libm_sincosl.s

glibc 2.9,最新版的C语言库函数

}
{ .mfb
      cmp.eq  p15, p0 = 0x1, GR_Cis
(p14) fms.s0 FR_ResultC = FR_FirstS, f1, FR_polyS
(p15) br.ret.sptk               b0
};;


{ .mmb       // exit for sincosl
      stfe  [sincos_pResSin] =  FR_ResultS
      stfe  [sincos_pResCos] =  FR_ResultC
      br.ret.sptk               b0
};;






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
      nop.i 0
}
{ .mfi
      ldfe FR_QQ_6 = [GR_ad_qq], 16    // Load QQ_6
      nop.f 0
      nop.i 0
};;

{ .mmi
      ldfe FR_PP_5 = [GR_ad_pp], 16    // Load PP_5
      ldfe FR_QQ_5 = [GR_ad_qq], 16    // Load QQ_5
      nop.i 0
};;



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

{ .mfi
      nop.m 0
      fma.s1 FR_polyS = FR_rsq, FR_PP_8, FR_PP_7 // poly = rsq*PP_8+PP_7
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_polyC = FR_rsq, FR_QQ_8, FR_QQ_7 // poly = rsq*QQ_8+QQ_7
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_r_cubed = FR_r, FR_rsq, f0  // rcubed = r * rsq
      nop.i 0
};;


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

{ .mfi
      ldfe FR_PP_4 = [GR_ad_pp], 16             // Load PP_4
      fma.s1 FR_polyS = FR_rsq, FR_polyS, FR_PP_6 // poly = rsq*poly+PP_6
      and GR_N_SinCos = 0x1, GR_N_Inc
}
{ .mfi
      ldfe FR_QQ_4 = [GR_ad_qq], 16             // Load QQ_4
      fma.s1 FR_polyC = FR_rsq, FR_polyC, FR_QQ_6 // poly = rsq*poly+QQ_6
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_corrS = FR_C_1, FR_rsq, f0       // corr = C_1 * rsq
      sub GR_N_SignS =  GR_N_Inc, GR_N_SinCos
}
{ .mfi
      nop.m 0
      fma.s1 FR_corrC = FR_S_1, FR_r_cubed, FR_r // corr = S_1 * r^3 + r
      add GR_N_SignC =  GR_N_Inc, GR_N_SinCos
};;

{ .mfi
      ldfe FR_PP_3 = [GR_ad_pp], 16             // Load PP_3
      fma.s1 FR_r_hi_sq = FR_r_hi, FR_r_hi, f0  // r_hi_sq = r_hi * r_hi
      tbit.z p7,p11       = GR_N_Inc, 0
}
{ .mfi
      ldfe FR_QQ_3 = [GR_ad_qq], 16             // Load QQ_3
      fms.s1 FR_r_lo = FR_r, f1, FR_r_hi        // r_lo = r - r_hi
      nop.i 0
};;

{ .mfi
      ldfe FR_PP_2 = [GR_ad_pp], 16             // Load PP_2
      fma.s1 FR_polyS = FR_rsq, FR_polyS, FR_PP_5 // poly = rsq*poly+PP_5
(p7)  tbit.z.unc p9,p10   = GR_N_SignC, 1
}
{ .mfi
      ldfe FR_QQ_2 = [GR_ad_qq], 16             // Load QQ_2
      fma.s1 FR_polyC = FR_rsq, FR_polyC, FR_QQ_5 // poly = rsq*poly+QQ_5
      nop.i 0
};;

{ .mfi
      ldfe FR_PP_1_lo = [GR_ad_pp], 16          // Load PP_1_lo
      fma.s1 FR_corrS = FR_corrS, FR_c, FR_c      // corr = corr * c + c
(p7)  tbit.z.unc p7,p8    = GR_N_SignS, 1
}
{ .mfi
      nop.m 0
      fnma.s1 FR_corrC = FR_corrC, FR_c, f0       // corr = -corr * c
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_U_loS = FR_r, FR_r_hi, FR_r_hi_sq // U_lo = r*r_hi+r_hi_sq
(p11) tbit.z.unc p13,p14  = GR_N_SignC, 1
}
{ .mfi
      nop.m 0
      fma.s1 FR_U_loC = FR_r_hi, f1, FR_r        // U_lo = r_hi + r
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_U_hiS = FR_r_hi, FR_r_hi_sq, f0  // U_hi = r_hi*r_hi_sq
(p11) tbit.z.unc p11,p12  = GR_N_SignS, 1
}
{ .mfi
      nop.m 0
      fma.s1 FR_U_hiC = FR_QQ_1, FR_r_hi_sq, f1  // U_hi = QQ_1*r_hi_sq+1
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_polyS = FR_rsq, FR_polyS, FR_PP_4 // poly = poly*rsq+PP_4
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_polyC = FR_rsq, FR_polyC, FR_QQ_4 // poly = poly*rsq+QQ_4
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_U_loS = FR_r, FR_r, FR_U_loS      // U_lo = r * r + U_lo
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_U_loC = FR_r_lo, FR_U_loC, f0     // U_lo = r_lo * U_lo
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_U_hiS = FR_PP_1, FR_U_hiS, f0     // U_hi = PP_1 * U_hi
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_polyS = FR_rsq, FR_polyS, FR_PP_3 // poly = poly*rsq+PP_3
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_polyC = FR_rsq, FR_polyC, FR_QQ_3 // poly = poly*rsq+QQ_3
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_U_loS = FR_r_lo, FR_U_loS, f0     // U_lo = r_lo * U_lo
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_U_loC = FR_QQ_1,FR_U_loC, f0      // U_lo = QQ_1 * U_lo
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_U_hiS = FR_r, f1, FR_U_hiS        // U_hi = r + U_hi
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_polyS = FR_rsq, FR_polyS, FR_PP_2 // poly = poly*rsq+PP_2
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_polyC = FR_rsq, FR_polyC, FR_QQ_2 // poly = poly*rsq+QQ_2
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_U_loS = FR_PP_1, FR_U_loS, f0     // U_lo = PP_1 * U_lo
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s1 FR_polyS = FR_rsq, FR_polyS, FR_PP_1_lo // poly =poly*rsq+PP1lo
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_polyC = FR_rsq, FR_polyC, f0      // poly = poly*rsq
      nop.i 0
};;


.pred.rel "mutex",p8,p14
{ .mfi
      nop.m 0
(p8)  fms.s0 FR_U_hiS   = f1, f0, FR_U_hiS
      nop.i 0
}
{ .mfi
      nop.m 0
(p14) fms.s0 FR_U_hiS   = f1, f0, FR_U_hiS
      nop.i 0
};;

.pred.rel "mutex",p10,p12
{ .mfi
      nop.m 0
(p10) fms.s0 FR_U_hiC   = f1, f0, FR_U_hiC
      nop.i 0
}
{ .mfi
      nop.m 0
(p12) fms.s0 FR_U_hiC   = f1, f0, FR_U_hiC
      nop.i 0
};;


{ .mfi
      nop.m 0
      fma.s1 FR_VS = FR_U_loS, f1, FR_corrS        // V = U_lo + corr
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_VC = FR_U_loC, f1, FR_corrC        // V = U_lo + corr
      nop.i 0
};;

{ .mfi
      nop.m 0
      fma.s0 FR_inexact = FR_PP_5, FR_PP_4, f0  // Dummy op to set inexact
      nop.i 0
};;


{ .mfi
      nop.m 0
      fma.s1 FR_polyS = FR_r_cubed, FR_polyS, f0  // poly = poly*r^3
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_polyC = FR_rsq, FR_polyC, f0      // poly = poly*rsq
      nop.i 0
};;


{ .mfi
      nop.m 0
      fma.s1 FR_VS = FR_polyS, f1, FR_VS           // V = poly + V
      nop.i 0
}
{ .mfi
      nop.m 0
      fma.s1 FR_VC = FR_polyC, f1, FR_VC           // V = poly + V
      nop.i 0
};;



.pred.rel "mutex",p7,p8
{ .mfi
      nop.m 0
(p7)  fma.s0 FR_ResultS = FR_U_hiS, f1, FR_VS
      nop.i 0
}
{ .mfi
      nop.m 0
(p8)  fms.s0 FR_ResultS = FR_U_hiS, f1, FR_VS
      nop.i 0
};;

.pred.rel "mutex",p9,p10
{ .mfi
      nop.m 0
(p9)  fma.s0 FR_ResultC = FR_U_hiC, f1, FR_VC
      nop.i 0
}
{ .mfi
      nop.m 0
(p10) fms.s0 FR_ResultC = FR_U_hiC, f1, FR_VC
      nop.i 0
};;



.pred.rel "mutex",p11,p12
{ .mfi
      nop.m 0
(p11) fma.s0 FR_ResultS = FR_U_hiC, f1, FR_VC
      nop.i 0
}
{ .mfi
      nop.m 0
(p12) fms.s0 FR_ResultS = FR_U_hiC, f1, FR_VC
      nop.i 0
};;

.pred.rel "mutex",p13,p14
{ .mfi
      nop.m 0
(p13) fma.s0 FR_ResultC = FR_U_hiS, f1, FR_VS
      nop.i 0
}
{ .mfb
      cmp.eq  p15, p0 = 0x1, GR_Cis
(p14) fms.s0 FR_ResultC = FR_U_hiS, f1, FR_VS
(p15) br.ret.sptk               b0
};;

{ .mmb       // exit for sincosl
      stfe  [sincos_pResSin] =  FR_ResultS
      stfe  [sincos_pResCos] =  FR_ResultC
      br.ret.sptk               b0
};;





SINCOSL_ZERO:

{ .mfi
      nop.m 0
      fmerge.s FR_ResultS = FR_Input_X, FR_Input_X // If sin, result = input
      nop.i 0
}
{ .mfb
      cmp.eq  p15, p0 = 0x1, GR_Cis
      fma.s0 FR_ResultC = f1, f1, f0    // If cos, result=1.0
(p15) br.ret.sptk               b0
};;

{ .mmb       // exit for sincosl
      stfe  [sincos_pResSin] =  FR_ResultS
      stfe  [sincos_pResCos] =  FR_ResultC
      br.ret.sptk               b0
};;


SINCOSL_DENORMAL:
{ .mmb
      getf.exp GR_signexp_x = FR_norm_x   // Get sign and exponent of x
      nop.m 999
      br.cond.sptk  SINCOSL_COMMON2        // Return to common code
}
;;


SINCOSL_SPECIAL:
//
//    Path for Arg = +/- QNaN, SNaN, Inf
//    Invalid can be raised. SNaNs
//    become QNaNs
//
{ .mfi
      cmp.eq  p15, p0 = 0x1, GR_Cis
      fmpy.s0 FR_ResultS = FR_Input_X, f0
      nop.i 0
}
{ .mfb
      nop.m 0
      fmpy.s0 FR_ResultC = FR_Input_X, f0
(p15) br.ret.sptk               b0
};;

{ .mmb       // exit for sincosl
      stfe  [sincos_pResSin] =  FR_ResultS
      stfe  [sincos_pResCos] =  FR_ResultC
      br.ret.sptk               b0
};;

GLOBAL_LIBM_END(__libm_sincosl)


// *******************************************************************
// *******************************************************************
// *******************************************************************
//
//     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 r62-63. 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
        mov   GR_N_Inc = 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 0
(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#