s_tanhf.s

glibc 2.9,最新版的C语言库函数

      fmerge.s       fSignumX = f8, f1            // signum(x)
      shl            rSignBit = rSignBit, 31      // mask for sign bit
}
{ .mfi
      adds           rBound = 0x3EA, r0
      nop.f          0
      addl           rSaturation = 0x4112, r0
}
;;

{ .mfi
      andcm          rOffset2 = rArg, rMask
      fclass.m       p6,p0 = f8, 0xc7             // is x [S,Q]NaN or +/-0 ?
      shl            rBound = rBound, 20          // 1.0f in GR
}
{ .mfb
      andcm          rAbsArg = rArg, rSignBit     // |x| in GR
      nop.f          0
(p7)  br.cond.spnt   tanhf_denormal               // branch out if x is denormal
}
;;

{ .mfi
      adds           rCoeffAddr2 = 352, rDataPtr
      fclass.m       p9,p0 = f8, 0x23            // is x +/- inf?
      shr            rOffset2 = rOffset2, 21
}
{ .mfi
      cmp.lt         p10, p8 = rAbsArg, rBound   // |x| < 0.3125? 
      nop.f          0
      adds           rCoeffAddr3 = 16, rDataPtr
}
;;

{ .mfi
(p8)  sub            rBias = rOffset2, rBias2
      fma.s1         fArg4 = fArgSqr, fArgSqr, f0 // x^4
      shl            rSaturation = rSaturation, 16
}
{ .mfb
(p10) adds           rBias = 0x14, r0
(p6)  fma.s.s0       f8 = f8,f1,f8                // NaN or +/-0
(p6)  br.ret.spnt    b0                           // exit for x = NaN or +/-0
}
;;

{ .mfi
      shladd         rCoeffAddr1 = rBias, 4, rDataPtr
      fma.s1         fArg3Sgn = fArgSqr, f8, f0  // sign(x)*|x|^3
      // is |x| < 9.125? 
      cmp.lt         p11, p12 = rAbsArg, rSaturation  
}
{ .mfi
      shladd         rCoeffAddr3 = rBias, 4, rCoeffAddr3
      fma.s1         fArg3 = fArgSqr, fAbsArg, f0 // |x|^3
      shladd         rCoeffAddr2 = rBias, 3, rCoeffAddr2
}
;;

{ .mfi
(p11) ldfpd          fC0, fC1 = [rCoeffAddr1]
(p9)  fmerge.s       f8 = f8,f1                   // +/- inf
(p12) adds           rDataPtr = 544, rDataPtr 
}
{ .mfb
(p11) ldfpd          fC2, fC3 = [rCoeffAddr3], 16
      nop.f          0
(p9)  br.ret.spnt    b0                           // exit for x = +/- inf
}
;;

{ .mfi
(p11) ldfpd          fA0, fA1 = [rCoeffAddr2], 16
      nop.f          0
(p8)  cmp.eq.unc     p13, p0 = rBias, rNearSaturation
}
{ .mfi
      add            rCoeffAddr1 = 48, rCoeffAddr1
      nop.f          0
      nop.i          0
}
;;

{ .mfi
(p11) ldfpd          fD0, fD1 = [rCoeffAddr3]
      nop.f          0
      nop.i          0
}
{ .mfb
(p11) ldfpd          fD2, fB0 = [rCoeffAddr1]
      // sign(x)*|x|^2
      fma.s1         fArgSqrSgn = fArgSqr, fSignumX, f0
(p10) br.cond.spnt   tanhf_near_zero
}
;;

{ .mfi
(p11) ldfpd          fA2, fA3 = [rCoeffAddr2], 16
      fcmp.lt.s1     p15, p14 = f8,f0
      nop.i          0
}
{ .mfb
(p12) ldfd           fA0 = [rDataPtr]
      fma.s1         fArg4Sgn = fArg4, fSignumX, f0 // sign(x)*|x|^4
(p12) br.cond.spnt   tanhf_saturation
}
;;
{ .mfi
      nop.m          0
      fma.s1         fArg7Sgn = fArg4, fArg3Sgn, f0  // sign(x)*|x|^7
      nop.i          0
}
{ .mfb
      nop.m          0
      fma.s1         fArg6Sgn = fArg3, fArg3Sgn, f0  // sign(x)*|x|^6
(p13) br.cond.spnt   tanhf_close_to_saturation      
}
;;

{ .mfi
      nop.m          0
      fma.s1         fPolC = fC3, fAbsArg, fC2    // C3*|x| + C2
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fPolCTmp = fC1, fAbsArg, fC0 // C1*|x| + C0
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fPolA = fA1, fAbsArg, fA0    // A1*|x| + A0
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      fma.s1         fPolD = fD1, fAbsArg, fD0    // D1*|x| + D0
      nop.i          0
}
{ .mfi
      nop.m          0
      // sign(x)*(|x|^7 + D2*x^6)
      fma.s1         fPolDTmp = fArg6Sgn, fD2, fArg7Sgn
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fPolATmp = fA3, fAbsArg, fA2  // A3*|x| + A2 
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fB0 = fB0, fArg4, f0          // B0*x^4
      nop.i          0
};;

{ .mfi
      nop.m          0
      // C3*|x|^3 + C2*x^2 + C1*|x| + C0
      fma.s1         fPolC = fPolC, fArgSqr, fPolCTmp  
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      // PolD = sign(x)*(|x|^7 + D2*x^6 + D1*|x|^5 + D0*x^4)
      fma.d.s1       fPolD = fPolD, fArg4Sgn, fPolDTmp  
      nop.i          0
}
;;

{ .mfi
      nop.m          0
      // PolA = A3|x|^3 + A2*x^2 + A1*|x| + A0 
      fma.d.s1       fPolA = fPolATmp, fArgSqr, fPolA 
      nop.i          0
}
;;                 

{ .mfi
      nop.m          0
      // PolC = B0*x^4 + C3*|x|^3 + C2*|x|^2 + C1*|x| + C0 
      fma.d.s1       fPolC = fPolC, f1, fB0 
      nop.i          0
}
;;     

{ .mfi
      nop.m          0
(p14) fma.s.s0       f8 = fPolC, fPolD, fPolA     // for positive x
      nop.i          0                           
}
{ .mfb
      nop.m          0
(p15) fms.s.s0       f8 = fPolC, fPolD, fPolA     // for negative x
      br.ret.sptk    b0                           // Exit for 0.3125 <=|x|< 8.0
};;


// Here if |x| < 0.3125
tanhf_near_zero:
{ .mfi
      nop.m          0
      fma.s1         fPolC = fC3, fArgSqr, fC2    // C3*x^2 + C2
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fPolCTmp = fC1, fArgSqr, fC0  // C1*x^2 + C0
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fPolC = fPolC, fArg4, fPolCTmp // C3*x^6 + C2*x^4 + C1*x^2 + C0
      nop.i          0
};;

{ .mfb
      nop.m          0
      // x + x^3*(C3*x^6 + C2*x^4 + C1*x^2 + C0)
      fma.s.s0       f8 = fPolC, fArg3Sgn, f8
      br.ret.sptk    b0                           // Exit for |x| < 0.3125
};;

// Here if 9.125 <= |x| < +inf
tanhf_saturation:
{ .mfb
      nop.m          0
      fma.s.s0       f8 = fA0, fSignumX, f0       // sign(x)*(1.0d - 2^(-52))
      // Exit for 9.125 <= |x| < +inf
      br.ret.sptk    b0                           // Exit for 9.125 <=|x|< +inf
}
;;
      
// Here if  8.0 <= |x| < 9.125
tanhf_close_to_saturation:
{ .mfi
      nop.m          0
      fma.s1         fPolATmp = fA1, fAbsArg, fA0 // A1*|x| + A0
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fPolA = fA3, fAbsArg, fA2    // A3*|x| + A2
      nop.i          0
}    
;;

.pred.rel "mutex", p14, p15
{ .mfi
      nop.m          0
      // for positive x
(p14) fma.s.s0       f8 = fPolA, fArgSqr, fPolATmp
      nop.i          0                           
}
{ .mfb
      nop.m          0
      // for negative x
(p15) fms.s.s0       f8 = fPolA, fArgSqrSgn, fPolATmp
      br.ret.sptk    b0                           // Exit for 8.0 <=|x|< 9.125
};;

// Here if x is single precision denormal
tanhf_denormal:
{ .mfi
      nop.m          0
      fclass.m       p7,p8 = f8, 0x0a             // is x -denormal ?
      nop.i          0
}
;;

{ .mfi
      nop.m          0
(p7)  fma.s.s0       f8 = f8,f8,f8                // -denormal
      nop.i          0
}
{ .mfb
      nop.m          0
(p8)  fnma.s.s0      f8 = f8,f8,f8                // +denormal
      br.ret.sptk    b0                           // Exit for denormal
}
;;

GLOBAL_LIBM_END(tanhf)