e_expf.s

Glibc 2.3.2源代码(解压后有100多M)

// If the input is really a single arg, then there will never be "Possible
// Underflow" or "Possible Overflow" arguments.
//

{ .mfi
     add       EXP_AD_M1 =    exp_GR_Ind1,EXP_AD_T1
     fcmp.ge.s1  p15,p14 = exp_norm_f8,EXP_MIN_SGL_OFLOW_ARG
     nop.i 999
}
{ .mfi
     shladd       EXP_AD_M2                =    exp_GR_Ind2,4,EXP_AD_T2
     fms.s1    exp_R                    =    f1,f8,exp_Mfloat
     nop.i 999 ;;
}

{ .mfi
     ldfe           exp_T1    =    [EXP_AD_M1]
     fcmp.le.s1  p11,p12 = exp_norm_f8,EXP_MAX_SGL_ZERO_ARG
     nop.i 999 ;;
}

{ .mfb
      ldfe           exp_T2   =    [EXP_AD_M2]
(p14) fcmp.gt.s1  p14,p0 = exp_norm_f8,EXP_MAX_SGL_NORM_ARG
(p15) br.cond.spnt L(EXP_CERTAIN_OVERFLOW) ;;
}

{ .mfb
      nop.m 999
(p12) fcmp.le.s1  p12,p0 = exp_norm_f8,EXP_MAX_SGL_UFLOW_ARG
(p11) br.cond.spnt L(EXP_CERTAIN_UNDERFLOW_ZERO)
}
;;

{ .mfi
      nop.m 999
(p13) fcmp.lt.s1  p13,p0 = exp_norm_f8,EXP_MIN_SGL_NORM_ARG
      nop.i 999
}
;;


{ .mfi
     nop.m                 999
     fma.s1    exp_Rsq   =    exp_R,exp_R,f0
     nop.i                 999
}
{ .mfi
     nop.m                 999
     fma.s1    exp_P3    =    exp_R,exp_coeff_P2,exp_coeff_P1
     nop.i                 999 
}
;;

{ .mfi
     nop.m                 999
     fma.s1    exp_P1    =    exp_R,exp_coeff_P6,exp_coeff_P5
     nop.i                 999 
}
{ .mfi
     nop.m                 999
     fma.s1    exp_P2    =    exp_R,exp_coeff_P4,exp_coeff_P3
     nop.i                 999
}
;;


{ .mfi
     nop.m                 999
     fma.s1    exp_P7    =    f1,exp_R,f1
     nop.i                 999
}
;;


{ .mfi
     nop.m                 999
     fma.s1    exp_P5    =    exp_Rsq,exp_P3,f0
     nop.i                 999
}
{ .mfi
     nop.m                 999
     fma.s1    exp_R4    =    exp_Rsq,exp_Rsq,f0
     nop.i                 999 
}
;;

{ .mfi
     nop.m                 999
     fma.s1    exp_T     =    exp_T1,exp_T2,f0
     nop.i                 999 
}
{ .mfi
     nop.m                 999
     fma.s1    exp_P4    =    exp_Rsq,exp_P1,exp_P2
     nop.i                 999 
}
;;

{ .mfi
     nop.m                 999
     fma.s1    exp_A     =    exp_T,exp_P7,f0
     nop.i                 999
}
{ .mfi
     nop.m                 999
     fma.s1    exp_P6    =    exp_R4,exp_P4,exp_P5
     nop.i                 999
}
;;

{ .bbb
(p12) br.cond.spnt L(EXP_CERTAIN_UNDERFLOW)
(p13) br.cond.spnt L(EXP_POSSIBLE_UNDERFLOW)
(p14) br.cond.spnt L(EXP_POSSIBLE_OVERFLOW)
}
;;

{ .mfb
     nop.m            999
     fma.s     f8   =    exp_T,exp_P6,exp_A
     br.ret.sptk     b0
}
;;

L(EXP_POSSIBLE_OVERFLOW):

// We got an answer. EXP_MAX_SGL_NORM_ARG < x < EXP_MIN_SGL_OFLOW_ARG
// overflow is a possibility, not a certainty
// Set wre in s2 and perform the last operation with s2

// We define an overflow when the answer with
//    WRE set
//    user-defined rounding mode
// is lsn +1

// Is the exponent 1 more than the largest single?
// If so, go to ERROR RETURN, else (no overflow) get the answer and
// leave.

// Largest single is FE (biased single)
//                   FE - 7F + FFFF = 1007E

// Create + largest_single_plus_ulp
// Create - largest_single_plus_ulp

// Calculate answer with WRE set.

// Cases when answer is lsn+1  are as follows:

//           midpoint
//              |
//  lsn         |         lsn+1
// --+----------|----------+------------
//              |
//    +inf          +inf      -inf
//                  RN         RN
//                             RZ
// exp_gt_pln contains the floating point number lsn+1.
// The setf.exp puts 0x1007f in the exponent and 0x800... in the significand.

// If the answer is >= lsn+1, we have overflowed.
// Then p6 is TRUE. Set the overflow tag, save input in FR_X,
// do the final calculation for IEEE result, and branch to error return.

{ .mfi
       mov         exp_GR_gt_ln    = 0x1007F 
       fsetc.s2    0x7F,0x42
       nop.i 999
}
;;

{ .mfi
       setf.exp      exp_gt_pln    = exp_GR_gt_ln
       fma.s.s2    exp_wre_urm_f8  = exp_T,  exp_P6, exp_A
       nop.i 999
}
;;

{ .mfi
       nop.m 999
       fsetc.s2 0x7F,0x40
       nop.i 999
}
;;

{ .mfi
       nop.m 999
       fcmp.ge.unc.s1 p6, p0       =  exp_wre_urm_f8, exp_gt_pln
       nop.i 999
}
;;

{ .mfb
       nop.m 999
       nop.f 999
(p6)   br.cond.spnt L(EXP_CERTAIN_OVERFLOW)  // Branch if really overflow
}
;;

{ .mfb
       nop.m 999
       fma.s        f8             = exp_T,  exp_P6, exp_A
       br.ret.sptk     b0                 // Exit if really no overflow
}
;;

L(EXP_CERTAIN_OVERFLOW):
{ .mmi
      sub   exp_GR_17ones_m1 = exp_GR_17ones, r0, 1 ;;
      setf.exp     f9 = exp_GR_17ones_m1
      nop.i 999 ;;
}

{ .mfi
      nop.m 999
      fmerge.s FR_X = f8,f8
      nop.i 999
}
{ .mfb
      mov        GR_Parameter_TAG = 16
      fma.s       FR_RESULT = f9, f9, f0    // Set I,O and +INF result
      br.cond.sptk  __libm_error_region ;;                             
}

L(EXP_POSSIBLE_UNDERFLOW): 

// We got an answer. EXP_MAX_SGL_UFLOW_ARG < x < EXP_MIN_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^-7f      1.1...11 2^-7f      1.0...00 2^-7e  
//   0.1...11 2^-7e                                     (biased, 1)
//    largest dn                               smallest normal

// If the answer is = 0, we have underflowed.
// Then p6 is TRUE. Set the underflow tag, save input in FR_X,
// do the final calculation for IEEE result, and branch to error return.

{ .mfi
       nop.m 999
       fsetc.s2 0x7F,0x41
       nop.i 999
}
;;

{ .mfi
       nop.m 999
       fma.s.s2     exp_ftz_urm_f8  = exp_T,  exp_P6, exp_A
       nop.i 999
}
;;


{ .mfi
       nop.m 999
       fsetc.s2 0x7F,0x40
       nop.i 999
}
;;

{ .mfi
       nop.m 999
       fcmp.eq.unc.s1 p6, p0     =  exp_ftz_urm_f8, f0
       nop.i 999
}
;;

{ .mfb
       nop.m 999
       nop.f 999
(p6)   br.cond.spnt L(EXP_CERTAIN_UNDERFLOW)  // Branch if really underflow 
}
;;

{ .mfb
       nop.m 999
       fma.s        f8             = exp_T,  exp_P6, exp_A
       br.ret.sptk     b0                  // Exit if really no underflow
}
;;

L(EXP_CERTAIN_UNDERFLOW):
{ .mfi
      nop.m 999
      fmerge.s FR_X = f8,f8
      nop.i 999
}
{ .mfb
      mov        GR_Parameter_TAG = 17
      fma.s       FR_RESULT  = exp_T, exp_P6, exp_A // Set I,U and tiny result
      br.cond.sptk  __libm_error_region ;;                             
}

L(EXP_CERTAIN_UNDERFLOW_ZERO):
{ .mmi
      mov   exp_GR_one = 1 ;;
      setf.exp     f9 = exp_GR_one
      nop.i 999 ;;
}

{ .mfi
      nop.m 999
      fmerge.s FR_X = f8,f8
      nop.i 999
}
{ .mfb
      mov        GR_Parameter_TAG = 17
      fma.s       FR_RESULT = f9, f9, f0    // Set I,U and tiny (+0.0) result
      br.cond.sptk  __libm_error_region ;;                             
}

.endp expf
ASM_SIZE_DIRECTIVE(expf)


.proc __libm_error_region
__libm_error_region:
.prologue
{ .mfi
        add   GR_Parameter_Y=-32,sp             // Parameter 2 value
	nop.f 999
.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
        nop.m 0
        nop.m 0
        add   GR_Parameter_RESULT = 48,sp
};;

{ .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
};; 

.endp __libm_error_region
ASM_SIZE_DIRECTIVE(__libm_error_region)


.type   __libm_error_support#,@function
.global __libm_error_support#