s_cosl.s

glibc 2.9,最新版的C语言库函数

}
;;

{ .mmi
(p9)  ldfe FR_S_4 = [GR_ad_se], -16    // Load S_4 if i_1=0
(p10) ldfe FR_C_4 = [GR_ad_ce], -16    // Load C_4 if i_1=1
      nop.i 999
}
;;

SINCOSL_SMALL_R_0:
// Entry point for 2^-3 < |x| < pi/4
.pred.rel "mutex",p9,p10
SINCOSL_SMALL_R_1:
// Entry point for pi/4 < |x| < 2^24 and |r| < 2^-3
.pred.rel "mutex",p9,p10
{ .mfi
(p9)  ldfe FR_S_3 = [GR_ad_se], -16    // Load S_3 if i_1=0
      fma.s1 FR_Z = FR_rsq, FR_rsq, f0 // Z = rsq * rsq
      nop.i 999
}
{ .mfi
(p10) ldfe FR_C_3 = [GR_ad_ce], -16    // Load C_3 if i_1=1
(p10) fnma.s1 FR_c = FR_c, FR_r, f0    // c = -c * r if i_1=0
      nop.i 999
}
;;

{ .mmf
(p9)  ldfe FR_S_2 = [GR_ad_se], -16    // Load S_2 if i_1=0
(p10) ldfe FR_C_2 = [GR_ad_ce], -16    // Load C_2 if i_1=1
(p10) fmerge.s FR_r = f1, f1
}
;;

{ .mmi
(p9)  ldfe FR_S_1 = [GR_ad_se], -16    // Load S_1 if i_1=0
(p10) ldfe FR_C_1 = [GR_ad_ce], -16    // Load C_1 if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_Z = FR_Z, FR_r, f0     // Z = Z * r if i_1=0
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_lo = FR_rsq, FR_S_5, FR_S_4 // poly_lo=rsq*S_5+S_4 if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly_lo = FR_rsq, FR_C_5, FR_C_4 // poly_lo=rsq*C_5+C_4 if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_hi = FR_rsq, FR_S_2, FR_S_1 // poly_hi=rsq*S_2+S_1 if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly_hi = FR_rsq, FR_C_2, FR_C_1 // poly_hi=rsq*C_2+C_1 if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
      fma.s1 FR_Z = FR_Z, FR_rsq, f0             // Z = Z * rsq
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_lo = FR_rsq, FR_poly_lo, FR_S_3 // p_lo=p_lo*rsq+S_3, i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly_lo = FR_rsq, FR_poly_lo, FR_C_3 // p_lo=p_lo*rsq+C_3, i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s0 FR_inexact = FR_S_4, FR_S_4, f0     // Dummy op to set inexact
      tbit.z p11,p12 = GR_N_Inc, 1     // p11 if i_0=0, N mod 4 = 0,2
                                       // p12 if i_0=1, N mod 4 = 1,3
}
{ .mfi
      nop.m 999
(p10) fma.s0 FR_inexact = FR_C_1, FR_C_1, f0     // Dummy op to set inexact
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_hi = FR_poly_hi, FR_rsq, f0     // p_hi=p_hi*rsq if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly_hi = FR_poly_hi, FR_rsq, f0     // p_hi=p_hi*rsq if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
      fma.s1 FR_poly = FR_Z, FR_poly_lo, FR_c        // poly=Z*poly_lo+c
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly_hi = FR_r, FR_poly_hi, f0       // p_hi=r*p_hi if i_1=0
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p12) fms.s1 FR_r = f0, f1, FR_r                     // r = -r if i_0=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
      fma.s1 FR_poly = FR_poly, f1, FR_poly_hi       // poly=poly+poly_hi
      nop.i 999
}
;;

//
//      if (i_0 == 0) Result = r + poly
//      if (i_0 != 0) Result = r - poly
//
{ .mfi
      nop.m 999
(p11) fma.s0 FR_Result = FR_r, f1, FR_poly
      nop.i 999
}
{ .mfb
      nop.m 999
(p12) fms.s0 FR_Result = FR_r, f1, FR_poly
      br.ret.sptk   b0                               // Exit for |r| < 2^-3
}
;;


SINCOSL_NORMAL_R:
//
// Here if 2^-3 <= |r| < pi/4
// THIS IS THE MAIN PATH
//
// Enter with r, c, and N_Inc having been computed
//
{ .mfi
      ldfe FR_PP_6 = [GR_ad_pp], 16    // Load PP_6
      fma.s1 FR_rsq = FR_r, FR_r, f0   // rsq = r * r
      tbit.z p9,p10 = GR_N_Inc, 0      // p9  if i_1=0, N mod 4 = 0,1
                                       // p10 if i_1=1, N mod 4 = 2,3
}
{ .mfi
      ldfe FR_QQ_6 = [GR_ad_qq], 16    // Load QQ_6
      nop.f 999
      nop.i 999
}
;;

{ .mmi
(p9)  ldfe FR_PP_5 = [GR_ad_pp], 16    // Load PP_5 if i_1=0
(p10) ldfe FR_QQ_5 = [GR_ad_qq], 16    // Load QQ_5 if i_1=1
      nop.i 999
}
;;

SINCOSL_NORMAL_R_0:
// Entry for 2^-3 < |x| < pi/4
.pred.rel "mutex",p9,p10
{ .mmf
(p9)  ldfe FR_C_1 = [GR_ad_pp], 16     // Load C_1 if i_1=0
(p10) ldfe FR_S_1 = [GR_ad_qq], 16     // Load S_1 if i_1=1
      frcpa.s1 FR_r_hi, p6 = f1, FR_r  // r_hi = frcpa(r)
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly = FR_rsq, FR_PP_8, FR_PP_7 // poly = rsq*PP_8+PP_7 if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly = FR_rsq, FR_QQ_8, FR_QQ_7 // poly = rsq*QQ_8+QQ_7 if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
      fma.s1 FR_r_cubed = FR_r, FR_rsq, f0  // rcubed = r * rsq
      nop.i 999
}
;;


SINCOSL_NORMAL_R_1:
// Entry for pi/4 <= |x| < 2^24
.pred.rel "mutex",p9,p10
{ .mmf
(p9)  ldfe FR_PP_1 = [GR_ad_pp], 16             // Load PP_1_hi if i_1=0
(p10) ldfe FR_QQ_1 = [GR_ad_qq], 16             // Load QQ_1    if i_1=1
      frcpa.s1 FR_r_hi, p6 = f1, FR_r_hi        // r_hi = frpca(frcpa(r))
}
;;

{ .mfi
(p9)  ldfe FR_PP_4 = [GR_ad_pp], 16             // Load PP_4 if i_1=0
(p9)  fma.s1 FR_poly = FR_rsq, FR_poly, FR_PP_6 // poly = rsq*poly+PP_6 if i_1=0
      nop.i 999
}
{ .mfi
(p10) ldfe FR_QQ_4 = [GR_ad_qq], 16             // Load QQ_4 if i_1=1
(p10) fma.s1 FR_poly = FR_rsq, FR_poly, FR_QQ_6 // poly = rsq*poly+QQ_6 if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_corr = FR_C_1, FR_rsq, f0       // corr = C_1 * rsq if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_corr = FR_S_1, FR_r_cubed, FR_r // corr = S_1 * r^3 + r if i_1=1
      nop.i 999
}
;;

{ .mfi
(p9)  ldfe FR_PP_3 = [GR_ad_pp], 16             // Load PP_3 if i_1=0
      fma.s1 FR_r_hi_sq = FR_r_hi, FR_r_hi, f0  // r_hi_sq = r_hi * r_hi
      nop.i 999
}
{ .mfi
(p10) ldfe FR_QQ_3 = [GR_ad_qq], 16             // Load QQ_3 if i_1=1
      fms.s1 FR_r_lo = FR_r, f1, FR_r_hi        // r_lo = r - r_hi
      nop.i 999
}
;;

{ .mfi
(p9)  ldfe FR_PP_2 = [GR_ad_pp], 16             // Load PP_2 if i_1=0
(p9)  fma.s1 FR_poly = FR_rsq, FR_poly, FR_PP_5 // poly = rsq*poly+PP_5 if i_1=0
      nop.i 999
}
{ .mfi
(p10) ldfe FR_QQ_2 = [GR_ad_qq], 16             // Load QQ_2 if i_1=1
(p10) fma.s1 FR_poly = FR_rsq, FR_poly, FR_QQ_5 // poly = rsq*poly+QQ_5 if i_1=1
      nop.i 999
}
;;

{ .mfi
(p9)  ldfe FR_PP_1_lo = [GR_ad_pp], 16          // Load PP_1_lo if i_1=0
(p9)  fma.s1 FR_corr = FR_corr, FR_c, FR_c      // corr = corr * c + c if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fnma.s1 FR_corr = FR_corr, FR_c, f0       // corr = -corr * c if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_U_lo = FR_r, FR_r_hi, FR_r_hi_sq // U_lo = r*r_hi+r_hi_sq, i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_U_lo = FR_r_hi, f1, FR_r        // U_lo = r_hi + r if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_U_hi = FR_r_hi, FR_r_hi_sq, f0  // U_hi = r_hi*r_hi_sq if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_U_hi = FR_QQ_1, FR_r_hi_sq, f1  // U_hi = QQ_1*r_hi_sq+1, i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly = FR_rsq, FR_poly, FR_PP_4 // poly = poly*rsq+PP_4 if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly = FR_rsq, FR_poly, FR_QQ_4 // poly = poly*rsq+QQ_4 if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_U_lo = FR_r, FR_r, FR_U_lo      // U_lo = r * r + U_lo if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_U_lo = FR_r_lo, FR_U_lo, f0     // U_lo = r_lo * U_lo if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_U_hi = FR_PP_1, FR_U_hi, f0     // U_hi = PP_1 * U_hi if i_1=0
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly = FR_rsq, FR_poly, FR_PP_3 // poly = poly*rsq+PP_3 if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly = FR_rsq, FR_poly, FR_QQ_3 // poly = poly*rsq+QQ_3 if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_U_lo = FR_r_lo, FR_U_lo, f0     // U_lo = r_lo * U_lo if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_U_lo = FR_QQ_1,FR_U_lo, f0      // U_lo = QQ_1 * U_lo if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_U_hi = FR_r, f1, FR_U_hi        // U_hi = r + U_hi if i_1=0
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly = FR_rsq, FR_poly, FR_PP_2 // poly = poly*rsq+PP_2 if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly = FR_rsq, FR_poly, FR_QQ_2 // poly = poly*rsq+QQ_2 if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_U_lo = FR_PP_1, FR_U_lo, f0     // U_lo = PP_1 * U_lo if i_1=0
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly = FR_rsq, FR_poly, FR_PP_1_lo // poly =poly*rsq+PP1lo i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly = FR_rsq, FR_poly, f0      // poly = poly*rsq if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
      fma.s1 FR_V = FR_U_lo, f1, FR_corr        // V = U_lo + corr
      tbit.z p11,p12 = GR_N_Inc, 1              // p11 if i_0=0, N mod 4 = 0,2
                                                // p12 if i_0=1, N mod 4 = 1,3
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s0 FR_inexact = FR_PP_5, FR_PP_4, f0  // Dummy op to set inexact
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s0 FR_inexact = FR_QQ_5, FR_QQ_5, f0  // Dummy op to set inexact
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p9)  fma.s1 FR_poly = FR_r_cubed, FR_poly, f0  // poly = poly*r^3 if i_1=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p10) fma.s1 FR_poly = FR_rsq, FR_poly, f0      // poly = poly*rsq if i_1=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p11) fma.s1 FR_tmp_result = f0, f1, f1// tmp_result=+1.0 if i_0=0
      nop.i 999
}
{ .mfi
      nop.m 999
(p12) fms.s1 FR_tmp_result = f0, f1, f1// tmp_result=-1.0 if i_0=1
      nop.i 999
}
;;

{ .mfi
      nop.m 999
      fma.s1 FR_V = FR_poly, f1, FR_V           // V = poly + V
      nop.i 999
}
;;

// If i_0 = 0  Result =  U_hi + V
// If i_0 = 1  Result = -U_hi - V
{ .mfi
        nop.m 999
(p11)        fma.s0 FR_Result = FR_tmp_result, FR_U_hi, FR_V
        nop.i 999
}
{ .mfb
        nop.m 999
(p12)        fms.s0 FR_Result = FR_tmp_result, FR_U_hi, FR_V
        br.ret.sptk   b0                     // Exit for 2^-3 <= |r| < pi/4
}
;;

SINCOSL_ZERO:
// Here if x = 0
{ .mfi
      cmp.eq.unc p6, p7 = 0x1, GR_Sin_or_Cos
      nop.f 999
      nop.i 999
}
;;

{ .mfi
      nop.m 999
(p7)  fmerge.s FR_Result = FR_Input_X, FR_Input_X // If sin, result = input
      nop.i 999
}
{ .mfb
      nop.m 999
(p6)  fma.s0 FR_Result = f1, f1, f0    // If cos, result=1.0
      br.ret.sptk   b0                  // Exit for x=0
}
;;


SINCOSL_DENORMAL:
{ .mmb
      getf.exp GR_signexp_x = FR_norm_x   // Get sign and exponent of x
      nop.m 999
      br.cond.sptk  SINCOSL_COMMON        // Return to common code
}
;;

SINCOSL_SPECIAL:
{ .mfb
        nop.m 999
//
//      Path for Arg = +/- QNaN, SNaN, Inf
//      Invalid can be raised. SNaNs
//      become QNaNs
//
        fmpy.s0 FR_Result = FR_Input_X, f0
        br.ret.sptk   b0 ;;
}

GLOBAL_IEEE754_END(cosl)

// *******************************************************************
// *******************************************************************
// *******************************************************************
//
//     Special Code to handle very large argument case.
//     Call int __libm_pi_by_2_reduce(x,r,c) for |arguments| >= 2**63
//     The interface is custom:
//       On input:
//         (Arg or x) is in f8
//       On output:
//         r is in f8
//         c is in f9
//         N is in r8
//     Be sure to allocate at least 2 GP registers as output registers for
//     __libm_pi_by_2_reduce.  This routine uses r59-60. These are used as
//     scratch registers within the __libm_pi_by_2_reduce routine (for speed).
//
//     We know also that __libm_pi_by_2_reduce preserves f10-15, f71-127.  We
//     use this to eliminate save/restore of key fp registers in this calling
//     function.
//
// *******************************************************************
// *******************************************************************
// *******************************************************************

LOCAL_LIBM_ENTRY(__libm_callout)
SINCOSL_ARG_TOO_LARGE:
.prologue
{ .mfi
        nop.f 0
.save   ar.pfs,GR_SAVE_PFS
        mov  GR_SAVE_PFS=ar.pfs                 // Save ar.pfs
};;

{ .mmi
        setf.exp FR_Two_to_M3 = GR_exp_2_to_m3  // Form 2^-3
        mov GR_SAVE_GP=gp                       // Save gp
.save   b0, GR_SAVE_B0
        mov GR_SAVE_B0=b0                       // Save b0
};;

.body
//
//     Call argument reduction with x in f8
//     Returns with N in r8, r in f8, c in f9
//     Assumes f71-127 are preserved across the call
//
{ .mib
        setf.exp FR_Neg_Two_to_M3 = GR_exp_m2_to_m3 // Form -(2^-3)
        nop.i 0
        br.call.sptk b0=__libm_pi_by_2_reduce#
};;

{ .mfi
        add   GR_N_Inc = GR_Sin_or_Cos,r8
        fcmp.lt.unc.s1        p6, p0 = FR_r, FR_Two_to_M3
        mov   b0 = GR_SAVE_B0                  // Restore return address
};;

{ .mfi
        mov   gp = GR_SAVE_GP                  // Restore gp
(p6)    fcmp.gt.unc.s1        p6, p0 = FR_r, FR_Neg_Two_to_M3
        mov   ar.pfs = GR_SAVE_PFS             // Restore ar.pfs
};;

{ .mbb
        nop.m 999
(p6)    br.cond.spnt SINCOSL_SMALL_R     // Branch if |r|< 2^-3 for |x| >= 2^63
        br.cond.sptk SINCOSL_NORMAL_R    // Branch if |r|>=2^-3 for |x| >= 2^63
};;

LOCAL_LIBM_END(__libm_callout)
.type   __libm_pi_by_2_reduce#,@function
.global __libm_pi_by_2_reduce#