e_sinhf.s

glibc 2.9,最新版的C语言库函数

      // x*(64/ln(2)) + Right Shifter
      fma.s1          fNint = fNormX, f64DivLn2, fRightShifter
      add             rTblAddr = 8, rTblAddr
}
{ .mfb
      cmp.gt          p7, p0 = -2, rExp_x        // Test |x| < 2^(-2)
      fma.s1          fXsq = fNormX, fNormX, f0  // x*x for small path
(p7)  br.cond.spnt    SINH_SMALL                 // Branch if 0 < |x| < 2^-2
}
;;

{ .mfi
      nop.m           0
      // check for overflow
      fcmp.ge.s1      p12, p13 = fAbsX, fMIN_SGL_OFLOW_ARG
      mov             rJ_mask = 0x3f             // 6-bit mask for J
}
;;

{ .mfb
      nop.m           0
      fms.s1          fN = fNint, f1, fRightShifter // n in FP register
      // branch out if overflow
(p12) br.cond.spnt    SINH_CERTAIN_OVERFLOW
}
;;

{ .mfi
      getf.sig        rNJ = fNint                   // bits of n, j
      // check for possible overflow
      fcmp.gt.s1      p13, p0 = fAbsX, fMAX_SGL_NORM_ARG
      nop.i           0
}
;;

{ .mfi
      addl            rN = 0xFFBF - 63, rNJ      // biased and shifted n-1,j
      fnma.s1         fR = fLn2Div64, fN, fNormX // R = x - N*ln(2)/64
      and             rJ = rJ_mask, rNJ          // bits of j
}
{ .mfi
      sub             rNJ_neg = r0, rNJ          // bits of n, j for -x
      nop.f           0
      andcm           rN_mask = -1, rJ_mask      // 0xff...fc0 to mask N
}
;;

{ .mfi
      shladd          rJ = rJ, 3, rTblAddr // address in the 2^(j/64) table
      nop.f           0
      and             rN = rN_mask, rN     // biased, shifted n-1
}
{ .mfi
      addl            rN_neg = 0xFFBF - 63, rNJ_neg // -x biased, shifted n-1,j
      nop.f           0
      and             rJ_neg = rJ_mask, rNJ_neg     // bits of j for -x
}
;;

{ .mfi
      ld8             rJ = [rJ]                    // Table value
      nop.f           0
      shl             rN = rN, 46 // 2^(n-1) bits in DP format
}
{ .mfi
      shladd          rJ_neg = rJ_neg, 3, rTblAddr // addr in 2^(j/64) table -x
      nop.f           0
      and             rN_neg = rN_mask, rN_neg     // biased, shifted n-1 for -x
}
;;

{ .mfi
      ld8             rJ_neg = [rJ_neg]            // Table value for -x
      nop.f           0
      shl             rN_neg = rN_neg, 46 // 2^(n-1) bits in DP format for -x
}
;;

{ .mfi
      or              rN = rN, rJ // bits of 2^n * 2^(j/64) in DP format
      nop.f           0
      nop.i           0
}
;;

{ .mmf
      setf.d          fT = rN            // 2^(n-1) * 2^(j/64)
      or              rN_neg = rN_neg, rJ_neg // -x bits of 2^n * 2^(j/64) in DP
      fma.s1          fRSqr = fR, fR, f0 // R^2
}
;;

{ .mfi
      setf.d          fT_neg = rN_neg    // 2^(n-1) * 2^(j/64) for -x
      fma.s1          fP = fA3, fR, fA2  // A3*R + A2
      nop.i           0
}
{ .mfi
      nop.m           0
      fnma.s1         fP_neg = fA3, fR, fA2  // A3*R + A2 for -x
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fma.s1          fP = fP, fRSqr, fR // P = (A3*R + A2)*R^2 + R
      nop.i           0
}
{ .mfi
      nop.m           0
      fms.s1          fP_neg = fP_neg, fRSqr, fR // P = (A3*R + A2)*R^2 + R, -x
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fmpy.s0         fTmp = fLn2Div64, fLn2Div64       // Force inexact
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fma.s1          fExp = fP, fT, fT                 // exp(x)/2
      nop.i           0
}
{ .mfb
      nop.m           0
      fma.s1          fExp_neg = fP_neg, fT_neg, fT_neg // exp(-x)/2
      // branch out if possible overflow result
(p13) br.cond.spnt    SINH_POSSIBLE_OVERFLOW
}
;;

{ .mfb
      nop.m           0
      // final result in the absence of overflow
      fms.s.s0        f8 = fExp, f1, fExp_neg  // result = (exp(x)-exp(-x))/2
      // exit here in the absence of overflow
      br.ret.sptk     b0              // Exit main path, 0.25 <= |x| < 89.41598
}
;;

// Here if 0 < |x| < 0.25.  Evaluate 9th order polynomial.
SINH_SMALL:
{ .mfi
      add             rAd1 = 0x200, rTblAddr
      fcmp.lt.s1      p7, p8 = fNormX, f0       // Test sign of x
      cmp.gt          p6, p0 = -60, rExp_x      // Test |x| < 2^(-60)
}
{ .mfi
      add             rAd2 = 0x210, rTblAddr
      nop.f           0
      nop.i           0
}
;;

{ .mmb
      ldfpd           fA4, fA3 = [rAd1]
      ldfpd           fA2, fA1 = [rAd2]
(p6)  br.cond.spnt    SINH_VERY_SMALL           // Branch if |x| < 2^(-60)
}
;;

{ .mfi
      nop.m           0
      fma.s1          fX3 = fXsq, fNormX, f0
      nop.i           0
}
{ .mfi
      nop.m           0
      fma.s1          fX4 = fXsq, fXsq, f0
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fma.s1          fA43 = fXsq, fA4, fA3
      nop.i           0
}
{ .mfi
      nop.m           0
      fma.s1          fA21 = fXsq, fA2, fA1
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fma.s1          fA4321 = fX4, fA43, fA21
      nop.i           0
}
;;

// Dummy multiply to generate inexact
{ .mfi
      nop.m           0
      fmpy.s0         fTmp = fA4, fA4
      nop.i           0
}
{ .mfb
      nop.m           0
      fma.s.s0        f8 = fA4321, fX3, fNormX
      br.ret.sptk     b0                // Exit if 2^-60 < |x| < 0.25
}
;;

SINH_VERY_SMALL:
// Here if 0 < |x| < 2^-60
// Compute result by x + sgn(x)*x^2 to get properly rounded result
.pred.rel "mutex",p7,p8
{ .mfi
      nop.m           0
(p7)  fnma.s.s0       f8 = fNormX, fNormX, fNormX // If x<0 result ~ x-x^2
      nop.i           0
}
{ .mfb
      nop.m           0
(p8)  fma.s.s0        f8 = fNormX, fNormX, fNormX // If x>0 result ~ x+x^2
      br.ret.sptk     b0                          // Exit if |x| < 2^-60
}
;;

SINH_POSSIBLE_OVERFLOW:

// Here if fMAX_SGL_NORM_ARG < x < fMIN_SGL_OFLOW_ARG
// This cannot happen if input is a single, only if input higher precision.
// Overflow is a possibility, not a certainty.

// Recompute result using status field 2 with user's rounding mode,
// and wre set.  If result is larger than largest single, then we have
// overflow

{ .mfi
      mov             rGt_ln  = 0x1007f // Exponent for largest single + 1 ulp
      fsetc.s2        0x7F,0x42         // Get user's round mode, set wre
      nop.i           0
}
;;

{ .mfi
      setf.exp        fGt_pln = rGt_ln  // Create largest single + 1 ulp
      fma.s.s2        fWre_urm_f8 = fP, fT, fT    // Result with wre set
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fsetc.s2        0x7F,0x40                   // Turn off wre in sf2
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fcmp.ge.s1      p6, p0 =  fWre_urm_f8, fGt_pln // Test for overflow
      nop.i           0
}
;;

{ .mfb
      nop.m           0
      nop.f           0
(p6)  br.cond.spnt    SINH_CERTAIN_OVERFLOW // Branch if overflow
}
;;

{ .mfb
      nop.m           0
      fma.s.s0        f8 = fP, fT, fT
      br.ret.sptk     b0                     // Exit if really no overflow
}
;;

// here if overflow
SINH_CERTAIN_OVERFLOW:
{ .mfi
      addl            r17ones_m1 = 0x1FFFE, r0
      fcmp.lt.s1      p6, p7 = fNormX, f0     // Test for x < 0
      nop.i           0
}
;;

{ .mmf
      alloc           r32 = ar.pfs, 0, 3, 4, 0 // get some registers
      setf.exp        fTmp = r17ones_m1
      fmerge.s        FR_X = f8,f8
}
;;

{ .mfi
      mov             GR_Parameter_TAG = 128
(p6)  fnma.s.s0       FR_RESULT = fTmp, fTmp, f0 // Set I,O and -INF result
      nop.i           0
}
{ .mfb
      nop.m           0
(p7)  fma.s.s0        FR_RESULT = fTmp, fTmp, f0 // Set I,O and +INF result
      br.cond.sptk    __libm_error_region
}
;;

// Here if x unorm
SINH_UNORM:
{ .mfb
      getf.exp        rSignexp_x = fNormX    // Must recompute if x unorm
      fcmp.eq.s0      p6, p0 = f8, f0        // Set D flag
      br.cond.sptk    SINH_COMMON            // Return to main path
}
;;

GLOBAL_IEEE754_END(sinhf)


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
      stfs [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
{ .mfi
      stfs [GR_Parameter_X] = FR_X            // Store Parameter 1 on stack
      nop.f 0
      add   GR_Parameter_RESULT = 0,GR_Parameter_Y // Parameter 3 address
}
{ .mib
      stfs [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
      ldfs  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#