e_acos.s

glibc 2.9,最新版的C语言库函数

      fma.s1             fB3 = fB3, fR, fB2
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fnma.s1            fD = fH, fS, fHalf // d0 = 1/2 - H0*S0
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fR8 = fRQuadr, fRQuadr, f0 // R^4
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fB9 = fB9, fR, fB8
      nop.i              0
}
;;
{.mfi
      nop.m              0
      fma.s1             fB12 = fB12, fRSqr, fB11
      nop.i              0
}
{.mfi
      nop.m              0
      fma.s1             fB7 = fB7, fRSqr, fB5
      nop.i              0
}
;;
{.mfi
      nop.m              0
      fma.s1             fB3 = fB3, fRSqr, fB1
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fH = fH, fD, fH // H1 = H0 + H0*d0
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fS = fS, fD, fS // S1 = S0 + S0*d0
      nop.i              0
}
;;
{.mfi
      nop.m              0
(p9)  fma.s1             fCpi = f1, f0, f0 // Cpi = 0 if x > 0
      nop.i              0
}
{ .mfi
      nop.m              0
(p8)  fma.s1             fCpi = fPiBy2, f1, fPiBy2 // Cpi = Pi if x < 0
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fB12 = fB12, fRSqr, fB9
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fB7 = fB7, fRQuadr, fB3
      nop.i              0
}
;;
{.mfi
      nop.m              0
      fnma.s1            fD = fH, fS, fHalf // d1 = 1/2 - H1*S1
      nop.i              0
}
{ .mfi
      nop.m              0
      fnma.s1            fSignedS = fSignX, fS, f0 // -signum(x)*S1
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fCloseTo1Pol = fB12, fR8, fB7
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fH = fH, fD, fH // H2 = H1 + H1*d1
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fS = fS, fD, fS // S2 = S1 + S1*d1
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      // -signum(x)* S2 = -signum(x)*(S1 + S1*d1)
      fma.s1             fSignedS = fSignedS, fD, fSignedS
      nop.i              0
}
;;
{.mfi
      nop.m              0
      fnma.s1            fD = fH, fS, fHalf // d2 = 1/2 - H2*S2
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      // Cpi + signum(x)*PolB*S2
      fnma.s1            fCpi = fSignedS, fCloseTo1Pol, fCpi
      nop.i              0
}
{ .mfi
      nop.m              0
      // signum(x)*PolB * S2
      fnma.s1            fCloseTo1Pol = fSignedS, fCloseTo1Pol, f0
      nop.i              0
}
;;
{ .mfb
      nop.m              0
      // final result for 0.625 <= |x| < 1
      fma.d.s0           f8 = fCloseTo1Pol, fD, fCpi
      // exit here for  0.625 <= |x| < 1
      br.ret.sptk        b0
}
;;


// here if |x| < 0.625
.align 32
acos_base_range:
{ .mfi
      ldfe               fCpi = [rPiBy2Ptr] // Pi/2
      fma.s1             fA33 = fA33, fXSqr, fA31
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fA15 = fA15, fXSqr, fA13
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fA29 = fA29, fXSqr, fA27
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fA25 = fA25, fXSqr, fA23
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fA21 = fA21, fXSqr, fA19
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fA9 = fA9, fXSqr, fA7
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fA5 = fA5, fXSqr, fA3
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fA35 = fA35, fXQuadr, fA33
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fA17 = fA17, fXQuadr, fA15
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fX8 = fXQuadr, fXQuadr, f0 // x^8
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fA25 = fA25, fXQuadr, fA21
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fA9 = fA9, fXQuadr, fA5
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fms.s1             fCpi = fCpi, f1, f8 // Pi/2 - x
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fA35 = fA35, fXQuadr, fA29
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fA17 = fA17, fXSqr, fA11
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fX16 = fX8, fX8, f0 // x^16
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fA35 = fA35, fX8, fA25
      nop.i              0
}
{ .mfi
      nop.m              0
      fma.s1             fA17 = fA17, fX8, fA9
      nop.i              0
}
;;
{ .mfi
      nop.m              0
      fma.s1             fBaseP = fA35, fX16, fA17
      nop.i              0
}
;;
{ .mfb
      nop.m              0
      // final result for |x| < 0.625
      fnma.d.s0           f8 = fBaseP, fXCube, fCpi
      // exit here for |x| < 0.625 path
      br.ret.sptk        b0
}
;;

// here if |x| = 1
// acos(1) = 0
// acos(-1) = Pi
.align 32
acos_abs_1:
{ .mfi
      ldfe               fPiBy2 = [rPiBy2Ptr] // Pi/2
      nop.f              0
      nop.i              0
}
;;
.pred.rel "mutex", p8, p9
{ .mfi
      nop.m              0
      // result for x = 1.0
(p9)  fma.d.s0           f8 = f1, f0, f0 // 0.0
      nop.i              0
}
{.mfb
      nop.m              0
      // result for x = -1.0
(p8)  fma.d.s0           f8 = fPiBy2, f1, fPiBy2 // Pi
      // exit here for |x| = 1.0
      br.ret.sptk        b0
}
;;

// here if x is a NaN, denormal, or zero
.align 32
acos_special:
{ .mfi
      // point to Pi/2
      adds               rPiBy2Ptr = 272, rTblAddr
      // set p12 = 1 if x is a NaN
      fclass.m           p12, p0 = f8, 0xc3
      nop.i              0
}
{ .mlx
      nop.m              0
      // smallest positive DP normalized number
      movl               rDenoBound = 0x0010000000000000
}
;;
{ .mfi
      ldfe               fPiBy2 = [rPiBy2Ptr] // Pi/2
      // set p13 = 1 if x = 0.0
      fclass.m           p13, p0 = f8, 0x07
      nop.i              0
}
{ .mfi
      nop.m              0
      fnorm.s1           fNormX = f8
      nop.i              0
}
;;
{ .mfb
      // load smallest normal to FP reg
      setf.d             fDenoBound = rDenoBound
      // answer if x is a NaN
(p12) fma.d.s0           f8 = f8,f1,f0
      // exit here if x is a NaN
(p12) br.ret.spnt        b0
}
;;
{ .mfi
      nop.m              0
      // absolute value of normalized x
      fmerge.s           fNormX = f1, fNormX
      nop.i              0
}
;;
{ .mfb
      nop.m              0
      // final result for x = 0
(p13) fma.d.s0           f8 = fPiBy2, f1, f8
      // exit here if x = 0.0
(p13) br.ret.spnt        b0
}
;;
// if we still here then x is denormal or unnormal
{ .mfi
      nop.m              0
      // set p14 = 1 if normalized x is greater than or
      // equal to the smallest denormalized value
      // So, if p14 is set to 1 it means that we deal with
      // unnormal rather than with "true" denormal
      fcmp.ge.s1         p14, p0 = fNormX, fDenoBound
      nop.i              0
}
;;
{ .mfi
      nop.m              0
(p14) fcmp.eq.s0         p6, p0 = f8, f0      // Set D flag if x unnormal
      nop.i              0
}
{ .mfb
      nop.m              0
      // normalize unnormal input
(p14) fnorm.s1           f8 = f8
      // return to the main path
(p14) br.cond.sptk       acos_unnormal_back
}
;;
// if we still here it means that input is "true" denormal
{ .mfb
      nop.m              0
      // final result if x is denormal
      fms.d.s0           f8 = fPiBy2, f1, f8 // Pi/2 - x
      // exit here if x is denormal
      br.ret.sptk        b0
}
;;

// here if |x| > 1.0
// error handler should be called
.align 32
acos_abs_gt_1:
{ .mfi
      alloc              r32 = ar.pfs, 0, 3, 4, 0 // get some registers
      fmerge.s           FR_X = f8,f8
      nop.i              0
}
{ .mfb
      mov                GR_Parameter_TAG = 58 // error code
      frcpa.s0           FR_RESULT, p0 = f0,f0
      // call error handler routine
      br.cond.sptk       __libm_error_region
}
;;
GLOBAL_LIBM_END(acos)



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
        stfd [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
        stfd [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
        stfd [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
        add   GR_Parameter_RESULT = 48,sp
        nop.m 0
        nop.i 0
};;
{ .mmi
        ldfd  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#