k)
//     poly = r * poly + A_1
//
(p0)   fadd.s1 FR_Wp1 = FR_W, f1 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p0)   fma.s1 FR_poly = FR_r, FR_poly, FR_A_1 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p0)   fma.s1 FR_poly = FR_rsq, FR_poly,FR_r  
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     Wp1 =  W + 1
//     poly = rsq * poly + rk
//
(p0)   fma.s1 FR_Y_lo = FR_Wp1, FR_poly, FR_W 
	nop.i 999 ;;
}
{ .mfb
	nop.m 999
//
//     Y_lo = poly * Wp1 + W 
//     Y_hi = T
//
(p0)   fmpy.s1 FR_Y_lo = FR_Y_lo, FR_T 
//
//     Y_lo = T * Y_lo
//
(p0)   br.cond.sptk L(EXPL_RETURN) ;; 
}

L(EXPL_SMALL): 

//
//     r4 = rsq * rsq
//

{ .mmi
      nop.m 999
(p0)  addl           GR_Table_Ptr1   = @ltoff(Constants_exp_64_P), gp
      nop.i 999
}
;;

{ .mmi
      ld8 GR_Table_Ptr1 = [GR_Table_Ptr1]
      nop.m 999
      nop.i 999
}
;;

{ .mmf
	nop.m 999
(p0)   ldfe FR_P_6 = [GR_Table_Ptr1],16 
//
//     Return  
//
(p0)   fadd.s1 FR_r = FR_X,f0 ;; 
}

{ .mmi
      nop.m 999
(p0)  addl           GR_Table_Ptr   = @ltoff(Constants_exp_64_Exponents#), gp
      nop.i 999
}
;;

{ .mmi
      ld8 GR_Table_Ptr = [GR_Table_Ptr]
(p0)  ldfe FR_P_5 = [GR_Table_Ptr1],16 
      nop.i 999
}
;;

//
//     Is input very small? 
//     Load P_5
//
{ .mii
(p0)   ldfe FR_P_4 = [GR_Table_Ptr1],16 
(p0)   add GR_Table_Ptr = 0x040,GR_Table_Ptr ;;  
(p0)   shladd GR_Table_Ptr = GR_Expo_Range,3,GR_Table_Ptr ;;  
}
{ .mmb
(p0)   ldfe FR_P_3 = [GR_Table_Ptr1],16 
//
//     Adjust ptr.
//
(p0)   ld8  GR_vsm_expo = [GR_Table_Ptr],0
	nop.b 999 ;;
}
{ .mfi
	nop.m 999
//
//     r = X (don't seem to need X_Cor) 
//     Load the threshold exponents
//
(p0)   fmpy.s1 FR_rsq = FR_r, FR_r 
	nop.i 999 ;;
}
//
//     Load the negative integer
//     Load P_5
//
{ .mfi
(p0)   cmp.lt.unc  p12, p0 =  GR_Expo_X, GR_vsm_expo 
	nop.f 999
	nop.i 999 ;;
}
{ .mfb
	nop.m 999
//
//     rsq = r * r
//     Offset into exponents
//
(p0)   fmpy.s1 FR_r4 = FR_rsq, FR_rsq 
(p12)  br.cond.spnt L(EXPL_VERY_SMALL) ;; 
}
{ .mfi
(p0)   ldfe FR_P_2 = [GR_Table_Ptr1],16 
//
//     Load p4,p3,p2,p1
//
(p0)   fma.s1 FR_poly_lo = FR_P_6, FR_r, FR_P_5 
//
//     Y_lo = r4 * poly_lo + poly_hi
//     Scale = 1.0
//
(p0)   add GR_temp = 0x1,r0 ;;  
}
{ .mmf
	nop.m 999
(p0)   ldfe FR_P_1 = [GR_Table_Ptr1],0 
(p0)   mov FR_Scale = f1 
}
//
//     Begin creating lsb to perturb final result
//
{ .mfi
(p0)   setf.sig FR_temp = GR_temp 
(p0)   mov FR_Y_hi = f1 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     poly_lo = p_5 + p_6 * r
//     poly_hi = p_1 + p_2 * r
//
(p0)   fma.s1 FR_poly_lo = FR_poly_lo, FR_r, FR_P_4 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     poly_lo = p_4 + poly_lo * r
//     poly_hi = r + poly_hi * rsq
//
(p0)   fma.s1 FR_poly_lo = FR_poly_lo, FR_r, FR_P_3 
	nop.i 999
}
{ .mfi
	nop.m 999
(p0)   fma.s1 FR_poly_hi = FR_P_2, FR_r, FR_P_1 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p0)   fma.s1 FR_poly_hi = FR_poly_hi, FR_rsq, FR_r 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//  
//     poly_lo = p_3 + poly_lo * r
//     Y_hi = 1, always
//  
(p0)   fma.s1 FR_Y_lo = FR_poly_lo, FR_r4, FR_poly_hi 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     Set lsb in fp register
// 
(p0)   for FR_temp = FR_Y_lo,FR_temp 
	nop.i 999 ;;
}
{ .mfb
	nop.m 999
//
//     Toggle on last bit of Y_lo
//
(p0)   fmerge.se FR_Y_lo = FR_Y_lo,FR_temp 
//
//     Set lsb of Y_lo to 1
//
(p0)   br.cond.sptk L(EXPL_RETURN) ;; 
}
L(EXPL_VERY_SMALL): 
{ .mfi
	nop.m 999
(p0)   mov FR_Y_lo = FR_r 
(p0)   cmp.eq.unc  p15, p0 =  r0, r0 
}
{ .mfi
       nop.m 999
(p0)   mov FR_Scale = f1
       nop.i 999
};;
{ .mfb
	nop.m 999
(p0)   mov FR_Y_hi = f1 
//
//     If flag_not_1, 
//     Y_hi = 1.0 
//     Y_lo = X + X_cor
//     PR_Safe = true
//
(p0)   br.cond.sptk L(EXPL_RETURN) ;; 
}
L(EXPL_HUGE): 
{ .mfi
	nop.m 999
//
//     Return for flag=2 
//
(p0)   fcmp.gt.unc.s1 p12, p13 =  FR_X, f0 
(p0)   cmp.eq.unc  p14, p15 =  r0, r0 ;; 
}
{ .mlx
	nop.m 999
//
//     Set Safe to false
//     Is x > 0
//
(p12)  movl GR_Mask = 0x15DC0 ;; 
}
{ .mlx
(p12)  setf.exp FR_Y_hi = GR_Mask 
(p13)  movl GR_Mask = 0xA240 ;; 
}
{ .mlx
(p13)  setf.exp FR_Y_hi = GR_Mask 
//     
//     x > 0: Create mask for Y_hi = 2**(24,000) 
//     x <= 0: Create mask for Y_hi = 2**(-24,000) 
//
(p13)  movl GR_temp = 0xA1DC ;; 
}
{ .mfi
(p13)  setf.exp FR_Y_lo = GR_temp 
//
//     x < =0: Create mask for 2**(-24,100)
//     x <= 0: Y_lo = w**(-24,100)
//
(p12)  mov FR_Y_lo = f1 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
(p12)  mov FR_Scale =  FR_Y_hi 
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     x > 0: Y_lo = 1.0
//     x > 0: Scale = 2**(24,000) 
//
(p13)  mov FR_Scale = FR_Y_hi 
	nop.i 999 ;;
}
L(EXPL_RETURN): 
{ .mfi
	nop.m 999
//
//     Scale = 2**(24,000)
//
//
//     exp(y *ln(x)) almost complete 
//     FR_Scale is Scale
//     f34 is Z_hi 
//     f35 is Z_lo 
//
(p0)   fmpy.s1 FR_Sgn = FR_Scale, FR_Sgn  
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     sgn * scale 
//
(p0)   fmpy.s1 FR_Y_lo = FR_Y_lo,FR_Sgn   
	nop.i 999 ;;
}
{ .mfb
	nop.m 999
//
//     Z_lo * (sgn * scale) 
//
(p0)   fma.s0 FR_Result = FR_Y_hi, FR_Sgn, FR_Y_lo  
//
//     Z_hi * (sgn * scale)  + Z_lo
//
(p15)  br.cond.sptk L(POWL_64_RETURN) ;;
}
{ .mfi
	nop.m 999
(p0)   fsetc.s3 0x7F,0x01
	nop.i 999
}
{ .mlx
	nop.m 999
//
//     Z_hi * (sgn * scale)  + Z_lo with wre & td
//     Z_hi * (sgn * scale)  + Z_lo with fz  & td
//
(p0)   movl GR_T1_ptr = 0x00000000013FFF ;;
}
{ .mfi
	nop.m 999
(p0)   fma.s3  FR_Result_small = FR_Y_hi, FR_Sgn, FR_Y_lo  
	nop.i 999
}
{ .mfi
	nop.m 999
(p0)   fsetc.s3 0x7F,0x40
	nop.i 999 ;;
}
{ .mfi
	nop.m 999
//
//     Return if no danger of over of underflow.
//
(p0)   fsetc.s2 0x7F,0x42
	nop.i 999;;
}
{ .mfi
	nop.m 999
//
//     S0 user supplied status
//     S2 user supplied status + WRE + TD  (Overflows)
//     S3 user supplied status + FZ + TD   (Underflows)
//
(p0)   fma.s2  FR_Result_big = FR_Y_hi, FR_Sgn, FR_Y_lo  
	nop.i 999 ;;