e_expf.s

glibc 2.9,最新版的C语言库函数

}
;;

{ .mfi
      nop.m           0
      // check for underflow and tiny (+0) result
      fcmp.le.s1      p11, p0 = fNormX, fMAX_SGL_ZERO_ARG
      nop.i           0
}
{ .mfb
      nop.m           0
      fms.s1          fN = fNint, f1, fRightShifter // n in FP register
      // branch out if overflow
(p15) br.cond.spnt    EXP_CERTAIN_OVERFLOW
}
;;

{ .mfb
      getf.sig        rNJ = fNint           // bits of n, j
      // check for underflow and deno result
      fcmp.lt.s1      p13, p0 = fNormX, fMIN_SGL_NORM_ARG
      // branch out if underflow and tiny (+0) result
(p11) br.cond.spnt    EXP_CERTAIN_UNDERFLOW
}
;;

{ .mfi
      nop.m           0
      // check for possible overflow
      fcmp.gt.s1      p14, p0 = fNormX, fMAX_SGL_NORM_ARG
      extr.u          rJ = rNJ, 0, 6        // bits of j
}
{ .mfi
      addl            rN = 0xFFFF - 63, rNJ // biased and shifted n
      fnma.s1         fR = fLn2Div64, fN, fNormX // R = x - N*ln(2)/64
      nop.i           0
}
;;

{ .mfi
      shladd          rJ = rJ, 3, rTblAddr  // address in the 2^(j/64) table
      nop.f           0
      shr             rN = rN, 6            // biased n
}
;;

{ .mfi
      ld8             rJ = [rJ]
      nop.f           0
      shl             rN = rN, 52           // 2^n bits in DP format
}
;;

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

{ .mfi
      setf.d          fT = rN               // 2^n * 2^(j/64)
      fma.s1          fP = fA3, fR, fA2     // A3*R + A2
      nop.i           0
}
{ .mfi
      nop.m           0
      fma.s1          fRSqr = fR, fR, f0    // R^2
      nop.i           0
}
;;

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

{ .mbb
      nop.m           0
      // branch out if possible underflow
(p13) br.cond.spnt    EXP_POSSIBLE_UNDERFLOW
      // branch out if possible overflow result
(p14) br.cond.spnt    EXP_POSSIBLE_OVERFLOW
}
;;

{ .mfb
      nop.m           0
      // final result in the absence of over- and underflow
      fma.s.s0        f8 = fP, fT, fT
      // exit here in the absence of over- and underflow
      br.ret.sptk     b0
}
;;

EXP_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    EXP_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
EXP_CERTAIN_OVERFLOW:
{ .mmi
      addl            r17ones_m1 = 0x1FFFE, r0
;;
      setf.exp        fTmp = r17ones_m1
      nop.i           0
}
;;

{ .mfi
      alloc           r32=ar.pfs,0,3,4,0
      fmerge.s        FR_X = f8,f8
      nop.i           0
}
{ .mfb
      mov             GR_Parameter_TAG = 16
      fma.s.s0        FR_RESULT = fTmp, fTmp, fTmp // Set I,O and +INF result
      br.cond.sptk    __libm_error_region
}
;;

EXP_POSSIBLE_UNDERFLOW:

// Here if fMAX_SGL_ZERO_ARG < x < fMIN_SGL_NORM_ARG
// Underflow is a possibility, not a certainty

// We define an underflow when the answer with
//    ftz set
// is zero (tiny numbers become zero)

// Notice (from below) that if we have an unlimited exponent range,
// then there is an extra machine number E between the largest denormal and
// the smallest normal.

// So if with unbounded exponent we round to E or below, then we are
// tiny and underflow has occurred.

// But notice that you can be in a situation where we are tiny, namely
// rounded to E, but when the exponent is bounded we round to smallest
// normal. So the answer can be the smallest normal with underflow.

//                           E
// -----+--------------------+--------------------+-----
//      |                    |                    |
//   1.1...10 2^-3fff    1.1...11 2^-3fff    1.0...00 2^-3ffe
//   0.1...11 2^-3ffe                                   (biased, 1)
//    largest dn                               smallest normal

{ .mfi
      nop.m           0
      fsetc.s2        0x7F,0x41                // Get user's round mode, set ftz
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fma.s.s2        fFtz_urm_f8 = fP, fT, fT // Result with ftz set
      nop.i           0
}
;;

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

{ .mfi
      nop.m           0
      fcmp.eq.s1      p6, p7 = fFtz_urm_f8, f0 // Test for underflow
      nop.i           0
}
{ .mfi
      nop.m           0
      fma.s.s0        f8 = fP, fT, fT          // Compute result, set I, maybe U
      nop.i           0
}
;;

{ .mbb
      nop.m           0
(p6)  br.cond.spnt    EXP_UNDERFLOW_COMMON     // Branch if really underflow
(p7)  br.ret.sptk     b0                       // Exit if really no underflow
}
;;

EXP_CERTAIN_UNDERFLOW:
// Here if  x < fMAX_SGL_ZERO_ARG
// Result will be zero (or smallest denorm if round to +inf) with I, U set
{ .mmi
      mov             rTmp = 1
;;
      setf.exp        fTmp = rTmp               // Form small normal
      nop.i           0
}
;;

{ .mfi
      nop.m           0
      fmerge.se       fTmp = fTmp, f64DivLn2    // Small with non-trial signif
      nop.i           0
}
;;
      
{ .mfb
      nop.m           0
      fma.s.s0        f8 = fTmp, fTmp, f0 // Set I,U, tiny (+0.0) result
      br.cond.sptk    EXP_UNDERFLOW_COMMON
}
;;

EXP_UNDERFLOW_COMMON:
// Determine if underflow result is zero or nonzero
{ .mfi
      alloc           r32=ar.pfs,0,3,4,0
      fcmp.eq.s1      p6, p0 =  f8, f0
      nop.i           0
}
;;

{ .mfb
      nop.m           0
      fmerge.s        FR_X = fNormX,fNormX
(p6)  br.cond.spnt    EXP_UNDERFLOW_ZERO
}
;;

EXP_UNDERFLOW_NONZERO:
// Here if  x < fMIN_SGL_NORM_ARG and result nonzero;
// I, U are set
{ .mfb
      mov             GR_Parameter_TAG = 17
      nop.f           0                         // FR_RESULT already set
      br.cond.sptk    __libm_error_region
}
;;

EXP_UNDERFLOW_ZERO:
// Here if x < fMIN_SGL_NORM_ARG and result zero;
// I, U are set
{ .mfb
      mov             GR_Parameter_TAG = 17
      nop.f           0                         // FR_RESULT already set
      br.cond.sptk    __libm_error_region
}
;;

GLOBAL_IEEE754_END(expf)


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#