s_tanh.s

glibc 2.9,最新版的C语言库函数

{.mfi
      ldfe           fA1         = [rCoeffAddr3], 32       // Load A1
      fma.s1         fRes        = fA19, fArgAbsNorm, fA18 // Polynomial
      nop.i          0
}
{.mfi
      ldfe           fA0         = [rCoeffAddr4], 32       // Load A0
      nop.f          0
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA17        = fA17, fArgAbsNorm, fA16  // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA15        = fA15, fArgAbsNorm, fA14  // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fTDeg7      = fTDeg3, fTQuadr, f0     // Polynomial
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fA13        = fA13, fArgAbsNorm, fA12 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA11        = fA11, fArgAbsNorm, fA10 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA9         = fA9, fArgAbsNorm, fA8   // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fRes        = fRes, fTSqr, fA17       // Polynomial
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fA7         = fA7, fArgAbsNorm, fA6 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA5         = fA5, fArgAbsNorm, f0  // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA15        = fA15, fTSqr, fA13     // Polynomial  
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fA4         = fA4, fArgAbsNorm, fA3 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA2         = fA2, fArgAbsNorm, fA1 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA11        = fA11, fTSqr, fA9 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0                                       
      fma.s1         fA7         = fA7, fTSqr, fA5  // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0                                       
      fma.s1         fRes        = fRes, fTQuadr, fA15 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0                                       
      fma.s1         fA4         = fA4, fTSqr, fA2     // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fRes        = fRes, fTQuadr, fA11 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0                                       
      fma.s1         fA4         = fA7, fTDeg3, fA4    // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fRes        = fRes,  fTDeg7, fA4  // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      // result for negative argument
(p15) fms.d.s0       f8          = fRes, fArgAbsNormSgn, fA0 // Polynomial
      nop.i          0
}
{ .mfb
      nop.m          0
      // result for positive argument
(p14) fma.d.s0       f8          = fRes, fArgAbsNormSgn, fA0 // Polynomial
      br.ret.sptk    b0
};;


// |x| < 0.25 Path /////////////////////////////////////////////////////////////
.align 32
tanh_near_zero:
{ .mfi
      adds           rCoeffAddr1 = 0xC80, rDataPtr      // address of A9
      fma.s0         fTSqr       = fArgSqr, fArgSqr, f0 // x^4 
      nop.i          0
}
{ .mfi
      adds           rCoeffAddr2 = 0xCB0, rDataPtr      // address of A7
      nop.f          0
      nop.i          0
};;

{ .mfi
      ldfpd          fA9, fA8    = [rCoeffAddr1], 16 // Load A9, A8
      nop.f          0
      nop.i          0
}
{ .mfi
      ldfpd          fA7, fA6    = [rCoeffAddr2], 16 // Load A7, A6
      nop.f          0
      nop.i          0
};;

{ .mfi
      ldfpd          fA5, fA4    = [rCoeffAddr1], 16 // Load A5, A4
      nop.f          0
      nop.i          0
}
{ .mfi
      ldfpd          fA3, fA2    = [rCoeffAddr2], 16 // Load A3, A2
      nop.f          0
      nop.i          0
};;

{ .mfi
      ldfe           fA1         = [rCoeffAddr1] // Load A1
      nop.f          0
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fTQuadr     = fTSqr, fTSqr, f0 // x^4
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fRes        = fA9, fArgSqr, fA8 // Polynomial
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fA7         = fA7, fArgSqr, fA6 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA3         = fA3, fArgSqr, fA2 // Polynomial
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fA5         = fA5, fArgSqr, fA4 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA1         = fA1, fArgSqr, f0 // Polynomial
      nop.i          0
}
{ .mfi
      nop.m          0
      fma.s1         fTQuadrSgn  = fTQuadr, f8, f0  // x^4 * x
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fRes        = fRes, fTSqr, fA7 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fA1         = fA3, fTSqr, fA1 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fRes        = fRes, fTSqr, fA5 // Polynomial
      nop.i          0
};;

{ .mfi
      nop.m          0
      fma.s1         fRes        = fRes, fTQuadr, fA1 // Polynomial
      nop.i          0
};;

{ .mfb
      nop.m          0
      fma.d.s0       f8          = fRes, f8, f8 // x+x*Polynomial
      br.ret.sptk    b0                         // Exit for |x| < 0.25
};;





// 19.0625 <= |x| < +inf Saturation path ///////////////////////////////////////
.align 32
tanh_saturation:
{ .mfi
      adds           rDataPtr    = 0xCD0, rDataPtr  // address of A0
      nop.f          0
      nop.i          0
};;

{ .mfi
      ldfe           fA0         = [rDataPtr]       // Load  A0 = 2^(-63)
      nop.f          0
      nop.i          0
};;

{ .mfb
      nop.m          0
      fma.d.s0       f8          = fA0, fSignumX, f0 // sign(x)*(1.0-2^(-63))
      br.ret.sptk    b0                       // Exit for 19.0625 <=|x|< +inf
};;




      
//  0, denormals and special IEEE numbers path /////////////////////////////////
_tanh_spec:

{ .mfi 
      cmp.lt         p15, p14    = rArg, r0 // Is arg negative (p15) 
                                            // or positive p14)
      fclass.m       p6,p0       = f8, 0x23 // To filter infinities
                                          // 0x23 = @pos|@neg|@inf 
      nop.i          0
};;

{ .mfi 
      nop.m          0
      fclass.m       p7,p0       = f8, 0xC7 // To filter NaNs & Zeros
                                 // 0xC7 = @pos|@neg|@zero|@qnan|@snan
      nop.i          0
};;

{ .mfb 
      nop.m          0
(p6)  fmerge.s       f8          = f8, f1     // +/-1 for INF args 
(p6)  br.ret.spnt    b0                       // exit for x = INF
};;

{ .mfb 
      nop.m          0
(p7)  fma.d.s0       f8          = f8, f1, f8    // +/-0 for 0 args 
                                                 // and NaNs for NaNs
(p7)  br.ret.spnt    b0                          // exit for x = NaN or +/-0
};;

{ .mfi 
      nop.m          0
      fnorm.s0       f8          = f8            // Normalize arg
      nop.i          0
};;

.pred.rel "mutex",p14,p15
{ .mfi 
      nop.m          0
(p14) fnma.d.s0      f8          = f8, f8, f8  // res = r-r^2
      nop.i          0
}
{ .mfb 
      nop.m          0
(p15) fma.d.s0       f8          = f8, f8, f8  // res = r+r^2
      br.ret.sptk    b0          // 0, denormals, specials return
};;

GLOBAL_LIBM_END(tanh)