s_log1pl.s

glibc 2.9,最新版的C语言库函数

;;

{ .mmi
      ld4 GR_Z_2 = [GR_ad_z_2]                // Load Z_2
      shladd GR_ad_tbl_2 = GR_Index2, 4, GR_ad_tbl_2  // Point to G_2
      and GR_M = GR_exp_mask, GR_M            // Get exponent of w = x
}
;;

{ .mmi
      ldfps  FR_G2, FR_H2 = [GR_ad_tbl_2],8   // Load G_2, H_2
      cmp.lt  p8, p9 =  GR_M, GR_exp_2tom7    // Test |x| < 2^-7
      cmp.lt  p7, p0 =  GR_M, GR_exp_2tom80   // Test |x| < 2^-80
}
;;

// Small path is separate code
//  p7 is for the small path: |x| < 2^-80
// near1 and regular paths are merged.
//  p8 is for the near1 path: |x| < 2^-7
//  p9 is for regular path:   |x| >= 2^-7

{ .mfi
      ldfd  FR_h2 = [GR_ad_tbl_2]             // Load h_2
      nop.f 999
      nop.i 999
}
{ .mfb
(p9)  setf.exp FR_2_to_minus_N = GR_minus_N   // Form 2^(-N)
(p7)  fnma.s0  f8 = FR_X_Prime, FR_X_Prime, FR_X_Prime // Result x - x*x
(p7)  br.ret.spnt  b0                         // Branch if |x| < 2^-80
}
;;

{ .mmi
(p8)  ldfe FR_P7 = [GR_ad_p],16               // Load P_7 for near1 path
(p8)  ldfe FR_P4 = [GR_ad_p2],16              // Load P_4 for near1 path
(p9)  pmpyshr2.u GR_X_2 = GR_X_1,GR_Z_2,15    // Get bits 30-15 of X_1 * Z_2
}
;;

//
//    For performance, don't use result of pmpyshr2.u for 4 cycles.
//
{ .mmf
(p8)  ldfe FR_P6 = [GR_ad_p],16               // Load P_6 for near1 path
(p8)  ldfe FR_P3 = [GR_ad_p2],16              // Load P_3 for near1 path
(p9)  fma.s1  FR_S_lo = FR_S_lo, f1, FR_BB    // S_lo = S_lo + BB
}
;;

{ .mmf
(p8)  ldfe FR_P5 = [GR_ad_p],16               // Load P_5 for near1 path
(p8)  ldfe FR_P2 = [GR_ad_p2],16              // Load P_2 for near1 path
(p8)  fmpy.s1 FR_wsq = FR_W, FR_W             // wsq = w * w for near1 path
}
;;

{ .mmi
(p8)  ldfe FR_P1 = [GR_ad_p2],16 ;;           // Load P_1 for near1 path
      nop.m 999
(p9)  extr.u GR_Index3 = GR_X_2, 1, 5         // Extract bits 1-5 of X_2
}
;;

{ .mfi
(p9)  shladd GR_ad_tbl_3 = GR_Index3, 4, GR_ad_tbl_3  // Point to G_3
(p9)  fcvt.xf FR_float_N = FR_float_N
      nop.i 999
}
;;

{ .mfi
(p9)  ldfps  FR_G3, FR_H3 = [GR_ad_tbl_3],8   // Load G_3, H_3
      nop.f 999
      nop.i 999
}
;;

{ .mfi
(p9)  ldfd  FR_h3 = [GR_ad_tbl_3]             // Load h_3
(p9)  fmpy.s1 FR_G = FR_G, FR_G2              // G = G_1 * G_2
      nop.i 999
}
{ .mfi
      nop.m 999
(p9)  fadd.s1 FR_H = FR_H, FR_H2              // H = H_1 + H_2
      nop.i 999
}
;;

{ .mmf
      nop.m 999
      nop.m 999
(p9)  fadd.s1 FR_h = FR_h, FR_h2              // h = h_1 + h_2
}
;;

{ .mfi
      nop.m 999
(p8)  fmpy.s1 FR_w4 = FR_wsq, FR_wsq          // w4 = w^4 for near1 path
      nop.i 999
}
{ .mfi
      nop.m 999
(p8)  fma.s1 FR_p87 = FR_W, FR_P8, FR_P7      // p87 = w * P8 + P7
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1  FR_S_lo = FR_S_lo, FR_2_to_minus_N, f0 // S_lo = S_lo * 2^(-N)
      nop.i 999
}
{ .mfi
      nop.m 999
(p8)  fma.s1 FR_p43 = FR_W, FR_P4, FR_P3      // p43 = w * P4 + P3
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fmpy.s1 FR_G = FR_G, FR_G3              // G = (G_1 * G_2) * G_3
      nop.i 999
}
{ .mfi
      nop.m 999
(p9)  fadd.s1 FR_H = FR_H, FR_H3              // H = (H_1 + H_2) + H_3
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fadd.s1 FR_h = FR_h, FR_h3              // h = (h_1 + h_2) + h_3
      nop.i 999
}
{ .mfi
      nop.m 999
(p8)  fmpy.s1 FR_w6 = FR_w4, FR_wsq           // w6 = w^6 for near1 path
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p8)  fma.s1 FR_p432 = FR_W, FR_p43, FR_P2    // p432 = w * p43 + P2
      nop.i 999
}
{ .mfi
      nop.m 999
(p8)  fma.s1 FR_p876 = FR_W, FR_p87, FR_P6    // p876 = w * p87 + P6
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fms.s1 FR_r = FR_G, FR_S_hi, f1         // r = G * S_hi - 1
      nop.i 999
}
{ .mfi
      nop.m 999
(p9)  fma.s1 FR_Y_hi = FR_float_N, FR_log2_hi, FR_H // Y_hi = N * log2_hi + H
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_h = FR_float_N, FR_log2_lo, FR_h  // h = N * log2_lo + h
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_r = FR_G, FR_S_lo, FR_r        // r = G * S_lo + (G * S_hi - 1)
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p8)  fma.s1 FR_p4321 = FR_W, FR_p432, FR_P1      // p4321 = w * p432 + P1
      nop.i 999
}
{ .mfi
      nop.m 999
(p8)  fma.s1 FR_p8765 = FR_W, FR_p876, FR_P5      // p8765 = w * p876 + P5
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_lo = FR_r, FR_Q4, FR_Q3      // poly_lo = r * Q4 + Q3
      nop.i 999
}
{ .mfi
      nop.m 999
(p9)  fmpy.s1 FR_rsq = FR_r, FR_r                 // rsq = r * r
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p8)  fma.s1 FR_Y_lo = FR_wsq, FR_p4321, f0       // Y_lo = wsq * p4321
      nop.i 999
}
{ .mfi
      nop.m 999
(p8)  fma.s1 FR_Y_hi = FR_W, f1, f0               // Y_hi = w for near1 path
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_lo = FR_poly_lo, FR_r, FR_Q2 // poly_lo = poly_lo * r + Q2
      nop.i 999
}
{ .mfi
      nop.m 999
(p9)  fma.s1 FR_rcub = FR_rsq, FR_r, f0           // rcub = r^3
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p8)  fma.s1 FR_Y_lo = FR_w6, FR_p8765,FR_Y_lo // Y_lo = w6 * p8765 + w2 * p4321
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_hi = FR_Q1, FR_rsq, FR_r     // poly_hi = Q1 * rsq + r
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_lo = FR_poly_lo, FR_rcub, FR_h // poly_lo = poly_lo*r^3 + h
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fadd.s1 FR_Y_lo = FR_poly_hi, FR_poly_lo    // Y_lo = poly_hi + poly_lo 
      nop.i 999
}
;;

// Remainder of code is common for near1 and regular paths
{ .mfb
      nop.m 999
      fadd.s0  f8 = FR_Y_lo,FR_Y_hi               // Result=Y_lo+Y_hi
      br.ret.sptk   b0                       // Common exit for 2^-80 < x < inf
}
;;


// Here if x=-1
LOG1P_EQ_Minus_1: 
//
//    If x=-1 raise divide by zero and return -inf
//  
{ .mfi
      mov   GR_Parameter_TAG = 138
      fsub.s1 FR_Output_X_tmp = f0, f1 
      nop.i 999
}
;;

{ .mfb
      nop.m 999
      frcpa.s0 FR_Output_X_tmp, p8 =  FR_Output_X_tmp, f0 
      br.cond.sptk __libm_error_region
}
;;

LOG1P_special: 
{ .mfi
      nop.m 999
      fclass.m.unc p8, p0 =  FR_Input_X, 0x1E1  // Test for natval, nan, +inf
      nop.i 999
}
;;

//     
//    For SNaN raise invalid and return QNaN.
//    For QNaN raise invalid and return QNaN.
//    For +Inf return +Inf.
//    
{ .mfb
      nop.m 999
(p8)  fmpy.s0 f8 =  FR_Input_X, f1 
(p8)  br.ret.sptk   b0                          // Return for natval, nan, +inf
}
;;

//    
//    For -Inf raise invalid and return QNaN.
//    
{ .mfb
      mov   GR_Parameter_TAG = 139
      fmpy.s0 FR_Output_X_tmp =  FR_Input_X, f0 
      br.cond.sptk __libm_error_region
}
;;


LOG1P_unsupported: 
//    
//    Return generated NaN or other value.
//    
{ .mfb
      nop.m 999
      fmpy.s0 f8 = FR_Input_X, f0 
      br.ret.sptk   b0
}
;;

// Here if -inf < x < -1
LOG1P_LT_Minus_1: 
//     
//    Deal with x < -1 in a special way - raise
//    invalid and produce QNaN indefinite.
//    
{ .mfb
      mov   GR_Parameter_TAG = 139
      frcpa.s0 FR_Output_X_tmp, p8 =  f0, f0
      br.cond.sptk __libm_error_region
}
;;


GLOBAL_IEEE754_END(log1pl)

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
        stfe [GR_Parameter_Y] = FR_Y,16         // Save 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
        stfe [GR_Parameter_X] = FR_X            // Store Parameter 1 on stack
        add   GR_Parameter_RESULT = 0,GR_Parameter_Y
        nop.b 0                                 // Parameter 3 address
}
{ .mib
        stfe [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 999
        nop.m 999
        add   GR_Parameter_RESULT = 48,sp
};;
{ .mmi
        ldfe  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#